CultLib/docs/cultnet-selection-cut.md

CultNet typed selection: cut map

Date: 2026-09-17. Imagination output. Intended path: F:\Projects\CultLib\docs\cultnet-selection-cut.md. No code in this document has been written; nothing here is committed by Imagination. This is its own campaign with its own map; it does not touch docs/typescript-quic-realtime-cut.md or any code in that tree.

Progress, 2026-09-22. Cut 1 is running (Hands, Sonnet) on the branch cultnet/selection-cut1. It will not merge to main until Soul has passed it.

• Commit 0, e420410: the cache's public read surface. Twelve tests, six mutation entries, all killed.
• Commit 1, 17d10e0: CultNetSelection and its evaluator, the three v1 messages, and the three §1 selector engines deleted. v0 now lowers into a selection. Test results: Networking 162/162, Mesh 252/253 (one skip that was already there), Caching 188/188. Nine networking mutation entries, all killed.
• Mutation harness: scripts/mutate-dotnet.ps1 is a shared dotnet/NUnit harness and works for any NUnit project in this repository.
• Not run through mutations yet: S1, S2, S6, S9–S15, S17, S20 and S23. A test pins each one, but none has a mutation entry.

Two things were found that this map did not name:

• CultNetDocumentBinding can carry a schema id that differs from its descriptor's. The deleted engines checked for this; the evaluator reads only the descriptor and cannot see it. The fix is at the call sites (WithBindingSchemaAlias, ExpandSchemaBindingAliases), so the evaluator stays independent of the registry.
• A fourth copy of schema matching, at CultNetDatabase.cs:1707, which serves CreateShardSnapshotResponse. §1 counted three engines; there are four. Self's ruling, 2026-09-22: it collapses into the evaluator in commit 2. Leaving it would make the claim of one evaluator false. It is a deletion and takes a mutation entry like the others.

Q-J was not implemented, and the map is the reason. Commits 0 and 1 shipped number as double? (CultNetSelection.cs:71), CompareNumber as IEEE comparison (CultNetSelectionEvaluator.cs:103-113), a JSON schema that says "type": "number", and a cache NumberOf that returns double. The operator had already ruled decimal strings, and §16 recorded it, but §2 and D8 still specified float64, and nothing specified the canonical form or the comparison that the ruling explicitly required. Hands built what §2 said. The Rust Hands caught the conflict before writing a line and stopped, correctly. Self swept §2, D8 and the §7 table on 2026-09-22, and §2 "Numbers" now specifies the form. The Q-J fix lands on the branch before the Rust runtime, as its own commits. Commit 1 is not wire-final until it does.

Commit 2 landed (Sonnet) in five parts:

• 0787608 deletes _projectRecord. Its two tests now drive a real change of identity instead of a projection. It also collapses the fourth matcher.
• c5b03cb collapses the CultMesh alias checks.
• 4167cc4 adds the fourth matcher's mutation entry.
• 5ba48f0 finishes the Mesh collapse. CultMeshSnapshotRequestOptions and CultMeshPeerSnapshotDocumentOptions become records that carry one Selection, and their clones collapse to with. CleanSnapshotFilter and the triple rule give way to two owners, ResolveDefaultSelection and OverlaySelection. CultMeshHotBodySubscription carries one Selection. The four-tier read fallback becomes one read anchored on the record key, with the two tiers that could return the wrong key's record deleted.
• 4c33a62 adds the Mesh entries: three reverts killed and two loosenings survived. The survivals are fixture gaps, since no test passes an empty non-null recordKeys or a record key with different casing. They go to the fix batch.

Mesh is net −83 against §11's estimate of about −326. The shortfall has four parts:

• about −41 is the payload-sniff fallback, which was kept;
• IsSameCultDocumentSchema stays, because it already delegates entirely and inlining it would duplicate the resolution at four sites;
• CultMeshClient's three subscription sites still call the v0 SubscribeAsync, which this cut does not retype. They move with Cut 2, the watch cut;
• the read keeps two tiers, because a test enforces that.

This is an honest miss.

Two corrections to Hands' report. First, the payload-sniff fallback was not "ruled kept by the operator". Hands kept it in c5b03cb as a discrepancy, and Soul is judging whether it is a second selection decision or a decode capability. Second, Hands reports that one Networking test crashes the test host on a disposed-socket race in CultNetRudpSchemaServer/CultNetTransport and calls it pre-existing. Networking was 162/162 at 17d10e0, so that claim must be checked against 17d10e0 and main before anyone believes it. It goes to Soul by name.

Q-J fix landed at e5348d3 (commit 2f, C#). The Rust runtime is next.

Soul on commit 2, 2026-09-22 (Opus, pinned at 4c33a62). Commit 2 does not pass.

• S2-1, medium, predates the range. The crash is real, and Hands was right that it is older than this commit. It is a disposed-socket race: the ack SendTo at CultNetTransport.cs:2549/2593 runs on a test's background poll thread. About 18 tests start IsBackground poll threads, and only one of them (:1885) joins its thread before the server disposes. Crash rates:

  Checkout	Suite	Crashes
  17d10e0	full	1 of 6
  17d10e0	subscription subset	5 of 12
  main	subscription subset	1 of 12

  "162/162" was luck. A second defect sits in the harness: mutate-dotnet.ps1:43/61, under $ErrorActionPreference='Stop' on PowerShell 5.1, throws the whole run away when a host crash writes to stderr, where it should record the mutant as NO VERDICT.

• S2-2, high: the rewritten reconcile tests lost what the old ones proved. With _projectRecord gone, a source key always equals its record key, and the delete-then-put now arrives through the live Watch path, so Reconcile() has nothing left to do.

  • Two mutants that 17d10e0 killed now survive: dropping Reconcile's identity branch (:261), and returning early from Reconcile.
  • Dropping the add loop (:274) and dropping the live identity branch (:340) survive both before and after the rewrite.
  • The RecordKey comparisons at :261 and :340, and the throw at :311, are now dead code.

• S2-3, high: the fourth matcher was collapsed into the alias matcher, not into the evaluator. CreateShardSnapshotResponse (CultNetDatabase.cs:633-670) still loops on its own.

  • A probe sent {schemas:[], keys:[]} to both paths: the shard path returned 0 rows, CreateRawSnapshotResponse returned 2.
  • The binding-alias expansion does not run on the shard path either. That part is plausible, not probed.
  • §2 says an empty list is refused at the door. The evaluator reads [] as "no filter" (CultNetSelectionEvaluator.cs:84) and the shard path reads it as "nothing". Neither follows the map.

• S2-4, medium-high: the Mesh one-pass read returns the wrong schema's record. At CultMesh.cs:2730-2750, the first candidate at the key that passes the exact match, the alias match or a trial decode wins. With [MeshOther, MeshNote] at one key, a MeshNote read returned MeshOther. Before 5ba48f0, the exact schema was searched first.

• S2-5, medium: Mesh has two decode rules that disagree. The read accepts a record by trial decode. DecodeSnapshotDocuments rejects the same record by payload sniff. RawSnapshotPayloadMatchesSchema is a real decode capability, since both of its tests are keys-only. It is not a second selection decision. It is, though, a second copy of the registry's resolver (CultNetDocumentRegistry.TryResolveDescriptorByPayloadSchema / TryReadSchemaVersion / InferSchemaName, :596-670).

• S2-6, high: the public CultMeshSnapshotRequestOptions.Selection honours only Schemas and Keys. Mesh sends v0 only. A request with a limit, a field predicate and a header projection went out as schemas=null keys=null, and two full rows came back. A caller's Selection also skips the old cleaning, so Keys=[""] now returns nothing where it used to return everything.

• S2-7, low, predates the range: selection meaning is decided in two more places. ResolveDefaultSelection (:835) checks the raw Count before cleaning, so [" "] selects everything. ResolveDefaultSelection reads null as "no filter", and OverlaySelection (:840) reads it as "inherit".

• S2-8, low. Both options types changed from class to record, which gives them value equality and is acceptable. CultMeshHotBodySubscription.Selection exposes mutable arrays on a contract that was immutable. The prebuilt Unity GameCult.Mesh.dll carries the old API: a release-time follow-up, rebuilt at the next release cut.

• S2-9 and S2-10: a miscount in the entries header. For the two Mesh survivors, Soul's fixtures kill both.

• Held: Mesh 3 killed and 2 survived, as reported. Networking 10 of 10 killed. Every restore verified by hash. The alias-matcher routing is right. IsSameCultDocumentSchema delegates.

Self's rulings, 2026-09-22, for the commit 2 fix batch:

• Empty lists and blank keys are refused at the one door, as §2 already says. CultNetSelection.Validate refuses keys: [], schemas: [], and any key or schema that is empty or whitespace, with selection_invalid. null means no filter, everywhere. The evaluator never sees []. v0 lowering carries over v0's old cleaning: an empty v0 list, or one made only of blanks, lowers to null. That is a v0 compatibility rule the lowering owns, not a meaning the evaluator carries. ResolveDefaultSelection and OverlaySelection stop deciding what null means. They build a CultNetSelection and let the door decide. Overlay means "the caller's selection replaces the default wholesale".
• The shard snapshot answers through the evaluator, binding aliases included, with no loop of its own.
• Mesh keeps sending v0 in this cut. §9 says v0-only peers refuse v1, and FU-v0 retires v0 later. So CultMeshSnapshotRequestOptions refuses, loudly and typed at construction or send, any Selection term v0 cannot carry: fields, cites, cited, limit, cursor, descending, projection other than the default. Refused, never silently dropped. Mesh sending v1 is a follow-up, FU-Mesh-v1, triggered by the first Mesh caller that needs a term beyond schemas and keys, most likely the Huginn consumer or Cut 2.
• The Mesh read searches the exact schema first, then the alias match, then a foreign id decoded by payload. It uses one decode rule shared with DecodeSnapshotDocuments, and that rule is the registry's resolver. The Mesh copy in CultMeshSnapshots.cs is deleted in favour of CultNetDocumentRegistry's.
• The dead identity code is deleted, not tested around. :261, :340 and :311 can no longer fire. Reconcile gets a test for what it still owns after the deletion: flipping authorizeRecord per key while the request stays authorized, and the add loop. Both surviving mutants must die.
• Test hygiene belongs to this batch, because a flaky M0 is not a control. Every test that starts a background poll thread joins it before its server is disposed. mutate-dotnet.ps1 records a host crash as NO VERDICT and carries on.
• Also owed by Cut 1 and not yet done: R-2. Python replies unsupported_schema_version to v1 (§9, interop_peer.py:360-387, cultmesh_py/server.py:341), with one test per site.

Rust commit 3 landed at b090ca9. Its parity vectors wait on the fix batch below.

The commit 2 fix batch landed (Sonnet), 568e8e3 and 021e3f1:

• Door and lowering. The door refuses blank and whitespace entries, and CultNetV0SelectionLowering owns the v0 cleaning.
• Shard snapshot. It answers through the evaluator (NET-D4-ShardMembership-Revert), with a fixture for exclusion outside the shard's prefix.
• Mesh v0 terms. CultMeshSnapshotRequestOptions refuses any term v0 cannot carry, and HotBodySubscription.Selection returns a copy.
• Mesh read order. The read goes exact, then alias, then payload, under one decode rule, and the Mesh copy of the resolver is deleted. The shared rule is CultNetDocumentRegistry.TryReadSchemaVersion, now public, matched against the caller's own descriptor. TryResolveDescriptorByPayloadSchema is ambiguous when two CLR types alias one schema id, which the foreign-id fixture does on purpose.
• Dead code and test hygiene. The dead identity code in Reconcile is deleted, and the poll threads are joined.
• Harness. mutate-dotnet.ps1 records a host crash as NO VERDICT.
• R-2. The Python v0 peer refuses v1.

Tests: Networking 195/195 over ten consecutive runs, with no crash. Mesh 253/254 with one skip that was already there. Caching 192/192. cultnet-py 54/54. Mutation entries 22/22 killed across the fix-batch, Mesh and Networking files, and both Mesh survivors now die. Delta: Mesh −3, Networking +62.

Open for Soul:

• It is unclear whether the door now refuses [], as ruled, or only blank entries. The report names blank entries only.
• Seven cultmesh-py daemon tests fail on ModuleNotFoundError: cultcache_py in a spawned child process. Hands calls this pre-existing; main must confirm it.
• The anchor-matching fragility of mutate-dotnet.ps1 (separator spelling) is recorded, not fixed.

Commit 4 landed (Sonnet) at 00f4c02, completing Cut 1's execution.

• Parity vectors (S12): contracts/cultnet/interop/selection-vectors.{cs,rs}-written.json, 8 vectors each, read in both directions. Six are evaluated selections. Two are door refusals: schemas: [], and a blank key.

• Door: keys: [] and schemas: [] are refused (CultNetSelection.cs:326/328), as ruled.

• Tests: cultnet-rs 111 → 174, cultcache-rs 64 (unchanged), Networking 196 + 1 skip, Caching 192, Mesh 252/253.

• Mutation entries, all killed:

  • Rust 23/23;
  • Caching 10/10, including 4 Q-J;
  • Networking 15/15, including 5 Q-J.

  S6 first survived. Its fixture now carries two references to the same target under different roles. S9–S15 and S17 do not apply to Rust (Rust has no subscription client).

• Open for Soul, highest first:

  • No cites vectors cross the runtimes. Hands says C#'s MatchesCitation compares Citation.Target.SchemaId against the real SHA-256 schema id with no alias fallback, so "this fixture pair cannot agree". That is either a fixture limit or a wire-parity gap in the hop, and parity is this cut's invariant. Soul settles which.
  • Size: selection.rs is 1,407 lines and contracts.rs +595, about 2,000 in all against §14's roughly 430.
  • Hands calls S20 "structurally vacuous" and S23 without a surface in Rust. Both are claims, not yet reached, and need checking.

• A flaky test, recorded as a follow-up. reactive_document_coalesces_direct_same_schema_alias_member_writes (packages/cultnet-rs/tests/cultnet.rs) fails under the harness's invocation of the full binary and passes in isolation. The v0-lowering entry's command was narrowed around it. It is not this cut's code. Its owner is the reactive-document path.

Two pre-merge additions from the Huginn read-side map, 2026-09-22. They join the next selection fix batch, whatever the final Soul pass adds.

• P-1: Rust's Evaluation gains matched. Rust cannot fill the page's matched today.
• P-2: derive Eq on the selection wire types, because Huginn's request derives it.

Load-bearing for Huginn: Cursor::parse is public and exposes as_of. The organ answers a cursor's page at the cursor's own asOf. If a finding makes parse private, expose cursor_as_of instead.

Soul's whole-cut pass, 2026-09-22 (Opus, at 00f4c02). Cut 1 does not merge. The notes are in the session scratchpad at soul-ss4-notes.md.

• High, confirmed:
  • The C# reference never answers v1. No v1 listener is registered (CultNetDatabaseServer.cs:49-55, Server.cs:562-571). CreateSelectionResponse and LastWriteSequence have no callers. §7's server handling, D6's subscription v1 and S9 were never built. Hands reported none of it missing.
  • cited returns the wrong edges in both runtimes. Edges are filtered by whether the citer is on the page, but under cited the page rows are the cited rows. exists:true gives no edges. exists:false gives edges into rows that are not on the page. The committed vector pins the wrong answer.
  • The C# door accepts "5\n". .NET's $ matches before a final newline. Rust and the schema refuse it.
• Parity broken on caller inputs:
  • Float rendering at ties. 394 f32 and 48 f64 values out of 200k render differently, because .NET rounds half to even and Rust does not. A shortest form is a stand-in for the value in any case: an f32 of 3e20 is stored as 300000006012263202816, bits 0x61821ab1. (Corrected 2026-09-22. Soul's report and this map first gave 300000002010536247296. The Rust Hands and an independent Python Decimal check both give ...6012263202816.)
  • schemas. C# matches through the alias matcher and Rust matches exact ids. ["leaf_a"] gives 8 rows against 0.
  • Tiebreak. C# compares UTF-16 code units and Rust compares UTF-8 bytes, so astral keys sort in opposite orders.
  • cited edge order. It comes from a Rust HashMap and is nondeterministic.
  • any_of on a numeric alias. C# goes through a culture-dependent ToString().
• The door is advisory. Both select implementations are public and never validate. Rust's matches only debug_asserts the hop, so a release build ignores cites and cited. Mesh EnsureV0Compatible lets Keys=[] and [""] through, and v0 lowering turns them into "every key".
• Not what the map specifies:
  • Matched counts the page, which breaks C5.
  • v0 and shard snapshots re-evaluate the whole cache on every page, which is O(N²/200), and cannot detect a write between pages.
  • Cursor digests collide: ["a|b"] digests the same as ["a","b"].
  • S20 has no C# test. Two S20 mutants survive 196/196. Rust select ignores projection, so the rule has no owner there. Nothing tests SelectPage or CreateSelectionResponse.
• The Networking host still crashes, 2 in 11 runs. The flaky HasConsumers assertion (NetworkingTests.cs:1720) predates this cut. Joins sit outside finally in about 18 tests, so a failed assertion becomes a host crash. The "10 runs, no crash" claim did not reproduce.
• Low:
  • The JSON schema disagrees with the runtimes on limit bounds and on minItems.
  • Python R-2 refuses only v1, so v2 and later are still dropped.
  • TryReadSchemaVersion is a public sniff heuristic, and two Mesh sites repeat the same line around it.
  • Rust v0 dedups before reject_duplicates, so that function is dead.
  • mutate-dotnet.ps1 throws on \ against / in paths.
• Deletions Soul named:
  • contracts.rs:1206-1541, about 335 lines of hand codecs that the serde derives already round-trip, verified both ways;
  • expand_scientific_notation, about 46 lines, dead: 400k renders produced no exponent;
  • reject_duplicates;
  • RawDocumentHeader::from_document, which has no callers;
  • C#'s WithBindingSchemaAlias/ExpandSchemaBindingAliases duplicate, about 60 lines, and a shared lower-and-page loop, about 20 lines.
• Settled: cites is a fixture limit, not a parity gap. With real SHA-256 ids, 40 selections agree byte for byte. The seven cultmesh-py failures are environmental: with PYTHONPATH set, 78/78 pass. The flaky Rust test waits 30×5 ms for a flusher, which is too short under contention, and it is not this cut's.
• Held:
  • Every mutation suite killed everything.
  • Door refusals are byte-identical apart from "5\n".
  • The Q-J comparator agrees with Python Decimal.
  • No float sits on either comparison path.
  • Negation, limit clamping and v0 dedup are at parity.
  • Mesh reads in the order exact, then alias, then payload.
  • Mesh is actually 253/254.

Self's rulings for the Cut 1 fix batch, 2026-09-22:

• R-A. Serving v1 is part of this cut. Both C# servers register v1 listeners. The snapshot answers through CreateSelectionResponse in last-write order, and the subscription server answers v1 as D6 and §7 say, with S9 tested. A v1 request that is not served is a defect, not a follow-up.
• R-B. Hop edges follow the hop's direction. Under cites, the edges are the ones from page rows. Under cited, they are the ones into page rows. Every edge touches a page row. The order is deterministic: the page-row order, then (from, role, to) in code-point order. Regenerate the vectors.
• R-C. Every string comparison in the vocabulary uses Unicode code-point order. In Rust that is UTF-8 byte order. C# compares by code point, not by UTF-16 unit. This covers the tiebreak, edge order and anything else ordered.
• R-D. A floating-point member renders as its exact decimal expansion, in canonical form, and not as a shortest form. Every finite f32 and f64 is an exact dyadic rational and has a finite decimal expansion. That expansion is the stored value, it is the same in both runtimes by construction, and it matches the reason for Q-J. Implement it from mantissa and exponent, with BigInteger in C# and a small bignum in Rust. Add no new dependency unless the expansion cannot be written in about 60 lines. decimal and the integers are unchanged. Flagged for the operator's review: this is Self's reading of Q-J, and it changes what ge "3e20" means for an f32 of 3e20. Accepted by the operator, 2026-09-22 ("I accept your recommendations").
• R-E. There is one schema-identity rule in both runtimes. Rust gets the alias matcher (the owner is cultnet-rs, ported from CultNetSchemaAliasMatching). schemas, cites.target.schemaId and the binding id emitted on records all go through it. A cites target that matches no schema is refused at the door, never answered with an empty page. ToRawRecord and ToEdge emit the same id.
• R-F. The door is inside select. select validates first, in both runtimes, and returns the typed refusal. A public entry point that skips validation does not exist. Rust's hop handling is not a debug_assert. Mesh's EnsureV0Compatible runs the door first.
• R-G. matched is the total number of matches, not the page count (C5). Rust's Evaluation carries it (P-1). A single evaluation answers v0 and shard paging, taken as one snapshot.
• R-H. Cursor digests use length-prefixed encoding for every string and list, in both runtimes.
• R-I. The projection has an owner in both runtimes. cultnet-rs gets select_page, which produces SelectionPage under projection, with matched and edges. C# SelectPage and CreateSelectionResponse get tests, and so does S20. Page bytes are compared across the runtimes in the parity vectors.
• R-J. any_of on a numeric alias compares the canonical rendering.
• R-K. Test hygiene is finished, not claimed. Every poll-thread join moves into finally, and the HasConsumers race gets a deterministic wait. The fix is proven by 20 consecutive full runs.
• R-L. The contract follows the runtimes. The schema gains minItems: 1 on schemas and keys, and limit is documented as clamped. Python refuses every cultnet.*.vN with N > 0. TryReadSchemaVersion goes back to internal, behind a registry method that takes the descriptor. mutate-dotnet.ps1 normalises path separators.
• R-M. Soul's named deletions all land. P-2 (Eq) lands too.

The C# half of the fix batch landed (Sonnet), 7903853..3d32c67.

What changed:

• R-A: v1 is served by both servers, and D6 reconcile routing is in place.
• R-C: CultNetCodePointComparer gives code-point order.
• R-D: floats render as their exact decimal expansion, computed with BigInteger.
• R-E: cites go through the alias matcher, and the wire id is one WireSchemaId.
• R-F: the door is unconditional inside Select.
• R-G: Matched is the total, computed once through EvaluateAll and SelectAll.
• R-H: the cursor digest is length-prefixed.
• R-J: numeric any_of compares the canonical form.
• R-M: WithBindingSchemaAlias and the duplicated paging loop are deleted.
• R-I: tests kill both S20 mutants. One real bug was fixed in the process: ToEdge serialised a plain-typed payload as a document.
• R-K: every join is in finally. HasConsumers now waits for the condition.
• R-L: the schema carries minItems. Python refuses every vN>0 (only one site exists). The harness normalises path separators.
• Vectors: cs-written.json is regenerated and gains the float-tie, cites-by-alias, astral-order and projection cases.

Tests: Networking 201/202, where the one failure is the stale Rust vector until the Rust half lands. Caching 192. Mesh 253/254. Python 54, 78 and 34. In 20 full runs there was no crash, but one run had an unexplained failure.

Not done, sent to a second Hands: mutation entries for the fix-batch rules (there are only regression tests), S9, and a diagnosis of the 1-in-20 failure over 30 runs.

The Rust half of the fix batch landed (Sonnet), 8b114ed..a088435:

• 36/36 entries killed.
• Three real bugs were found while it ran, among them a v0 fast path that had inherited the door's "empty matches all".
• The contracts.rs hand codecs are gone (−362/+35).
• The Rust Hands corrected f32 3e20 to 300000006012263202816.

Shared parity fixture (Sonnet), 3317ddd and defb9a1. Both runtimes now build their vectors from contracts/cultnet/interop/selection-vectors.fixture.json, which uses real SHA-256 ids. A test bug is fixed: both readers had compared the page size with matched. Of 15 vectors in each direction, the following agree byte for byte: exact ids, float tie, astral order, header and document projection, cites/cited, and the refusals. Two real parity defects:

• Two incompatible alias rules. C#'s CultNetSchemaAliasMatching.Matches (CultNetDatabase.cs:74-87) strips .vN and compares against the descriptor's SchemaName. Rust's schema_alias::matches (selection.rs:452-459) strips .vN and compares against the hash. No alias resolves in both.
• C# does not refuse an unresolved cites.target (CultNetSelection.cs:361-391), which R-E requires. Rust does refuse it.

Self's ruling, 2026-09-22: the C# reference's alias rule is the rule. CultLib keeps wire parity against the C# reference. An alias is a schema name with a .vN suffix, resolved against the descriptor's name. A hash-shaped alias is not a wire form. Rust ports the name rule, which means its RowSet must expose each schema's name, and it deletes the hash-stripping matcher. C# refuses an unresolved cites.target at the door. Each vector that fails today becomes a vector both runtimes must pass.

The Rust alias port landed (Sonnet). Rust's schema_alias::matches is now (candidate, schema_id, schema_name), and it matches the C# reference row for row: exact id, bare name, .vN on the name, case-sensitive, the last .v marker wins, and .v with no digits is taken literally. The C# table was produced by running the reference source. The hash-stripping overload is deleted. SchemaVersion and CompatibleSchemaIds are not ported, because no Rust call site needs them; that is a stated reduction. Rust entries are 40/40 killed. The whole suite had never run under -Repo before, and its commands now carry --manifest-path.

Reader	Vectors from	Pass
Rust	cs-written	15/15
C#	rs-written	14/16

The two C# failures both come from the pending C# door refusal of an unmatched cites.target, which sits with the matrix Hands.

The early Soul pass (commits 0–1, pinned at 17d10e0) reported late, on 2026-09-22. It was interrupted three times by API errors. Much of what it found is already covered by R-A to R-M. These findings are new, and still open against the current branch:

• F2, high: typed refusals never reach the wire. CultNetErrorMessage (CultNetSchemaMessages.cs:~498-512) has no code and no details. Nothing maps the selection or cursor exceptions to an error. This means selection_invalid, cursor_stale, cursor_invalid and reference_outside_target exist only in-process.

• F6, high: the cursor can be forged. The digest is an unkeyed SHA-256 that anyone can recompute, so any caller can mint a cursor. A record key containing the U+241F delimiter makes a server's own cursor fail to parse.

• F7, high: authorization sees two different ids. On the snapshot path, the authorizer receives the binding's wire id. On the live path it receives the descriptor id (CultNetDatabaseSubscriptionServer.cs:~309 against :~451). A row delivered by the snapshot can therefore be refused on its first live change, or leak through a denylist. No test pins the live id.

• F8, medium: LastWriteSequence is not the order a snapshot reflects.

  • Replica apply paths (CultNetDatabase.cs:~1416, :~1459) never set it.
  • Nothing rebuilds it after a restart.
  • Sequences are per shard, but asOf is a single ulong.

• F11, medium: v0 lowering changed what v0 answers. A snapshot with recordKeys: [] used to answer empty, and now answers every row. v0 rows now come back in schema-and-key order instead of the requested key order.

• F12, medium: the committed schemas do not describe the bytes C# writes.

  • Headers $ref a record schema that requires payload.
  • MessagePack-CSharp writes both headers and documents, one of them nil, so every C# page fails the oneOf.
  • fieldPredicate does not forbid values and number together.

• F14, low, now reachable because v1 is served: field predicates on a Removed change throw. TryGetIndexValue(null), at CultNetDatabaseServer.cs:~375.

• F15, low: cites skips the out-of-target check for edges that do not point at the requested key. EnsureWithinDeclaredTarget also returns early when TargetType is null.

• Surviving mutants:

  • S1-Loose (reachable becomes all descriptors);
  • S2-AllOf;
  • S17-FirstLeaf;
  • Door-NoValidate, since fixed by R-F;
  • S5-DigestNoFields;
  • Auth-LiveSchemaId, which is F7;
  • QL-OutOfTargetCites, which is F15.

  S4 has no entry. A clone at C:\ss1 was left behind: the Soul pass could not remove it.

Self's rulings for fix batch 3, 2026-09-22. Batch 3 goes to Hands once the C# matrix Hands and the Rust alias port have landed, so that no two C# Hands edit src/ at the same time.

• R-N, F2: refusals are wire messages. cultnet.error.v0 gains code and details. That is an additive field, so the version is unchanged, and v0 peers ignore it. Every selection and cursor refusal, and reference_outside_target, maps to its code in both servers and in Rust. Tests decode the bytes a peer actually receives.
• R-O, F6: cursors are keyed. The digest becomes an HMAC under a key the answering server holds. The key is random per process, so a cursor does not survive a restart and answers cursor_invalid. That is acceptable, because cursors are short-lived and §2 already calls them "minted by the answering server". The cursor body uses length-prefixed fields, not a delimiter. Rust does the same, since Huginn mints cursors.
• R-P, F7: one id reaches the authorizer, and it is the wire id, on the snapshot, live and reconcile paths alike. A test pins the id each path passes.
• R-Q, F8: every path that commits a row, whether local or a replica apply, sets the last-write sequence. It is rebuilt from the mutation log at startup. A multi-shard selection's asOf is defined, or refused: a selection that spans shards is answered only when every row comes from one shard's log. Otherwise it is refused with cursor_stale/selection_invalid as appropriate, and the choice is documented in §2. A single shard is the case today.
• R-R, F11: v0 lowering preserves v0's pre-cut answers exactly. v0 is a frozen contract for the TS, Python and Kotlin peers. This corrects the earlier ruling in the commit 2 fix batch, which lowered an empty v0 list to "no filter". Where the pre-cut v0 server answered empty for recordKeys: [], it still answers empty. Where pre-cut v0 kept the requested key order, it still does. The pre-cut behaviour comes from the code at b3d9cf7, and tests prove it by running the same requests against both.
• R-S, F12: the schemas describe the bytes. Add a header schema, make the page shape match what MessagePack-CSharp writes (nil keys included), and make fieldPredicate exclusive. A test decodes real C# bytes against the schemas.
• R-T, F14 and F15: removed changes are evaluated without reading a document. A removal matches through the selection's keys and schemas alone, and field predicates count as unknown. cites checks every edge against its declared target.
• R-U: every surviving mutant from the early pass becomes an entry, and it must die. That covers S1-Loose, S2-AllOf, S17-FirstLeaf, S5-DigestNoFields, Auth-LiveSchemaId, QL-OutOfTargetCites and an S4 entry.

Fix batch 3 is landing in pieces, 2026-09-22.

• The harness is retired (operator ruling). The branch's mutate-dotnet.ps1, the four *.entries.ps1 and the Rust entries file are deleted. Tests pin behaviour, and Soul runs Stryker.NET and cargo-mutants on the diff.
• C#, first pass (8d7bee9..f366791): R-R (v0 lowering answers exactly what pre-cut v0 answered, key order included) and the cites-refusal vectors. C# now reads 16 of 16 Rust vectors. It stopped there, so R-N, R-O, R-P, R-Q, R-S, R-T and R-U were left.
  • R-R meets R-F: the raw v0 wire answers an explicit empty list with an empty page, while Mesh's EnsureV0Compatible refuses that list before sending. Both layers keep their own ruling, which is right.
• Rust (5cf31d1..c606eb3): R-O (keyed cursor), R-T, R-R, and the entries file deleted.
  • The cursor API for Huginn: Cursor::parse stays public and key-free and still exposes as_of. Only digest verification needs the process key, so no second helper was needed.
  • Rust reads 17 of 17 C# vectors.
  • R-N is pending in Rust by design. An earlier guess at the wire shape was backed out; Rust matches C#'s field names byte for byte once they land.
• C# pass A is in Hands: R-N, R-P, R-T. Pass B follows: R-O, R-Q, R-S, R-U.
• C# pass A landed (66b5e60, e8a3dec): R-N, R-P and R-T.
  • R-N's wire shape: CultNetErrorMessage gains code and details. CultNetErrorDetails carries field, value, asOf and current. The codes are selection_invalid, cursor_stale, cursor_invalid and reference_outside_target. The schema gains both fields additively.
  • Two catch sites that previously let the exception escape now answer the peer.
  • R-P: the live fast path authorized on the descriptor id and now uses the wire id, so all three paths agree.
  • R-T: a removal skips field predicates instead of dereferencing a null document. cites no longer early-returns on a null target, and the root cause was in CultCache: a many-reference declared without an explicit attribute left its target type null, even though the element type was statically known, so the out-of-target check silently skipped.
  • Networking 223 pass and 2 skip; Mesh 255.
  • The Caching claim was wrong. There is one failure, not three, and it is identical at the branch head, at b3d9cf7 and on main. It expects a write to a read-only directory to fail, which cannot happen in a container running as root. Environmental.
• Rust R-N landed (eff62e5): 213 tests pass. Parity is pinned by captured C# bytes for all four codes, taken from a probe run on the dotnet image.
  • An encoding trap the vectors would not have caught: MessagePack C# always writes all five declared keys in declaration order, nil included, while Rust's generic path sorts keys alphabetically. Rust now builds that map by hand. A key added or reordered in either error type must be mirrored there.
  • A defect in C#, sent to pass B: ForReferenceOutsideTarget prefixes a message that already carries its own prefix, so a peer reads reference_outside_target: reference_outside_target: ….
  • The parity-vector machinery carries no concept of an error message, so R-N's parity lives in its own byte-level test. That is the right home.
• C# pass B landed (a4cf54f..52082ec): R-O, R-Q, R-S and R-U, plus the doubled reference_outside_target prefix.
  • Networking 238 pass and 2 skip; Mesh 255.
  • Hand probes: six of seven killed. S1-Loose, S2-AllOf, S17, S5-DigestNoFields, Auth-LiveSchemaId and S4 all die.
  • QL-OutOfTargetCites is a genuine gap. Since pass A fixed CultCache.ResolveReferenceTarget, the fixture's many-reference gets an inferred target type, so the guard never fires for it. TargetType is now null only for a scalar reference typed as the bare ICultRecordRef, and no fixture reaches that shape. Open for Soul: is that shape registerable at all? If it is, pin it. If it is not, the null branch is dead and goes.
  • Two fixes found while writing the tests:
    • the cursor's length prefix counted UTF-16 chars, so an astral key did not round-trip;
    • MiniJsonSchemaValidator had no not, so the exclusivity test found nothing.
  • cultnet.raw-document-record.schema.json was missing schemaName, schemaVersion and schemaContentHash entirely. Any real C# page would have failed it on additionalProperties: false before R-S's own bugs even mattered.
  • Self's correction: cursors need no cross-runtime byte parity. The brief for pass B said Rust and C# must encode a cursor identically, on the premise that one runtime parses another's cursor. That premise is wrong. A cursor is opaque, minted by the answering server and echoed back by the caller, and the digests cannot verify across processes because each server holds its own key. Both encodings are length-prefixed five-field bodies with an HMAC-SHA256 digest, and they differ in base64 flavour and separators. That is correct and stays. Rust's own doc said so first.
  • Settled by Soul: three Caching failures, identical at b3d9cf7 and 52082ec. All three expect a write to a read-only directory to fail, and the container is root. Environmental. Pass B's count was right.

Soul's whole-cut pass, 2026-09-22 (Opus, at 52082ec). Cut 1 may not merge.

• Held: Networking 238 and 2 skips, Mesh 255, cultnet-rs 213. R-Q holds (both replica applies, the startup rebuild, and a shard-spanning refusal). R-O holds against a stranger and across a restart, and the astral key round-trips. R-H's digest covers every selection term. R-R holds and is honest. R-A holds. R-F's door is inside EvaluateAll, which every path reaches. The canonical-number door is anchored \A..\z.
• S-1, high: the hop indexes rows by record key alone. byKey (CultNetSelectionEvaluator.cs:176-178) and Rust's HashMap<&str, &R> (selection.rs:1457) both keep the last row at a key. CultCache keys are unique per schema, and the Mesh tests already put two schemas at one key. With rows in one order the selection refuses valid data as reference_outside_target; in the other order it succeeds. Under cited the index is a set of bare keys, so a row of one schema counts as cited because something cites another schema at the same key.
• S-2, high: R-T's cites half did not land in C#. C# filters edges by the citation's key (:355) before EnsureWithinDeclaredTarget (:359); Rust checks first. A many-reference declared to one leaf but holding an edge into another is refused by Rust and accepted by C#.
• S-3, high: a selection with both hops loses its cites edges in Rust. selection.rs:1541 computes one direction for every edge; C# anchors each edge by its own hop. No vector covers both hops.
• S-4, high: the error message fails its own schema. Nothing decodes real CultNetErrorMessage bytes against cultnet.error.schema.json. routingHint is undeclared under additionalProperties: false, and nil details.asOf, details.current, code and details all fail their declared types. The header schema shows the nullable pattern was understood and simply not applied here.
• S-5, medium: Stryker leaves 85 survivors over the cut's two core files: 396 mutants, 310 killed, 85 survived, 75 with no coverage at all, score 66%. The old "every entry killed" was an artifact of a hand-picked list.
• S-6, medium: the cursor is malleable in its own position. The digest covers the selection but not the body's asOf, ordinal, schemaId or recordKey. A caller holding one valid cursor can rewrite its position and reuse the digest.
• S-7, medium: R-I's claim is false. The vectors compare ids, matched, hasNext and edges. They never compare page bytes, asOf, next, or a refusal's code, and rs-written.json carries no projection vector at all.
• S-8, low: the out-of-target null branch is dead. D11 refuses a bare ICultRecordRef in both scalar and many form, so every surviving shape has a non-null target.
• S-9, low: asOf is the maximum across all shards, so a single-shard page advertises another shard's number, and an unrelated write invalidates a live cursor.
• S-10, low: MiniJsonSchemaValidator silently ignores minimum, maximum, maxLength, maxItems, anyOf, allOf, if/then/else, uniqueItems and format. The committed schemas use minimum 34 times, including limit's bound. Its pattern also runs through .NET Regex, whose $ accepts a trailing newline.
• S-11, low: the v1 subscribe refuses a hop before running the door and always names field cites.
• S-12, low: the fixture's prose still calls the alias vectors a cross-runtime defect, which the alias port fixed. Rust's port also drops SchemaVersion and CompatibleSchemaIds, so R-E is two rules agreeing on a subset.
• Where the evidence lives. The pass ran against 52082ec. Its nine hand probes — eight C# and the Rust both-hops probe that proved S-3 on a matched pair — are on branch soul-probe in the scratch clone C:\ss5 (e0fd672..107ff5f), unpushed. Nothing reached origin and the pinned checkout at F:\Projects\CultLib was never touched. Notes are at scratchpad/soul-sel-final-notes.md. Probes are scratch by doctrine and are meant to be thrown away; this pointer exists because S-1 through S-3 were proven there and fix batch 4's tests have to reproduce them independently. If C:\ss5 is gone, the probes are gone with it, and that is acceptable — the findings are recorded here, which is the durable copy.
• Soul's own corrections, same pass. It first reported a wedged cargo mutants job on Yggdrasil and asked for it to be killed. Checking the box rather than its log, the work directory was already gone: what it had lost was its detached ssh's output. Self had by then killed the process, which was Soul's own run, so nothing else was affected. The shape is worth keeping: a log is not the machine, and Self acted before confirming the process was actually stalled.
• The Stryker triage is owed, after four attempts, none of them Stryker's fault. The dotnet/sdk:10.0 image has no python3; the JSON reporter threw on CreateDirectory; two runs with a different output path ran about fifty minutes without finishing, against three minutes for the mutation phase that did complete. The completed run stands, and the command is in Soul's notes: dotnet-stryker --project GameCult.Networking.csproj --mutate "**/CultNetSelection*.cs" --reporter json --reporter progress --concurrency 4. Read the HTML report Stryker writes beside the JSON, rather than parsing JSON in the container.
• The triage landed on the fifth attempt, 2026-09-22, and it sharpens S-5 rather than softening it. Most of the 85 are message-string mutations and are not worth chasing; roughly a dozen are real, and they sit on the rules this cut exists to enforce:
  • CultNetSelection.cs:439, Any() → All() survived. §2's "a comparison on an alias any reachable schema declares non-numeric is refused" degrades to "refused only if every reachable schema declares it non-numeric". No fixture has an alias numeric on one reachable leaf and a string on another — the exact case the rule exists for.
  • CultNetSelection.cs:387, || → && survived. A cites.target with a schemaId and a blank recordKey, or the reverse, stops being refused.
  • member.IndexAlias ?? member.MemberName → member.MemberName survived at all four sites (CultNetSelectionEvaluator.cs:352,384,404 and CultNetSelection.cs:445). §2's role-naming rule for the hop is untested everywhere it is implemented.
  • CultNetSelectionEvaluator.cs:203,204, ThenBy ↔ ThenByDescending survived on both paths: the (ordinal, schemaId, recordKey) tiebreak is pinned by nothing.
  • CultNetSelectionEvaluator.cs:290-297, five order flips in EdgesFor, all survived. R-B's "page-row order, then (from, role, to) in code-point order" has no test pinning a single component, though R-B exists because that ordering was wrong once already.
  • CultNetSelectionEvaluator.cs:590-605 has no coverage at all: ComputeDigest's cites and cited branches, plus the number-presence flag at :584-585. R-H's collision-freedom rests on a reading, not a test.
  • :673,678 survived: Cursor.Parse's length-prefix bounds checks, on attacker-supplied base64. :307,322 survived: cursor positioning under descending, and ComparePosition's schema-vs-key tiebreak. :710,715,729 survived: the code-point comparator's surrogate path, despite R-C having a vector for it.
  • Equivalent, not to be chased: Selection.cs:103 <=, :118 >=, :121 dot <= 0 are guarded upstream and unreachable for canonical input; :103,107 turn sign * x into sign / x with x in {-1,1}; :109 Max→Min is a no-op because PadRight to a shorter length does nothing; :80 is a RegexOptions flag; Evaluator:272 is equivalent because full.Edges is empty without a hop.
• Soul withdrew a finding of its own in the same pass. It had reported that C#'s cursor digest does not sort its lists while Rust's does, offered as a low-severity intra-runtime difference. CultNetSelectionEvaluator.cs:637 sorts by code point, same as Rust. The remark belongs in no fix list.
• cargo-mutants cannot run on cultnet-rs at all. schema_discovery.rs uses include_str!("../../../contracts/…"), which escapes the package, so the tool's copied tree fails its baseline build. There is no mutation evidence for the Rust half, and the package's own source shape is why.

Self's rulings for fix batch 4, 2026-09-22:

• R-V (S-1). A row's identity is (schemaId, recordKey), never the key alone. Both runtimes index rows and incoming edges by the pair. RecordRef already carries both. Tests: two schemas at one key, in both row orders, give the same answer; cited does not match a row of another schema at the same key.
• R-W (S-2 and S-3). The hop behaves identically in both runtimes. The target check runs before any key filtering, and each edge is anchored by its own hop, so a selection carrying both cites and cited returns both sets. C# adopts Rust's ordering for the first; Rust adopts C#'s per-edge anchoring for the second. Vectors cover both.
• R-X (S-4 and S-10). A message that fails its own schema is a defect. cultnet.error.schema.json declares routingHint and every nullable field the way the header schema does. A test decodes real bytes for every code. MiniJsonSchemaValidator refuses a schema containing a keyword it does not implement, rather than ignoring it, and implements minimum, maximum, maxLength and maxItems. Its pattern anchors so a trailing newline cannot pass.
• R-Y (S-6). The cursor's HMAC covers the body as well as the selection, so the position cannot be rewritten. Test: a rewritten ordinal or key is refused.
• R-Z (S-7). The vectors compare page bytes, plus asOf, next and a refusal's code, and carry a header-projection and a document-projection vector on both sides. Where a field is legitimately per-server, such as a cursor, the vector states that and compares the rest.
• R-AA (S-5). Triage all 85 survivors by name. The 75 uncovered mutants mean whole regions have no test; write behavioural tests for them. A survivor that is genuinely equivalent gets a one-line reason in the report. The triage half is done (see the list under S-5, 2026-09-22). What is owed is a behavioural test for each named live survivor: the alias disagreement across reachable schemas, the half-blank cites.target, the four index-alias role sites, the row tiebreak, EdgesFor's ordering, ComputeDigest's hop branches, Cursor.Parse's bounds, descending cursor positioning, and the surrogate comparator. The equivalent mutants are recorded and are not to be chased.
• R-AB. Make cultnet-rs self-contained enough to mutate. Replace the escaping include_str! with a build script that stages the contracts into OUT_DIR, or another route that keeps every source path inside the package. Prove it by running cargo-mutants on the crate.
• S-8: delete the dead branch.
• S-9: asOf is the watermark of the shards the selection actually covers.
• S-11: run the door first, and name the field that is actually set.
• S-12: correct the fixture's prose, and state Rust's alias reduction where R-E is described.

Fix batch 4, C# half, landed (Sonnet) on cultnet/selection-cut1, 52082ec..b7d77b7:

• e99bf46 — R-V, R-W, S-8, R-Y: row identity is (schemaId, recordKey); the cursor's HMAC covers its own body.
• b3a8f6d — S-9, S-11: asOf is scoped to the shards the selection covers; the v1 subscribe door runs before the hop refusal.
• b7d77b7 — R-X, S-4, S-10: the error schema matches the wire; the validator refuses keywords it does not implement.

Counts on Yggdrasil: Networking 247 pass / 2 skip (was 238, so +9 tests), Mesh 255 pass. Caching's three failures are the read-only-directory tests that cannot fail under the stopgap's root container — unproven, not passing.

R-V proven in both row orders. Before the fix, a narrow reference declared to SelLeafA resolving a key shared with a SelLeafB row depended on insertion order: one order refused a valid reference as reference_outside_target, the other silently resolved to the wrong schema's row for cited membership. After it, both orders resolve through the leaf set and cited never matches the other schema's row at that key.

What R-X's new refusal exposed. Twelve uses of unimplemented keywords across three committed schemas, none reached by any existing test: cultnet.database-change-raw.schema.json uses allOf and if/then/else; both cultmesh.verse-catalog-*.schema.json use uniqueItems six times and anyOf once. Those three schemas are now formally unvalidatable until either the keywords are implemented or the schemas are rewritten. Nothing broke, because nothing ran them through the validator — which is the finding.

Two gaps Hands named rather than papered over:

• S-11 has no dedicated regression test. HandleSubscribeV1Async is private and reachable only through a live LiteNetLib round trip. Fixed by inspection, covered only by the suite passing. Owed.
• The three unvalidatable schemas above need a decision: implement the keywords, or rewrite them.

Ledger correction. Commits 0 and 1 came in at about twice the §14 estimate:

Package	Actual (net)	§14 estimate
Caching	+297	about +150
Networking	+851	about +400
contracts	163	about 200

The two new Networking files alone are 714 lines, against about 360. The hop's incoming index, the cursor's digest and mint, and the per-field door messages all cost more than the prose assumed. The binding-alias fix was not in the ledger at all. The estimate was wrong; the scope did not grow. The deletions in commits 2 to 4 (Mesh collapse, _projectRecord, the Rust snapshot_query.rs) are still to be counted against the total.

Status: cut map for two cuts. Cut 1: the selection vocabulary in the C# reference runtime and the Rust runtime (sections 1-17), with the follow-on runtime cuts and the Huginn consumer cut named and not mapped. Cut 2: watching a selection (section 18), mapped to the same standard, landing after Cut 1 and depending on it. Cut 1 goes to Hands without waiting on anything in Cut 2.

Rulings this map is written under (operator, 2026-09-17):

• The vocabulary lands in CultLib at the CultNet layer, not in Huginn. It fills a hole: the query side of the substrate was never dropped by decision, only never asked for. contracts/cultnet/cultnet-distributed-database.md:10-11 names RethinkDB changefeeds with "point, table/schema, and filtered watches" as prior art to keep; schema and key allowlists were built, filtered selection never was.
• The first cut covers the C# reference (src\GameCult.Networking) and Rust (packages\cultnet-rs, what Huginn consumes). TypeScript, Python and Kotlin follow when they have a caller.
• Deletes first: the two untyped stand-ins are replaced, not sat beside.
• The CultMesh subtraction audit is a required part of the map, with real line counts, and an honest miss beats a padded ledger.
• Nothing on the wire is serde-untagged (rule of this cut; evidence below).

What survives unchanged from the Huginn-targeted specification (scratchpad/cut10b-read-side-spec.md, D1-D4): the five capabilities, the chosen representation and the rejected ones with their reasons, the order and snapshot cursor, the role-typed hop, the one negation, and the bound. What changes: owner, repo, the value model of a field predicate (strings, because the cache's indexes are strings), the parity obligation, and the subtraction.

Pins

Repo	Branch	HEAD	State
CultLib	main	51fb449	dirty in native/GameCult.Mesh.Quic.Native/* and scripts/mutate-cultmesh.mjs from the live QUIC campaign; not touched. Every file:line below is against 51fb449.
Huginn	eureka/memory-organ	25a841d	the consumer; its cut is named in section 12 and mapped in the Epiphany campaign map.
Epiphany leaf	5cda0886	—	untouched.

Baselines: src/GameCult.Networking 14,848 lines; src/GameCult.Mesh 21,720 lines; tests/GameCult.Networking.Tests/NetworkingTests.cs 141 [Test]s; packages/cultnet-rs 111 #[test]s; cultnet-rs depends on serde, rmp-serde 1.3, rmpv, serde_json, cultcache-rs and no schemars (packages/cultnet-rs/Cargo.toml:14-31).

1. Body facts, each by source read

The reference's read side today.

• CultNetDatabase (src/GameCult.Networking/CultNetDatabase.cs): change kinds :15-45; Watch<T> :1052, WatchAllChanges :1063, WatchRecord<T> :1072; GetAll<T> :776; per-shard mutation log with sequence (:465-466, :528-590, :631 a snapshot carries ShardLogSequence). CultNetSchemaAliasMatching :47-97 is the one schema-alias matcher, internal static.
• Three separate selector engines evaluate the same two allowlists: CultNetDocumentRegistry.CreateRawSnapshotResponse :263-349 (snapshot, over cache.AllEntries by reflection at :325-342) with its own MatchesRequestedSchema :505-524; CultNetDatabaseServer.CreateChangeMessage :347-384 with Matches/MatchesSchema/InferSchemaName :401-468 (the LiteNetLib path, the one production construction site, CultNetLocal.cs:154); CultNetDatabaseSubscriptionServer.CreateChangeCore :442-485 with MatchesRequestedSchema :493-500 (the RUDP path). Two of them carry an identical reflection copy of CreateRawRecord (CultNetDatabaseServer.cs:386-399, CultNetDatabaseSubscriptionServer.cs:502-515).
• The two stand-ins. CultNetDatabaseSubscribeMessage (CultNetSchemaMessages.cs:788-826): schemaIds :805, recordKeys :809, includeSnapshot :813, then consumerRuntimeId, bodyIds, supportedBodyTransports (body-plane demand, not selection). The same pair on CultNetSnapshotRequestMessage :737-741, CultNetShardCatalogRequestMessage :949-953 (shard routing, stays), and CultNetSchemaCatalogRequestMessage :1591 (catalog, stays). _projectRecord on the subscription server (:27, :44, :50, :239-240, :314): no production caller constructs the subscription server at all — grep "new CultNetDatabaseSubscriptionServer" over src/ is empty; it is built only in tests (NetworkingTests.cs:1596-1605 passes projectRecord). Projection is authority that exists in the reference and reaches no wire and no caller.
• The cache already owns addressing. CultIndexAttribute (src/GameCult.Caching/CultDocumentContracts.cs:34, Alias :41) names a member as an index; the cache's index values are strings by construction (CultCache.cs:1084 Getter = document => getValue(document)?.ToString(); GetByIndex<T>(string alias, string value) :1503-1509). CultReferenceAttribute (:45, TargetType :53, Many :55) names a member as a reference; the descriptor records IsReference, IsMany, TargetSchemaName (CultCache.cs:1069-1080). So a field predicate's key is an index alias and a hop's role is a reference member, both declared by the document's owner.
• The declarations are public; the values are not. ToCatalogEntry() (CultCache.cs:151-176) is public and carries every member's Slot, MemberName, TypeName, IsReference, IsMany, TargetSchemaName, IsName, IndexAlias. But CultDocumentDescriptor.Members (:149) and IndexAccessors (:148) are internal, and GameCult.Caching declares no InternalsVisibleTo for GameCult.Networking (the only InternalsVisibleTo in src/ outside the vendored Unity tree is GameCult.Caching.MessagePack → GameCult.Caching.Tests). That is why CreateRawSnapshotResponse reaches for reflection at CultNetDocumentRegistry.cs:325-342. An evaluator outside the cache assembly cannot read a declared index's value today, by any means the cache offers. The earlier claim in this map that no cache change is needed for addressing was wrong: addressing is declared publicly and readable only inside the cache. Section 6 carries the read surface as an add; sections 7 and 14 carry its cost.
• Schemas are hand-written JSON in contracts/cultnet/*.schema.json, listed by file name in CultNetSchemaRegistry.cs:360-385; TypeScript compiles them with Ajv and refuses unknown fields (cultnet.database-subscribe.schema.json has additionalProperties: false); Python lists required fields per version by hand (packages/cultnet-py/src/cultnet_py/schema_catalog.py:283, :407); Rust types are a serde enum tagged on schemaVersion (packages/cultnet-rs/src/contracts.rs:252) with hand validation (require_optional_string_vec, :548-594). C# encodes with MessagePack-CSharp string keys ([Key("name")]), i.e. named maps; Rust encodes with rmp_serde::to_vec_named.
• Runtime coverage of the subscribe message today. C# server ×2, C# client (CultNetDatabaseSubscriptionClient.cs:138-151); TypeScript contract only (packages/cultnet-ts/src/contracts.ts:237-247, no server fanout, interop tests send it at test/interop/cultnet-interop.test.ts:719,798); Python server ×2 (cultnet_py/interop_peer.py:390-401 with DatabaseSubscription.matches :110-125, and cultmesh_py/server.py:315-331 with _subscription_matches_document :449-470, _schema_matches_request :699-716, and a third copy in cultmesh_py/node.py:1322-1345); Rust: none — cultnet-rs has SnapshotRequest with the two allowlists (contracts.rs:314-321) and no DatabaseSubscribe variant; Kotlin: none (a snapshot-request builder, CultMesh.kt:2111, and a require(schemaVersion == "…snapshot_request.v0") at :2195).
• How each runtime treats a message it does not know, which is the wire-parity fact section 9 rests on: C# throws (CultNetSchemaMessageSerialization.cs:85 _ => throw); TypeScript parseCultNetMessage (contracts.ts:728-744) looks the version up in its Ajv validators and throws when absent; Kotlin requires the exact version (CultMesh.kt:2195-2196); Rust's tagged enum fails to deserialize on an unknown tag; Python returns [] (interop_peer.py:360-387 falls off the chain; cultmesh_py/server.py:341 delegates to handle_message) — a silent drop, no reply, and for a known version with unknown fields Python reads only the fields it expects (message.get(...)), so it would answer a selector it did not understand with the wrong rows.
• Every committed write has a shard-log sequence — assumed from CultNetDatabase.cs:465 (_nextLogSequences) and the invariant "raw snapshot/put/delete messages must pass through CultNetDatabase" (cultnet-distributed-database.md:115-116); Hands confirms that a local PutAsync outside any shard also takes a sequence, because the order in section 3 rests on it. If it does not, the row's ordinal is its storedAt and the cursor refuses on any advance (section 3 says what that costs).

The Aetheria catalog as a hostile example, verified at F:\Projects\Aetheria\Assets\Scripts\ServerShared:

• Inheritance. ItemData.cs:272 abstract class ItemData (Name :275, Mass :284), :313 abstract class CraftedItemData : ItemData (which re-declares Name at :327, hiding the base's), and only the leaves carry [CultDocument]: SimpleCommodityData :300, CompoundCommodityData :330, GearData :449, and more. The cache's DiscoverMembers (CultCache.cs:838-875) walks current.BaseType, so a leaf's descriptor carries every inherited member with its attributes (GetCustomAttribute reads the member's own attribute wherever it is declared). Inherited addressing therefore already resolves per leaf. Corrected on a second read: duplicate slots are refused (:931-935), and a hiding member is already refused too — :937-943 groups keyed members by Member.Name and rejects any name carried at two different slots (HiddenMemberMessage), while a hiding member re-using one slot is caught by the duplicate-slot rule. So CraftedItemData's Name at ItemData.cs:327 hiding :275 is not registrable in CultCache as it stands, however the aliases fall; that is a finding for the Aetheria campaign, not a hole here. What is not refused by name is two members with different names carrying the same index alias (base [CultIndex("mass")] Mass, leaf [CultIndex("mass")] Weight). That does not collide silently either: the alias map is built by ToDictionary at :462-468, which throws ArgumentException("An item with the same key has already been added") — a real refusal, from the wrong owner, naming neither schema nor member nor alias. D10 is therefore a renaming of an existing refusal, not a new rule, and is smaller than it first looked.
• A reference to an abstract target. FactionProduct.cs:24 CultRecordRef<CraftedItemData> Design. PersistedMember.FromMember (CultCache.cs:1069-1071) takes the target type and reads CultDocumentAttribute from it — an abstract base has none, so TargetSchemaName is null today for exactly this reference. The registry can compute the leaf set by assignability (it already does for index lookups, :1506, and store routing, :1828) but records nothing.
• A dictionary reference with a value. ItemData.cs:333-334 [CultReference(typeof(PersonalityAttribute), many: true)] Dictionary<CultRecordRef<PersonalityAttribute>, float> DemandProfile. The descriptor records IsMany (:1080) and nothing else: no code in the cache enumerates a many-reference's targets, and the only accessor is Getter = document => getValue(document)?.ToString() (:1082), which on a dictionary yields the CLR type name. The edge and its payload are declared and unreadable.
• Numbers. PersistedMember.MemberType (:1053) knows the member is a float, and the catalog persists it as TypeName (:38, :192, compared by CompareSchemaShapes :605-629), so a member's numeric type is already declared and persisted; what is missing is a typed accessor beside the string one. Adding an accessor is a runtime change with no catalog-shape or hash change, because TypeName is already in the canonical shape.

The cache's index maps are unique indexes, and the evaluator must not use them. Index(stored) (CultCache.cs:1849-1862) writes MapOf(_indexes, (type, alias))[value] = key — one key per value, last writer wins. Two rows of one type sharing an index value do not collide loudly; the second silently displaces the first in the map. GetByIndex<T> (:1502-1508) reads that map and Single<T> (:1838-1848) throws only when several types assignable to T each contribute a key, never for two rows of one type. So GetByIndex answers "the row that most recently claimed this value", not "the rows with this value".

The consequence for this cut is a rule, not a caveat: the evaluator evaluates every predicate per row, through the accessor, and never through _indexes or GetByIndex. The declared alias is an addressing name — the cache's own lookup map is a different, lossier thing that happens to be built from the same declaration. A "use the index, it is already there" optimisation would turn kind any_of [weapon] into a one-row answer. S23 pins it, WatchByIndex already has the bug (section 18), and the unique-index behaviour itself is left exactly as it is: it is the cache's own contract for a single-record lookup and this cut does not renegotiate it.

A fourth selector engine, in Rust, that this map's first draft missed. packages/cultnet-rs/src/snapshot_query.rs (339 lines, pub use at lib.rs:29) already serves and asks snapshot requests:

• serve_read_only_raw_snapshot :64-130 is an engine — it walks the source's records, refuses a duplicate identity and an unregistered schema, applies a CultNetReadOnlySnapshotPolicy allowlist of (schema_id, record_key) pairs, then filters by the request's schema_ids and record_keys (:112-125). The policy is authorization and stays; :112-125 is the fifth copy of the same two allowlists and goes, as the C# three do.
• CultNetRawSnapshotQuery :158-285 is the client side, and it exposes the sharpest limit of the two-allowlist model: request() :195-217 turns a set of exact (schema, key) pairs into the cross product of their schemas and their keys, and accept_response :219-284 then errors on any record the cross product dragged in ("snapshot response contains unexpected record"). Asking for (A,x) and (B,y) asks for four and refuses two. v1 does not fix this: schemas × keys is still a cross product (section 13, Q-M). What v1 does change is that accept_response's check is a declared expectation, not a re-filter of a page the server got right, and section 8's forbidden writer is worded to keep that distinction honest.

The CultMesh audit (Explore pass over src/GameCult.Mesh and CultNetDocumentRegistry.cs, 92 sites classified) is in section 11 with its numbers.

2. The vocabulary, exactly

One shape, cultnet.selection.v1, carried by two request messages at a new version and answered by one page message. Field names are the wire spelling; C# and Rust type names follow each runtime's convention.

Selection {
  schemas:     string[]?              // kinds: schema ids or aliases, matched by CultNetSchemaAliasMatching; absent = every schema
  keys:        string[]?              // record-key allowlist; absent = every key
  fields:      FieldPredicate[]?      // conjunction; each is any-of over one declared index alias
  cites:       Citation?              // rows whose declared reference `role` (or any reference) names `target`
  cited:       Incoming?              // rows that some row does / does not name in `role`
  projection:  "header" | "document"  // default "header"
  descending:  bool                   // default false; the one order, reversed
  limit:       uint32?                // clamped 1..=200
  cursor:      string?                // opaque; minted by the answering server
}
FieldPredicate {                                           // one operator per predicate, named by `op`; no union, no untagged
  index:   string                                          // a declared index alias, reachable on some schema the selection can reach, else refused
  op:      "any_of" | "lt" | "le" | "ge" | "gt"
  values:  string[]?                                       // present iff op = any_of; non-empty
  number:  string?                                         // present iff op is a comparison; a canonical decimal (Q-J, below); the alias must be declared numeric on every reachable schema that declares it
}
Citation       { target: RecordRef, role: string? }        // target validated as a reference the row owner recognises
Incoming       { role: string, exists: bool }
RecordRef      { schemaId: string, recordKey: string }

SelectionPage {                                            // cultnet.snapshot_response_raw.v1 carries this
  matched:  uint32
  asOf:     uint64                                          // the snapshot the page is exact for
  next:     string?                                         // absent on the last page
  headers:  RawDocumentHeader[]?  |  documents: RawDocumentRecord[]?   // exactly one present, by projection
  edges:    Edge[]?                                         // present iff the selection has `cites` or `cited`: the edges the hop traversed
  shardId?, shardEpoch?, shardLogSequence?                   // as v0 carries them
}
Edge {
  from:     RecordRef                                       // the citing row
  role:     string
  to:       RecordRef                                       // the cited row, its schema resolved to the leaf actually stored
  payloadEncoding: "messagepack"?, payload: bytes?          // the value attached to the edge (a dictionary's value), typed by the member's declared value type; present under `document`, absent under `header`
}
RawDocumentHeader = RawDocumentRecord minus payload, payloadEncoding

Messages: cultnet.snapshot_request.v1 { messageId, selection, shardId?, shardEpoch? }; cultnet.database_subscribe.v1 { messageId, subscriptionId, selection, includeSnapshot, consumerRuntimeId?, bodyIds?, supportedBodyTransports? }; cultnet.snapshot_response_raw.v1 as above. database_change_raw.v0 and database_unsubscribe.v0 are unchanged: a change is a change.

Rules, each one owner:

• Order is (ordinal, schemaId, recordKey) ascending, or the reverse under descending. ordinal is the sequence of the commit that last wrote the row (the shard-log sequence in the reference; the receipt sequence in Huginn). The row owner supplies it; the vocabulary never reads a clock.
• Snapshot. A page is exact as of asOf. A cursor carries asOf, the last (ordinal, schemaId, recordKey), and a digest of the selection with cursor and limit cleared. A server that can present rows as of the cursor's asOf answers the next page over that snapshot; a server that cannot refuses cultnet.error.v0 { code: "cursor_stale", details: { asOf, current } }. The reference refuses when the shard log has advanced (rows mutate, and a replay is not this cut's); Huginn never refuses (append-only, one writer). One contract, two honest servers. A cursor that does not decode, or whose digest is not this selection's, is cursor_invalid.
• Fields are predicates over one declared index each, conjoined across predicates: any_of over string values, or one of four comparisons (lt, le, ge, gt) against one number over a declared numeric member. A row of a schema that does not declare the index never matches it. An index no reachable schema declares, an empty values, a comparison on an alias any reachable schema declares non-numeric, an empty keys or schemas list, and a role no schema declares are refused typed at the door (selection_invalid { field, value }), never answered as an empty page. any_of values are strings because the cache's indexes are strings (CultCache.cs:1084). Comparison numbers are canonical decimal strings under the operator's Q-J ruling. The rule is specified in full in the Numbers item below. No range operator (two predicates on one alias make a range), no ne, and no nesting.
• Numbers (Q-J; Self's specification, 2026-09-22, which the ruling asked for and this map had not yet written).
  • Canonical form. A number on the wire matches ^-?(0|[1-9][0-9]*)(\.[0-9]*[1-9])?$, and -0 is excluded. That fixes one spelling per value. It forbids leading zeros, trailing fractional zeros, a bare trailing point, an explicit +, exponent notation, negative zero, whitespace, and culture separators.
  • Non-canonical spellings are refused. Any other spelling is refused at the door with selection_invalid { field: "number", value }. It is never normalised. The reasons: the cursor digest and the parity vectors need one byte form per value, and a lenient parser differs between runtimes at exactly these edges. The ruling says "a non-canonical spelling of an equal value must not change an answer". That holds because such a spelling never reaches the evaluator. It gets a typed refusal, not a different page.
  • Row side. The cache renders a numeric member as its exact canonical decimal:
    • An integer type renders as its exact decimal. There is no 2^53 limit.
    • decimal renders with its trailing zeros stripped.
    • float and double render in the runtime's established shortest round-trip form ("R" in .NET, the standard shortest form in Rust), rewritten from exponent notation to positional digits. Parity pins this.
    • A NaN or infinite member matches no comparison.
  • Comparison. Both operands are canonical, so they are compared numerically by sign, then by the integer part's length, then by the integer digits, then by the fraction digits padded on the right. It is one small pure function per runtime, with no numeric parse and no precision limit.
  • Mutants that must die:
    • a lexicographic comparison ("10" < "9");
    • a comparison that ignores sign;
    • a comparison that ignores the integer part's length;
    • a door that accepts +1, 1.0, 01, 1e3 or -0;
    • a row rendering through float64 (a long member of 2^53+1 must compare greater than "9007199254740992").
  • Refusing a non-canonical number at the door, rather than normalising it, was Self's reading of the ruling. The operator accepted it on 2026-09-22.
• Inheritance. A schema is a leaf: only a type with [CultDocument] has a schema name, and schemas names leaves. A predicate on an alias applies to every reachable leaf whose descriptor carries a member with that alias, inherited or own — the cache already discovers inherited members per leaf, so "gear or commodities with mass over 10" is schemas: [geardata, simplecommoditydata], fields: [{index: mass, op: gt, number: 10}] with mass declared once on the abstract base. Two leaves that each declare a member with the same alias independently are two members; the predicate reads each leaf's own. One leaf that resolves the same alias to two members (a base member and a hiding re-declaration both carrying the alias) is refused at registration (D10). A numeric comparison on an alias that is numeric on one reachable leaf and a string on another is refused at the door, because the row owner cannot compare a string.
• The hop reads edges the row owner declares: in the reference, members with CultReferenceAttribute (role = the member's index alias or name); a one-reference yields one edge, a many-reference yields one edge per element, and a dictionary keyed by references yields one edge per key with the value as the edge's payload. In Rust, Row::references() yields (role, target, payload). A reference's target is a set of leaves: the declared target type's own schema if it is a leaf, else every registered leaf assignable to it (CraftedItemData → compoundcommoditydata, consumableitemdata, …). The hop follows an edge into any row whose schema is in that set. A stored edge naming a row whose schema is outside the set refuses the selection (reference_outside_target { from, role, to }), not skips it: an edge the declaration forbids is corrupt data and a read does not paper over it. The citer's status, fields and authorization are not consulted by the hop; one hop, no second.
• Edges in the answer. A selection with cites or cited returns, beside the rows, the edges the hop traversed, each with its payload under document projection and without it under header. A weighted reference and a bill of materials are the same shape: follow the edge, return the payload, filter the rows. Filtering on the payload as part of the traversal is not expressible (section 13): that is a join condition.
• Projection. header is the record without its payload; document is the record. The consumer decides what a header means for typed rows (Huginn's typed summary is its header; the substrate carries whichever type the page was instantiated with).
• Not on the wire: the admission window (admittedAfter/Before) of Huginn's Cut 9 — descending plus the cursor answer "the latest"; a date window is a range, and ranges are out (R-3).

The bound, carried forward in force, with one widening the operator ruled. Closed key set (declared aliases), a closed operator set (any-of and four comparisons, one per predicate, no nesting), no path language, no expression terms, no OR, one negation, one hop with its edges returned, no depth field, no sort field, no text, no predicate on an edge's payload in the traversal. Section 13 lists what is deliberately impossible. The capability set is now: C1 typed predicates over declared members, string any-of and numeric comparison; C2 projection; C3 order and cursor; C4 one hop, both directions, returning typed edge payloads; C5 existence and count without bodies.

Nothing on the wire is serde-untagged — rule, with evidence. Probe scratchpad/untagged-probe/ (result in untagged-probe-result.txt), run 2026-09-17 against rmp-serde 1.3.1 / rmp 0.8.15 / serde 1.0.229 (the versions Huginn's lockfile resolves; cultnet-rs pins rmp-serde 1.3): every #[serde(untagged)] enum with struct variants fails to round-trip under positional MessagePack (data did not match any variant of untagged enum) and passes only under named maps; externally tagged shapes pass under both. The reference encodes named maps and so does Rust today, so an untagged shape would work by accident of an undocumented encoding choice. grep untagged packages/cultnet-rs/src packages/cultcache-rs/src is empty at 51fb449; this cut keeps it empty and pins it (section 10).

3. Decisions, with the reasons

D1. Strings for field values, not typed enums. In the Huginn-targeted design a predicate was Option<Vec<FindingSeverity>>, closed at compile time. The substrate cannot know a consumer's enums, and the cache's own index model is string-valued; so the wire carries strings and the row owner refuses a value outside its enum at the door. The loss is compile-time closure; what is kept is runtime closure with a typed refusal, which is the same guarantee to a caller.

D2. Declared aliases, not a field-path language. index and role are names the document's owner declared with CultIndexAttribute and CultReferenceAttribute; an undeclared name is refused. A dot-path over the payload would be the string language the bound excludes, and the cache already has the declaration mechanism.

D3. A new message version, not new fields on v0. Section 1's fact: Python reads only the fields it knows and would answer a v0 selector it did not understand with the wrong rows. A v1 version string makes every runtime not in this cut refuse or drop by its own dispatch rather than mis-answer. The cost is that v0 stays answerable by the reference until the follow-ups land (R-1, a scheduled deletion with a named trigger).

D4. One evaluator in the reference, three engines deleted. Select over (descriptor, key, document) rows from the cache replaces CreateRawSnapshotResponse's loop, CultNetDatabaseServer.CreateChangeMessage's Matches, and CultNetDatabaseSubscriptionServer.CreateChangeCore. CultNetSchemaAliasMatching (CultNetDatabase.cs:47-97) becomes the one alias matcher and goes public; the four Mesh copies and two Networking copies go.

D5. Projection becomes a wire value; _projectRecord dies. The delegate ran after authorization and reached no caller. projection on the selection is the same authority as a value; _authorizeRequest and _authorizeRecord stay, because authorization is not selection.

D6. Subscriptions evaluate the selection on the snapshot and on every change, and hop-bearing selections reconcile. cites/cited are set-dependent: a change to row B can change whether row A matches. The subscription server already owns a diff-against-delivered reconcile (Reconcile :251-292, DeliveredBySourceRecordKey); a hop-bearing subscription runs it on every change instead of the per-change fast path. That is the mechanism by which "watch a question" falls out, and this cut does not add a local CultNetDatabase.Watch(Selection); section 11 records what that later cut would collapse; it is Cut 2, section 18 (R-5).

D7. The Rust runtime carries the evaluator, not only the types. Huginn evaluates selections over its own rows; if select lived only in C#, Rust consumers would each write one. packages/cultnet-rs/src/selection.rs holds the types, the cursor, and select over a Row trait, with no subscription client in this cut (Huginn does not subscribe).

D8. Numbers are addressable through the type the cache already declares; the cache gains accessors, not attributes. A comparison needs a numeric value the string getter cannot give. CultCache.cs:1053 holds the CLR MemberType and the catalog already persists TypeName in the canonical shape, so numeric-ness is declared; what is added is runtime: on each member, ValueOf(document) -> object? and NumberOf(document) -> string? (the exact canonical decimal, Q-J; was double? before 2026-09-22) (non-null exactly for the closed set of CLR numeric types), and the public member entry exposes IsNumeric derived from TypeName. No new attribute, no catalog-shape change, no schema hash moves, no migration. Rust mirrors it as Row::number(index) and RowSet::is_numeric(schema, index); the consumer decides from its own types.

Verified by source read, because everything downstream rests on it: CultDocumentMemberDescriptor.TypeName is set at CultCache.cs:474 from CultSchemaTypeNames.FromType(member.MemberType), which renders a CLR full name (System.Single, System.Nullable<System.Single>, System.Collections.Generic.Dictionary<GameCult.Caching.CultRecordRef<…>, System.Single>); ToCatalogEntry() :168 copies it into CultSchemaMemberCatalogEntry.TypeName :42; CompareSchemaShapes compares it inside CanonicalSchemaMember :605-629. Numeric-ness is declared and persisted today, and the claim holds.

Two corrections to the estimate. First, IsNumeric has two possible authorities — the CLR MemberType in hand and the TypeName string read back from a stored catalog. They must agree, or a schema validates one way locally and another way against a persisted catalog. The CLR set is the authority (a closed set: the eight integer types, float, double, decimal, and Nullable<T> over any of them); the string set is generated from it by CultSchemaTypeNames.FromType, never hand-written. S22 pins that.

Second, and larger: the accessor is not the only thing missing. Per section 1, IndexAccessors and Members are internal and GameCult.Networking is not a friend assembly, so the evaluator cannot read any declared value — string or number — without the reflection the map is deleting. The cache add is therefore a public read surface over declared members, of which the numeric accessor is one method: TryGetIndexValue(object document, string alias, out string? value), TryGetIndexNumber(object document, string alias, out string canonicalDecimal) (revised 2026-09-22 under Q-J: NumberOf returns the member's exact canonical decimal string, never double; see §2 "Numbers"; commit 0 shipped double and is corrected in the Q-J fix), ReferencesOf (D11), and a public member view carrying IndexAlias, IsReference, IsMany, IsNumeric and the target Type (D9 needs the type, not the name). Section 14 carries the revised cost.

D9. A reference's target set is resolved by the registry at evaluation, not persisted. Persisting TargetSchemaNames would change the canonical member shape and every stored catalog's hash for such schemas. Resolving at evaluation by assignability over registered leaves (:1506, :1828 do this today) costs nothing and is correct as of the registry in hand. The catalog's TargetSchemaName stays what it is (null for an abstract target) and is not read by the hop.

The resolution needs a Type, and CultDocumentMemberDescriptor carries only TargetSchemaName, a string that is null in exactly the case that matters. So the public member view exposes the target Type (held already on PersistedMember.MemberType / the attribute's TargetType, CultCache.cs:1069-1070), and the leaf set is AllDescriptors (public, :254-255) filtered by target.IsAssignableFrom(d.DocumentType) — the same predicate GetByIndex<T> uses at :1506 and store routing at :1828. It must be computed against the registry at evaluation, not at descriptor build: BuildDescriptor runs per type as it is registered, and a leaf registered later would be missing from a set frozen early. A cached set keyed by target type is legitimate only if it is invalidated by Refresh() (:257-265) and by RegisterDescriptor (:381-410); simplest is not to cache it.

D10. The duplicate-alias refusal is renamed, not invented. The Aetheria hiding case turned out to be refused already (section 1: HiddenMemberMessage at CultCache.cs:937-943), and two differently-named members sharing one alias already throw — from ToDictionary at :462-468, as an ArgumentException that names neither the schema, nor the members, nor the alias. D10 is to replace that with a rejection in DiscoverMembers' rejections list, in the shape of its neighbours: one grouping over IndexAlias beside the two at :931-943, one message beside DuplicateSlotMessage / HiddenMemberMessage. The behaviour barely changes; what changes is that the refusal can be read. Nothing in the selection depends on it — a colliding catalog never registers either way — so it is in this cut only because the cut is already in this file and the message is four lines.

D11. Many-references are enumerated by the cache, with payloads. On each reference member, ReferencesOf(document) -> IEnumerable<(CultRecordKey target, object? payload)>: a CultRecordRef<T> yields one; an IEnumerable<CultRecordRef<T>> yields each with no payload; an IDictionary<CultRecordRef<T>, V> yields each key with its value. Anything else declared many is refused at registration (unsupported_reference_shape), so a declaration the cache cannot walk is never silently walked as nothing.

The enumeration reads ICultRecordRef.Key (CultDocumentContracts.cs:92-95), which every CultRecordRef<T> implements (:97-99), so it needs no generic dispatch. Two shape rules fall out of ResolveReferenceTarget (CultCache.cs:1100-1113) and belong in the same refusal: a member declared many whose element type is not a CultRecordRef<> and whose attribute names no TargetType has no resolvable target at all (IsReference is true only because the attribute exists, and TargetSchemaName is null for a reason that is not inheritance); and a many member that is neither IEnumerable<CultRecordRef<T>> nor IDictionary<CultRecordRef<T>, V> cannot be walked. Both are unsupported_reference_shape at registration. A many: false member whose type is not CultRecordRef<T> is the same fault and takes the same refusal.

Agreement with our own abandoned prior art (ThreadEpiphanyGraphQuery, Epiphany 5c62a650, deleted with its host 5f6f2441): direction and edge kind as data, schema derived — agreed; differs as before (predicate not subgraph read; roles are declared names not free strings; one negation added; no depth field).

Rejected, unchanged in reason: Qdrant's Filter (open dot-path keys, polymorphic values, OR/threshold/recursion advertised then refused, the untagged hazard; Cut 11 lowers FieldPredicate to must/match.any in ~30 lines on the daemon side); ReQL term trees; Firestore CompositeFilter; Mongo/ES magic-key documents; every string-parsed evaluator surveyed.

4. Deletes first

Path:lines (at 51fb449)	Lines	What
src/GameCult.Networking/CultNetDatabaseServer.cs:347-468	122	CreateChangeMessage, CreateRawDocumentRecord, Matches, MatchesSchema, InferSchemaName — the LiteNetLib selector engine; the handler calls the shared evaluator and the shared raw-record projection
src/GameCult.Networking/CultNetDatabaseSubscriptionServer.cs:424-515	92	CreateAuthorizedChange, CreateChange, CreateChangeCore, ResolveWireSchemaId, MatchesRequestedSchema, CreateRawRecord — the RUDP selector engine
CultNetDatabaseSubscriptionServer.cs:27, :44, :50, :239-240, :314	6	_projectRecord, its parameter and its two call sites
CultNetDatabaseSubscriptionServer.cs:294-322	29	CreateProjectedSnapshot body: becomes evaluate + authorize + project (~12 lines stay)
src/GameCult.Networking/CultNetDocumentRegistry.cs:263-349	87	CreateRawSnapshotResponse: the loop and its reflection over cache.AllEntries (:325-342); the raw-record construction (:297-311) survives as ToRawRecord
CultNetDocumentRegistry.cs:505-524	20	MatchesRequestedSchema, a copy of CultNetSchemaAliasMatching.Matches(candidate, descriptor)
CultNetDocumentRegistry.cs:247-258	12	CreateSnapshotRequest (v0 builder); v1 takes a Selection
Networking total	368
src/GameCult.Mesh/CultMeshSnapshots.cs:638-661, 750-773, 889-911	71	three verbatim clones of the selector-carrying options; one with copy over a record type remains
src/GameCult.Mesh/CultMesh.cs:2721-2748, 1941-1969	57	the fourth and fifth copies (ToSnapshotRequestOptions, WithPublicationBindingOptions)
CultMeshSnapshots.cs:973-983	11	CleanSnapshotFilter: normalisation is the selection's own validation
CultMeshSnapshots.cs:1000-1009, 1033-1047, 1078-1088	36	the client-side re-filter of a snapshot by schema and InferSchemaName; the server's v1 answer is exact, decode stays
CultMesh.cs:2750-2800	51	ReadDocumentFromSnapshotResponse's four-tier fallback (which returns a record with the wrong key when only the schema matches) and TryDecodeSnapshotDocument; replaced by one exact read (~10 stay)
CultMesh.cs:3474-3497	24	IsSameCultDocumentSchema: CultNetSchemaAliasMatching, now public
CultMeshSnapshots.cs:625-636 and the rule at :198, CultMesh.cs:2716	~15	the "no schema filter when keys are given" rule, three times; one place
CultMeshClient.cs:574-604, 830-860, 932-962 selector construction	~25 net	{recordKeys, schemaIds} literals become Selection literals; the retry loops stay
CultMeshBodyDemand.cs:177-221	~15 net	CultMeshHotBodySubscription builds two allowlists; builds one Selection
CultMesh.cs:2701-2719, 2838-2858, 2875-2915	~20 net	selector construction for peer snapshots and body subscriptions
CultMeshSnapshots.cs:139-201, 434-516, 611-623 overlay logic	~40 net	overlaying caller schemaIds/recordKeys onto defaults, four times
Mesh total	≈ 365 gross, ≈ 331 net	see section 11 for what does not go
packages/cultnet-rs/src/snapshot_query.rs:100-130	~31	the fifth selector engine: the requested_schemas / requested_keys sets and the two is_some_and filters inside serve_read_only_raw_snapshot become one select over a lowered v0 selection. The duplicate-identity and unregistered-schema refusals (:90-110) and the whole CultNetReadOnlySnapshotPolicy (:132-157, authorization) stay
src/GameCult.Caching/CultCache.cs	0	the cache deletes nothing. The string Getter/GetterNullable pair (:1081-1082) stays: it is what any_of compares and what the index maps are built from (:1858). The numeric accessor is beside it, not instead of it
tests/GameCult.Networking.Tests/NetworkingTests.cs: CultNetDatabaseServer_Creates_Filtered_SnapshotResponse :3124, …_ForCompatibleSchemaAlias :3158, CultNetDatabaseServer_Creates_Filtered_SubscriptionChange :5547, …_ForSchemaAlias :5601, DatabaseSubscription_FiltersLiveChangesByWireSchemaBinding :2256, …ByRequestedSchemaAlias :2310, and the projectRecord lambda in …AuthorizesRequestAndFiltersSnapshotAndLiveRecords :1601-1605	~250	rewritten over v1 selections against the one evaluator (section 10); the rules they pin survive

Not deleted in this cut, by ruling (R-1), and scheduled for deletion on FU-v0's trigger: the v0 message classes and their three schema files, the v0 arms in CultNetSchemaMessageSerialization, and the v0 handling in the two servers — kept as a lowering: a v0 request lowers to Selection { schemas: schemaIds, keys: recordKeys, projection: document } and is answered through the same evaluator, so v0 has no engine of its own. They die with the last runtime follow-up.

5. Keeps and moves

• CultNetSchemaAliasMatching (CultNetDatabase.cs:47-97), internal → public, the one alias matcher for every runtime-side match.
• _authorizeRequest, _authorizeRecord, Reconcile, Watch, ApplyProjectedChange, PublishDemand, DemandChanged and the body-plane fields of the subscribe message: authorization, reconciliation and demand are not selection.
• CultNetDocumentRegistry's raw-record construction (:297-311) as ToRawRecord(descriptor, binding, key, document, options); the payload schema resolution (:452-503) untouched.
• The v0 messages, schemas and arms as a lowering (above).
• CultMeshSnapshotRequestOptions as a record { Selection, ShardId, ShardEpoch, …transport knobs }.
• Every (B)-class site of the audit (section 11): transport racing, authority, freshness, body routing, provider-session validation.
• The interop lanes and CI witness (.github/workflows/cultnet-interop.yml) green throughout, because v0 still answers.

6. Adds

Add	Owner	Live consumer	Protected invariant	What it replaces
src/GameCult.Caching/CultCache.cs: a public read surface over declared members — CultDocumentDescriptor.DeclaredMembers (a public view: MemberName, Slot, IndexAlias, IsName, IsReference, IsMany, IsNumeric, TargetType), TryGetIndexValue(document, alias, out string?), TryGetIndexNumber(document, alias, out string canonicalDecimal) (Q-J, revised 2026-09-22), ReferencesOf(document, member) (D11), ResolveTargetLeaves(targetType) over AllDescriptors (D9)	GameCult.Caching	the evaluator; CultNetDocumentRegistry; later any runtime-side reader	the cache owns addressing and its values; a reader outside the assembly reads declarations and values through one surface instead of reflecting over documents	the reflection loop at CultNetDocumentRegistry.cs:325-342, and the absence that forced it
CultCache.cs registration refusals: one alias-collision rejection beside :931-943 (D10), one unsupported_reference_shape rejection for a many/reference member the cache cannot walk (D11)	GameCult.Caching	every registering catalog	a declaration the cache cannot read is refused when it is declared, not answered as nothing when it is read	an ArgumentException from ToDictionary :462-468 that names nothing; a silent unwalkable reference
contracts/cultnet/cultnet.selection.schema.json ($id, referenced by $ref), cultnet.snapshot-request.v1.schema.json, cultnet.database-subscribe.v1.schema.json, cultnet.snapshot-response-raw.v1.schema.json; registry entries in CultNetSchemaRegistry.cs	CultLib contracts	C#, Rust; TS/Py/Kotlin follow-ups	one published shape of a selection, hand-written like its siblings, pinned by vectors in both runtimes	schemaIds/recordKeys/includeSnapshot as the wire's only selection
src/GameCult.Networking/CultNetSelection.cs: CultNetSelection, CultNetFieldPredicate, CultNetCitation, CultNetIncoming, CultNetSelectionProjection, CultNetRecordRef, CultNetRawDocumentHeader, CultNetSelectionPage, CultNetEdge (From, Role, To, PayloadEncoding?, Payload?), CultNetSelectionOperator (AnyOf, Lt, Le, Ge, Gt, serialised as the op string, never as a union); Validate(IReadOnlyList<CultDocumentDescriptor>) — declared aliases and roles, non-empty lists, exactly one of values/number per op, and every reachable leaf declaring a compared alias declares it numeric	GameCult.Networking	the three message classes, the evaluator, Mesh	a selection is typed data with one validation	_projectRecord, the allowlist pair
src/GameCult.Networking/CultNetSelectionEvaluator.cs: Select(cache, descriptors, selection, ordinals, asOf) → ordered, hopped, cursored, projected page; Matches(selection, descriptor, key, document) for a single change; the cursor mint/parse with cursor_stale / cursor_invalid	GameCult.Networking	CultNetDocumentRegistry (snapshot), both servers (changes), Reconcile	one evaluator; order, snapshot and hop have one owner	the three engines (section 4)
CultNetDatabase.LastWriteSequence(schemaId, key) (a per-key ordinal kept on apply, ~20 lines)	CultNetDatabase	the evaluator's order and cursor	a row's ordinal is its last commit's sequence	storedAt as the only time
v1 message classes beside the v0 ones; the v0→Selection lowering (one function)	GameCult.Networking	both servers	v0 has no engine of its own	—
packages/cultnet-rs/src/selection.rs: the types (Serialize, Deserialize, no untagged; SelectionOperator an externally tagged / string enum, Edge { from, role, to, payload_encoding, payload }), Row trait (id, ordinal, values(index) -> &[String], number(index) -> Option<f64>, references() -> Vec<(role, RecordRef, Option<Vec<u8>>)>), RowSet trait (declared_indexes, declared_roles, is_numeric(schema, index), target_leaves(role)), Malformed, select, cursor mint/parse, LIMIT_MAX	cultnet-rs	Huginn (section 12), later Rust services	the same evaluator semantics as the reference, pinned by vectors both ways	Huginn's own select.rs from the earlier design
snapshot_query.rs: serve_read_only_raw_snapshot calls select over a lowered v0 selection; the policy check stays where it is	cultnet-rs	its existing callers and tests	the fifth engine has no rule of its own	the filter at :112-125
contracts.rs: SnapshotRequestV1, DatabaseSubscribeV1, SnapshotResponseRawV1 variants	cultnet-rs	Huginn's daemon	the wire spelling equals the reference's	—
contracts/cultnet/selection-vectors.cs-written.json and selection-vectors.rs-written.json: a fixture row set (cultnet.interop-selection-row document type; declared indexes kind, severity, tags[] and mass (numeric, declared on the abstract middle type, not on the leaves); references parent, related[] and components (a dictionary keyed by reference with a float payload) and design (declared target = the abstract middle, resolving to two leaves); two leaf schemas under one abstract middle, so inheritance, the abstract target, the dictionary edge and the comparison are all exercised by the fixture rather than argued about in prose), twenty-eight selections, and each selection's expected page as ordered ids + matched + next presence + the expected edges with their payload bytes. The fixture must also carry, because tests depend on it and a loosening that cannot fail is the fixture's defect: a row whose mass equals a compared number exactly (S16), three rows sharing one index value (S23), one edge pointing outside its declared target (S18), two dictionary entries with different payload bytes (S19), and a nullable numeric member and a decimal one (S22). Expected to grow (R-4): it is chosen for coverage now, not pinned as final, and adding rows or selections to it is a normal change	tests, both runtimes	NetworkingTests, cultnet-rs/tests/selection.rs	vectors written by the reference decode and evaluate identically in Rust, and vectors written by Rust decode and evaluate identically in the reference	a within-runtime round trip, which pins nothing about parity
contracts/cultnet/interop/cultnet.interop-selection-row.schema.json	contracts	the vectors	—	—
docs/cultnet-selection-cut.md (this file), a paragraph in contracts/cultnet/cultnet-distributed-database.md under "Implemented", and the runtime table row in docs/runtime-parity-scope.md	docs	—	describe the live system	—

No dependency in either runtime (schemars is not added to cultnet-rs: the schema is hand-written JSON like every CultNet schema, and Huginn's derived schema for its own request embeds Selection through a #[schemars(schema_with)] field attribute in Huginn. The attribute returns a $ref to this cut's published selection schema $id. It does not go through the substrate. (Corrected 2026-09-22: a hand-written impl JsonSchema for a foreign type fails to compile with E0117, the orphan rule. The read-side Imagination probed it.) — see section 12). No new package, binary, transport, or store format. packages/cultcache-rs is untouched: the C# cache change buys the evaluator a read surface it does not have, and in Rust the row owner already supplies exactly that through Row/RowSet. The asymmetry is real and intended — the reference reads documents the cache owns, Rust reads rows the consumer owns — and it is the reason the parity obligation is on the vectors and not on the two implementations' shapes.

7. Per-file changes (code that exists, anchored; new code by name and rule)

src/GameCult.Networking/CultNetSchemaMessages.cs: after :826, the three v1 classes; CultNetSchemaVersions (:77, :93 and the SnapshotResponseRaw constant) gains SnapshotRequestV1, DatabaseSubscribeV1, SnapshotResponseRawV1. v0 classes untouched.

CultNetSchemaMessageSerialization.cs:9-85: three arms before the _ => throw.

CultNetSchemaRegistry.cs:360-385: four SchemaResourceSpec entries.

CultNetDatabase.cs: :47 internal static class → public static class; a Dictionary<(string, string), long> maintained where a mutation is appended to the shard log (Hands anchors on the append site; section 1's assumption is confirmed there) and LastWriteSequence; :1052-1075 untouched.

CultNetDocumentRegistry.cs: :247-258 → CreateSnapshotRequest(messageId, CultNetSelection); :263-349 → CreateRawSnapshotResponse(cache, messageId, selection, ordinals, asOf, options) calling the evaluator and ToRawRecord; :505-524 deleted.

CultNetDatabaseServer.cs: :276-316 handles v0 (lowered) and v1; :332-345 CreateSubscription calls evaluator.Matches(selection, change) then ToRawRecord; :347-468 deleted.

CultNetDatabaseSubscriptionServer.cs: :27, :44, :50 _projectRecord deleted; :112-167 handles v0 (lowered) and v1, and answers a snapshot_response_raw.v1 page with matched/asOf/next; :218-249 Watch: a selection with cites/cited runs Reconcile on every change, otherwise Matches on the changed row; :294-322 evaluate + authorize + project; :424-515 deleted. Module doc :14-18 rewritten: projection is a value on the request.

src/GameCult.Caching/CultCache.cs (the only cache file this cut touches):

• :112-150 CultDocumentDescriptor's constructor and fields: Members :149 and IndexAccessors :148 stay internal; a public DeclaredMembers view and the three value accessors are added beside them. Do not widen Members itself — CultDocumentMemberDescriptor is a mutable internal DTO whose shape ToCatalogEntry :151-176 depends on, and publishing a mutable descriptor list makes the catalog's shape a public contract by accident. The public view is a readonly struct or sealed record built from it.
• :445-490 BuildDescriptor: the alias ToDictionary at :462-468 is the site of the D10 refusal's symptom; the refusal itself belongs in DiscoverMembers so it lands in rejections with its neighbours.
• :838-946 DiscoverMembers: two new groupings beside the two at :931-943 — one over IndexAlias (D10), one over reference shape (D11) — and two message helpers beside DuplicateSlotMessage / HiddenMemberMessage. Nothing in the existing walk changes.
• :1050-1114 PersistedMember: Getter/GetterNullable :1081-1082 untouched; NumberOf (non-null exactly for the closed CLR numeric set, Nullable<T> unwrapped) and ReferencesOf added beside them; MemberType :1052 and the attribute's TargetType surface through the public view as the target Type.
• :254-255 AllDescriptors is the input to ResolveTargetLeaves; the set is computed per call, not cached, because Refresh() :257-265 and RegisterDescriptor :381-410 can both add a leaf after any given descriptor was built.
• CultSchemaTypeNames.FromType (CultSchemaTypeNames.cs:9-31) is internal and generates the TypeName strings; the IsNumeric string set is generated from the CLR set through it, never written by hand (S22).

packages/cultnet-rs/src/snapshot_query.rs: :100-130 the two allowlist sets and their filters become one select over the lowered selection; :64-99 and :132-157 untouched; :158-285 CultNetRawSnapshotQuery untouched in this cut and named in Q-M.

CultNetSelection.cs, CultNetSelectionEvaluator.cs (new): as section 6. Rules that must die under their own mutation are in section 10.

src/GameCult.Mesh/CultMeshSnapshots.cs: :16-62 the options record carries Selection; :139-201, :434-516, :611-636 overlay through one WithSelection; :638-661, :750-773, :889-911 deleted; :973-983, :1000-1009, :1033-1047, :1078-1088 deleted. CultMesh.cs: :1941-1969, :2721-2748, :2750-2800 (one exact read remains), :3474-3497 deleted; :2701-2719, :2838-2858, :2875-2915 build a Selection. CultMeshClient.cs:574-604, 830-860, 932-962, CultMeshBodyDemand.cs:177-221: Selection literals.

packages/cultnet-rs/src/selection.rs (new), contracts.rs:314-331 region gains the three v1 variants, lib.rs re-exports. tests/selection.rs (new): the vectors both directions, the evaluator, the cursor, the refusals, the untagged grep.

tests/GameCult.Networking.Tests/NetworkingTests.cs: the seven tests in section 4 rewritten; new tests in section 10; the vector writer under CULTNET_WRITE_VECTORS=1 following scripts/write-cultmesh-realtime-frame-vectors.mjs and the C#-written pattern the QUIC campaign established (docs/typescript-quic-realtime-cut.md:15-21).

8. Authority map

• Owner. CultNet's database organ, src/GameCult.Networking, as the reference: CultNetSelection owns the shape and its validation; CultNetSelectionEvaluator owns matching, the hop, the order, the snapshot and the cursor; CultNetSchemaAliasMatching owns schema identity; CultNetDatabase owns a row's ordinal. The cache owns addressing and every value of it: which names are indexes, which members are references, whether a member is numeric, what leaves a reference may target, what a row's value at an alias is, and what edges and payloads a row carries — through CultIndexAttribute and CultReferenceAttribute, the persisted TypeName, and the public read surface of section 6. The vocabulary reads those declarations and adds none: no new attribute, no persisted field, no schema-hash movement, and no opinion in the evaluator about what a CLR type means. The evaluator never reflects over a document; if it needs something the cache does not expose, the cache gains a method, not the evaluator a GetProperty. packages/cultnet-rs is the Rust spelling of the same owner, at wire parity, with the row owner supplying what the cache supplies in C# (Row, RowSet).
• Inputs. The cache's rows and descriptors, the shard log's sequences, the selection, the peer's authorization answers. Not inputs: a clock, a payload path, a transport.
• Outputs. Pages (headers or documents, matched, asOf, next), per-change match decisions, typed refusals (selection_invalid, cursor_stale, cursor_invalid) as cultnet.error.v0 with a code, the form the reference already uses for routing errors.
• Derived state. The page, the order, the incoming-edge index per evaluation, the cursor. Demotions: schemaIds/recordKeys are no longer a selection language; they are a lowering of v0 into one. _projectRecord is no longer an authority; projection is a request value. CreateRawSnapshotResponse, CreateChangeMessage and CreateChangeCore are no longer engines; they call one. IsSameCultDocumentSchema, the Mesh InferSchemaName and the registry's MatchesRequestedSchema are no longer owners of schema identity.
• Forbidden writers. Neither server may match a record by a rule of its own; serve_read_only_raw_snapshot may not filter by schema or key (its policy is authorization and stays; its allowlist filter goes); Mesh may not re-filter a page the server answered — a caller may verify a page against expectations it declared before the request (CultNetRawSnapshotQuery::accept_response, which errors rather than quietly dropping, and so cannot hide a server that answered wrongly), but no caller may narrow a page and call the narrowed result the answer; nothing may mint a cursor a caller could construct; nothing may answer a page over a different asOf than its cursor's; the evaluator may not read a payload by path, only declared indexes and references, and may not reflect over a document at all; the evaluator may not read the cache's _indexes map or call GetByIndex — those answer a unique-index lookup, not a predicate (section 1, S23); no rule may consult an edge's payload while traversing (section 13); nothing outside GameCult.Caching may decide whether a member is numeric or what leaves a target resolves to; no type on the wire is untagged; the Rust select may not diverge from the reference on any vector.
• Shared paths. The evaluator under snapshot, change, and reconcile; CultNetSchemaAliasMatching under every schema match in both assemblies; ToRawRecord under snapshot and change; the v0 lowering under both servers.
• Deletion line. Commit 0: the cache's public read surface and its two registration refusals, with S17-S22 green and GameCult.Networking untouched — the cache change is independently true and lands first, so that if the rest of the cut is cut, the substrate is not left with half a surface. Commit 1: the shape, the evaluator and the v1 messages beside v0, with the three engines rewired and deleted and the seven tests rewritten — green. Commit 2: _projectRecord and the Mesh collapse. Commit 3: Rust. Commit 4: vectors both ways and entries. Nothing lands with two engines alive.
• Where the line to CultMesh falls. Mesh is a client that builds selections and reads pages; it evaluates nothing. Where Mesh watches a cache locally by name or index (audit sites #2, #4-9, #16-21), it keeps its own loops until CultNetDatabase.Watch(Selection) exists, which is Cut 2 (section 18).

9. Parity: what this cut covers, and what the other runtimes do

In this cut: the C# reference (both servers, the client, the registry) and Rust (cultnet-rs: types, evaluator, cursor; no subscription client, because no Rust caller subscribes). Parity is pinned by vectors in both directions (section 10), not by a round trip inside one runtime.

Runtimes not in this cut, stated answer per runtime, by source read:

Runtime	On v0 (unchanged)	On a v1 request it receives	On a v1 page it receives
TypeScript	unaffected; the reference still answers v0	refuses: parseCultNetMessage finds no validator for the version and throws (contracts.ts:728-744)	refuses the same way
Kotlin	unaffected (client only)	refuses: require(schemaVersion == …v0) throws (CultMesh.kt:2195)	n/a as a server; a Kotlin client sends v0 and gets v0
Python	unaffected	silently drops, no reply: handle_server_message returns [] for an unknown version (interop_peer.py:387), so a caller times out rather than reading a refusal; it cannot answer with wrong rows, because v1 is a version it does not dispatch, which is why D3 chose a version over new fields	as above
Rust	in this cut	in this cut	in this cut

The one runtime that mis-answers today is none: nothing sends v1 to a runtime that does not speak it until its follow-up lands, and a stray v1 is refused (TS, Kotlin, C#) or dropped (Python). Python's drop is fixed in this cut, not in its follow-up (R-2): after this cut the table's Python row reads "refuses with unsupported_schema_version", and every runtime on the wire answers an unknown version with something a caller can read. The build budget gains one Python test run for it.

Follow-ups, recorded with triggers:

• FU-TS. cultnet-ts: the three v1 contracts, Ajv registration, a Selection type; cultmesh-ts builders. Trigger: the first TypeScript caller that needs a filtered snapshot or subscription (the browser Eve runtime, when it reads pipeline state). Until then TS refuses v1 loudly.
• FU-Py. Its first line is not a follow-up: by R-2, cultnet-py and cultmesh-py reply cultnet.error.v0 { code: "unsupported_schema_version" } for an unknown version in this cut (five lines at interop_peer.py:360-387 and the delegating path at cultmesh_py/server.py:341, with one test per site that a v1 message produces a refusal and not an empty list). The rest of FU-Py is the v1 contracts, and the three copies of the selector match (interop_peer.py:110-125, cultmesh_py/server.py:449-470, cultmesh_py/node.py:1322-1345) collapsed into one evaluator. Trigger: the first Python peer asked a v1 question, or the interop lane gaining a v1 case.
• FU-Kt. cultmesh-kotlin: v1 snapshot request builder and page reader. Trigger: an Android/JVM client reading filtered state.
• FU-v0. Retire v0's classes, schemas and lowering once FU-TS, FU-Py, FU-Kt have landed and the interop lanes send v1. Trigger: the last of the three.
• Not a follow-up any more: CultNetDatabase.Watch(Selection) is Cut 2, mapped in section 18 with its own deletes, authority map, verification and numbers. It lands after this cut. The audit's ≈ 240 local-watch lines are counted there, and the real figure turned out larger than the audit's estimate (≈ −378 in Mesh, section 18.3), because the audit did not count the nine Collection* overloads as selection.

docs/runtime-parity-scope.md's table gains, per runtime, "typed selection v1: C# and Rust claimed; TS, Python, Kotlin not claimed (refuse / refuse / drop)". The claim is written where the parity claims live, not here alone.

10. Verification

Builds. dotnet build of GameCult.Caching, GameCult.Networking, GameCult.Mesh and their test projects — the cache is now in the build budget, and it is the assembly the rest of CultLib depends on, so dotnet test tests/GameCult.Caching.Tests is the gate on commit 0 and runs on every later commit; cargo test -p cultnet-rs; the TypeScript workspace untouched but npm run test --workspace packages/cultnet-ts run once to prove the interop lane still passes on v0. Host = target = workstation for C# and Rust; no native code. Nothing cleaned.

Tests, each named for the rule it pins, with the mutation that kills it — a loosening where one exists.

#	Test (C# unless marked)	Pins	Revert kills	Loosening kills
S1	Selection_RefusesUndeclaredIndexRoleAndEmptyLists	D2, door refusals typed	validation dropped (empty page)	index checked against any schema's aliases rather than the reachable ones (a predicate on an alias only an excluded schema declares matches nothing silently)
S2	Evaluator_ConjoinsAnyOfPredicatesOverDeclaredIndexes	fields	an arm dropped	any-of treated as all-of
S3	Evaluator_OrdersByLastWriteSequenceThenIdentityAndReverses	order, descending	sort by storedAt	descending ignored
S4	Evaluator_PagesExactlyOnceAndTheLastPageSaysSo	cursor walk	next always present	position >= (last item repeats)
S5	Evaluator_RefusesAStaleCursorAfterTheShardLogAdvances	cursor_stale typed	advance ignored (a page over a moved set)	digest check dropped (a cursor accepted under another selection)
S6	Evaluator_HopsOneEdgeByDeclaredReferenceAndRole	hop, roles from CultReferenceAttribute, Many	role ignored	a second hop followed
S7	Evaluator_NegatedIncomingEdgeIsTheOnlyNegation	cited { exists: false }	exists ignored	citer's status consulted (an unauthorized citer's edge dropped)
S8	Evaluator_HeaderProjectionCarriesNoPayload	projection	payload carried	payloadEncoding carried without payload
S9	Subscription_EvaluatesSelectionOnSnapshotAndEveryChange_AndReconcilesHopBearingOnes (rewrite of :1587)	D6	a change matched by v0 allowlists	a hop-bearing subscription taking the per-change fast path (a change to the citer not reconciled)
S10	V0Request_LowersToASelectionAndHasNoEngineOfItsOwn (rewrites of :3124, :3158, :5547, :5601, :2256, :2310)	D3/D4	v0 matched by its own loop	alias matching inlined again in one server
S11	SchemaAliasMatching_IsTheOneMatcherInBothAssemblies (grep-shaped, negative)	D4	—	—
S12	SelectionVectors_WrittenByTheReference_DecodeAndEvaluateIdenticallyInRust (Rust reads selection-vectors.cs-written.json) and SelectionVectors_WrittenByRust_DecodeAndEvaluateIdenticallyInTheReference (C# reads selection-vectors.rs-written.json)	parity	any semantic drift on either side	a vector whose expected ids are compared as a set (order lost) — the vectors carry ordered ids and next presence
S13 (Rust)	selection_is_never_untagged (grep over packages/cultnet-rs/src for untagged, empty) and selection_round_trips_named_and_positional (both encodings)	the encoding rule	an untagged shape added	—
S14 (Rust)	select_over_a_toy_row_set_matches_orders_hops_pages_and_refuses	D7, no CultCache knowledge in the evaluator	as S1-S8	as S1-S8
S15	Mesh_ReadsASnapshotPageWithoutReFilteringIt (Mesh tests)	the client trusts the page	the four-tier fallback restored	schema re-filter restored
S16	Evaluator_ComparesNumbersAtTheBoundaryForEachOfTheFourOperators	D8's operators; one operator per predicate; two predicates make a range	the comparison ignored (every row matches)	le evaluated as lt (and ge as gt) — kills only if the fixture carries a row whose mass equals the compared number exactly; the fixture must, and if it does not, that is the fixture's defect, not licence to mutate something easier
S17	Evaluator_AppliesAnInheritedAliasToEveryReachableLeaf	the inheritance rule; the cache's BaseType walk reaching the evaluator	the walk stopped at the leaf (CultCache.cs:847) — kills S17 and S12 together	"reachable" narrowed from every leaf matching schemas to the first descriptor consulted, so a two-leaf selection on an alias declared only on the abstract middle answers with one leaf's rows and no refusal — requires the fixture to declare mass only on the middle and to carry rows of both leaves
S18	Evaluator_ResolvesAnAbstractTargetToItsLeavesAndRefusesAnEdgeOutsideThem	D9; the typed reference_outside_target	the target set taken as {TargetSchemaName}, which is null for an abstract target, so the hop follows nothing	the out-of-target edge skipped instead of refused — the option section 3 rejected, and the mutant that proves the rejection is load-bearing; requires the fixture to carry one edge pointing at a schema outside the declared target
S19	Cache_EnumeratesADictionaryReferenceAsEdgesCarryingItsValues (cache tests)	D11	ReferencesOf yields nothing for a dictionary — today's behaviour, restored	the keys enumerated and the values dropped (payload null), which is the shape a careless IEnumerable<ICultRecordRef> implementation would have; requires two dictionary entries with different payload bytes, or the assertion passes on shape alone
S20	Evaluator_ReturnsTraversedEdgesOnlyForHoppingSelections_AndNoPayloadUnderHeader	the edges rule	edges never emitted	payload carried under header; and, separately, edges emitted for a selection with neither cites nor cited — a widening rather than a loosening, and killed by the same test
S21	Cache_RefusesADuplicateIndexAliasAndAnUnwalkableReferenceByName (cache tests)	D10, D11's shape rule	the rejections removed — kills only if the test asserts the message names the schema, the two members and the alias, since without them ToDictionary still throws an ArgumentException and a shallow Assert.Throws passes either way. That is the entire content of D10	the alias grouping keyed by MemberName instead of IndexAlias, which refuses nothing the existing HiddenMemberMessage (:937-943) did not already refuse
S23	Evaluator_MatchesEveryRowSharingAnIndexValue_NotTheCacheIndexWinner	the evaluator reads values per row, never _indexes / GetByIndex	the predicate answered from GetByIndex, so an any_of over a value three fixture rows share answers with one	the per-row read kept but short-circuited after the first match when matched is all the caller asked for — a plausible optimisation that makes matched a boolean in disguise; requires three rows sharing one index value, which the fixture must carry
S22	Cache_DerivesIsNumericFromTheClrSetAndTheTypeNameSetAgrees (cache tests)	D8's two authorities agreeing	IsNumeric true for every member, so a comparison on a string alias validates and compares ToString() lexically	the TypeName-string set hand-written and missing System.Nullable<System.Single> or System.Decimal, so a member is numeric against the live descriptor and non-numeric against the same schema read back from a persisted catalog — requires the fixture to declare a nullable numeric member and a decimal one

On the parity vectors and the four findings. S12 is not a new test but a widened one: the fixture of section 6 carries the abstract middle, the two leaves, the inherited numeric alias, the dictionary reference with float payloads and the abstract-target reference, and the vector set includes at least one selection per finding — a comparison on the inherited alias, a hop into the abstract target, a hop over the dictionary returning payloads, and the same hop under header. Both directions: vectors written by the reference are evaluated in Rust, and vectors written by Rust are evaluated by the reference. A round trip inside one runtime pins nothing and is not counted. The Rust side's fixture rows implement Row over the same declared values, so the vectors test the semantics both ways while each runtime keeps its own source of rows.

Negative greps: rg -n "GetProperty|GetField|GetCustomAttribute" src/GameCult.Networking empty (the evaluator reflects over nothing); rg -n "untagged" packages/cultnet-rs/src packages/cultcache-rs/src empty; rg -n "InferSchemaName" src/GameCult.Mesh empty and src/GameCult.Networking only in CultNetSchemaAliasMatching and the registry's payload-schema resolution; rg -n "MatchesRequestedSchema|MatchesSchema\(" src empty; rg -n "_projectRecord|projectRecord" src empty; rg -n "SchemaIds = .*RecordKeys = " src/GameCult.Mesh empty (no allowlist pair assembled by hand); rg -n "CloneSnapshot(Facade)?RequestOptions" src at most one.

Mutations. tools-style entries in scripts/mutate-cultmesh.mjs's format with three new targets, caching (killer dotnet test tests/GameCult.Caching.Tests, carrying S19, S21 and S22), networking (killer dotnet test tests/GameCult.Networking.Tests) and rust (killer cargo test -p cultnet-rs), one entry per rule above with revert and loosening; the runner's byte-rewrite/verified-restore discipline applies unchanged. The runner is the QUIC campaign's file and is dirty in the tree; this campaign adds its entries in its own file (scripts/mutate-cultnet-selection.mjs or an entries module the runner imports) and does not edit the QUIC campaign's targets — Self decides the exact split at landing.

11. The CultMesh subtraction audit, with real numbers

Method: an Explore pass classified 92 sites in src/GameCult.Mesh and CultNetDocumentRegistry.cs as (A) mechanical selection/projection over schema ids, keys or fields, or (B) a domain rule (authority, freshness, transport, session). (A) totalled 1,158 lines, (B) 429. (A) is not a reduction figure; it is what could be expressed as a selection. What collapses is what has a replacement in this cut.

Category	Sites (audit #)	Lines now	After	Net
Selector-carrying options cloned five times	#11, #3, #74, #75, #77	128	~30 (with on a record)	−98
"No schema filter when keys are given" rule, three times	#65, #73, #10	~16	~4	−12
CleanSnapshotFilter	#79	11	0	−11
Client-side re-filter of a snapshot by schema, plus InferSchemaName	#80 (part), #81, #82	~51	~10 (decode only)	−41
Four-tier read fallback + decode gate	#12, #13	50	~10	−40
IsSameCultDocumentSchema	#22	24	0 (public matcher)	−24
Selector literals in client bindings	#32, #34, #35	93	~68	−25
Hot-body subscription allowlists	#25	45	~30	−15
Peer snapshot / body subscribe selector construction	#10, #14, #15	81	~61	−20
Snapshot facade overlay logic	#64, #66-70, #72	~110	~70	−40
Collapses in this cut				≈ −326
Local cache/database watches by name and index and their predicates	#2, #4-9, #16-21	≈ 240	≈ 240	0 in this cut; ≈ −378 in Cut 2 (section 18.3 — larger than this audit estimated, because the nine Collection* overloads at CultMesh.cs:2278-2480 were not classified as selection and are)
Handle catalogs by type/schema, surface catalog, operation payload	#41-49	233	233	0 — these index handles, not records
Verse discovery selection and reshaping	#36, #38 (part), #39, #40, #83, #85	~110	~110	0 — over Verse observations, not documents
Single-file catalog membership	#58-62	81	81	0 — file-format membership
Provider session wire decode gate	#50	12	12	0
(B) sites	24 sites	429	429	0 by definition

Plain statement. src/GameCult.Mesh shrinks by about 326 lines of 21,720 (1.5%) in this cut, and a further ≈ 378 in Cut 2 — the two together are about 3.2% of the assembly, and Cut 2 is where the audit's expectation of a real reduction is met. About 590 of the audit's (A) lines stay because they select handles, Verse observations or file-catalog entries, not database records. The operator's expectation of a net reduction in CultMesh holds, modestly; it does not hold for CultLib as a whole in this cut, because the Rust runtime is new capability (section 14). If the expectation was a large reduction, this audit says it is not there without the watch cut, and it is not there at all for the catalogs and discovery.

12. Huginn: the consumer cut, and what moved

Moves to CultLib (this map): Selection, Citation, Incoming, Projection, Page, the cursor, select, the Row/RowSet traits, the untagged rule and its probe, the cap, the order rule and the snapshot contract. Huginn's earlier select.rs is never written.

Stays with the organ (a smaller, later cut in the Epiphany campaign map, against a cultnet-rs dependency Huginn already has):

• impl Row for PipelineDocumentView: ordinal = admission.sequence (Q-B A, still recommended; under Q-B B Huginn refuses cursor_stale on advance exactly as the reference does, which is now a legitimate second option), values(index) for the declared aliases campaign, repo, cut, kinds→(schema), in_force, faculty, severity, confidence, origin, authority, claim_outcome, outcome; references() from docs::citations with the fifteen roles; RowSet::declared_* as the organ's closed lists, so an unknown alias or a value outside an enum refuses typed.
• docs::citations, CitationRole, ResolutionOutcomeKind, the summary as the organ's header, the as_of restriction of the image to receipts, the semantic refusal, the ref validation of cites.target through the leaf, and the deletion of open_items/history with the relation table unchanged (history(Subject) = kinds: [resolution], cites { target, role: subject }; the open lists = five selections).
• Deleted from the organ's wire (R-3, ruled, not proposed): PipelineQuery; admitted_after and admitted_before, which existed only to approximate an order and are replaced by descending plus the cursor — they are a delete in the Huginn cut's own deletes-first table, with their query-builder call sites, their schema fields and the tests that pin them, not a deprecation; and QUERY_LIMIT_MAX. Huginn's Query { instance, selection } carries the substrate's type; its published request schema embeds Selection through a #[schemars(schema_with)] attribute that returns a $ref to the published selection schema. A hand-written impl fails with E0117 (corrected 2026-09-22), and the substrate publishes JSON, not schemars. It is pinned by the existing byte-for-byte test.
• Cut 11's lowering to Qdrant and the after-index application of the hop, as before.

The Huginn cut is ordered after this cut's commit 3 (Rust) and before Cut 11 and Cut 13, as 10b was.

13. Deliberately not possible, and why

OR across predicates (ask twice); negation beyond cited { exists: false }; a predicate on the far end of the hop; a second hop; ranges, dates, text, paths, sort fields, counts beyond matched; a page not tied to asOf; selection across shards or minds; a cursor the caller constructs. Each for the reason the earlier specification gave, unchanged.

Named individually, because each has an obvious-looking one-line implementation and a caller who will ask for it:

Not expressible	The obvious ask	Why it is the door
A predicate on an edge's payload during traversal — "components whose quantity is over 4", "demand profile entries weighted above 0.5"	one if inside the hop loop, over a payload the evaluator is already holding	It is a join condition. The moment a predicate can read a value that belongs to neither the row nor the selection, the evaluator is combining two sources under a condition, and everything that follows is a query planner: which side to drive from, whether to build the incoming-edge index before or after filtering, how matched counts a row whose only surviving edge was filtered, what the cursor means when the page is a join result, and whether the payload's type is comparable at all when its declared type differs per leaf. None of that is answerable in a bounded vocabulary. The caller follows the edge, receives the payload in edges, and filters the result — one extra step for the caller, an entire category of machinery this organ never grows
A set of exact (schema, key) pairs	keys already exists; let an entry name its schema	schemas × keys is a cross product and stays one. CultNetRawSnapshotQuery (section 1) shows the cost honestly: it over-requests and verifies. The fix is either a second list shape or a pair type on the wire, and both are the beginning of a term language. A caller who needs exact pairs asks once per schema, or accepts the cross product and verifies against what it declared. Recorded as Q-M, recommendation A: state the limit
A predicate that depends on how many edges a row has — "hulls with more than three components"	cited already counts to one	exists: true/false is a boolean, and a threshold is a range over a derived aggregate. Aggregation is a second evaluator with its own ordering and paging semantics
edges without rows, or rows the hop excluded	a projection: "edges"	the page's rows are the answer and the edges explain it; an edge-only page is a different question (a graph read) and would need its own order, cursor and bound

14. Subtraction ledger (estimate)

	Removed	Added
src/GameCult.Caching	0	~150: the public member view ~40, TryGetIndexValue / TryGetIndexNumber + the CLR numeric set ~35, ReferencesOf over the three shapes ~40, ResolveTargetLeaves ~10, two rejections with their messages ~20
src/GameCult.Networking	368	~400 (CultNetSelection + CultNetEdge + the operator 170, evaluator + cursor + the incoming-edge index 190, ordinals 20, v1 classes and lowering ~40 minus shared)
src/GameCult.Mesh	≈ 365 gross	≈ 40 (one WithSelection, one exact read)
packages/cultnet-rs/src	~31 (snapshot_query.rs's filter)	~430 (selection.rs 390 with edges and numbers, contracts.rs 40)
packages/cultnet-py, cultmesh-py	0	~5 (the unsupported_schema_version reply, ruled into this cut)
contracts/cultnet	0	4 schema files (~200 lines JSON), 1 interop row schema, 2 vector files (generated)
tests, C#	~250 rewritten	~600 (S1-S12, S15-S18, S20 in Networking; S19, S21, S22 in Caching; the vector writer)
tests, Rust	0	~340 (S12-S14 with the edge and number vectors)
entries	0	~38 entries across three targets
docs	—	this map; two paragraphs
net source outside tests		≈ +32 in C# Networking, +150 in C# Caching, +399 in Rust; ≈ +585 across the repo

The estimate moved, and in the wrong direction. The committed version of this ledger said ≈ +370 and ≈ −18 in C#. Both were wrong, for one reason: the map assumed the cache needed no change, and the cache needs about 150 lines of new public surface before the evaluator can read a single declared value from outside the assembly (section 1). The honest figure is ≈ +585. What the repo buys for it is unchanged in kind and larger in size: one evaluator where five stood (three in C#, one in Mesh's re-filter, one in Rust), a read surface where the substrate's own consumers were reflecting over documents, edges and their payloads where a declared reference was unreadable, numeric comparison over a type the catalog has persisted all along, and a Rust runtime with the capability at all. Nothing here is a reason to shrink the cut; it is a reason not to claim the cut shrinks the repo.

The repo grows by about 370 source lines for: a typed selection with a derived-by-vector schema where two allowlists were the only language; projection as a value where it was a delegate no caller reached; an exact snapshot cursor where paging did not exist; one evaluator where three engines and six schema matchers stood; a Rust runtime that did not have the capability at all. CultMesh proper shrinks ≈ 326. The liability retired is not in the line count: five copies of one options record, four spellings of schema identity, a client that re-filtered what the server answered and a read that could return the wrong key.

15. Build budget

C#: three assemblies and three test projects (GameCult.Caching joins the budget; it is a dependency of the other two, so touching it rebuilds them), debug, workstation. Rust: cultnet-rs lib + tests, warm target dir, +0 to +60 paths. Nothing cleaned. No native, no TypeScript build beyond one interop-lane run.

16. Operator questions, one batch

Ruled 2026-09-17 and no longer questions. Four of the batch are settled; they are recorded here as rulings and carried in the sections they govern.

• R-1 (was Q-F). v0 retires on a named trigger, not "eventually." The reference keeps answering v0 by lowering it into a Selection, with no engine of its own. That lowering is a scheduled deletion: it is deleted when FU-TS, FU-Py and FU-Kt have all landed and the interop lanes send v1, and that is FU-v0's trigger in section 9. A shim with no deletion date is a permanent shim; this one has a date written as a condition, and the condition is checkable by grep over the three runtimes' contract files. Section 4's "not deleted in this cut" row and section 5's keeps both point here.
• R-2 (was the second half of Q-F). Python's silent drop is fixed in this cut. Every other runtime refuses a message it does not understand; Python returns [] (interop_peer.py:387) and the caller reads a hang rather than a rejection. The five lines that reply cultnet.error.v0 { code: "unsupported_schema_version" } land here, ahead of FU-Py's own trigger, because a silence is a worse failure than a missing feature. Carried in section 9's FU-Py entry and section 14's ledger.
• R-3 (was Q-G). descending is the order; the time windows go. admittedAfter / admittedBefore existed to approximate an order the vocabulary now states outright, and two filters standing in for an order are worse than the order. They are deleted from Huginn's wire in the consumer cut (section 12), not kept as an organ-side alias any-of. "The latest N" is descending, limit: N.
• R-4 (was Q-I). The fixture is chosen for coverage and expected to grow. The vector row type is the dedicated cultnet.interop-selection-row family of section 6, shaped to exercise the four hostile findings rather than describe them: shared fields on an abstract middle with two leaves under it, a reference whose declared target is that middle, and a many-reference carrying a payload. It is not pinned as final — the fixture is expected to gain rows and selections as the vocabulary meets more catalogs, and a change to it is a normal change, not a contract break. What is pinned is the parity obligation over whatever the fixture holds: both directions, ordered ids, matched, next presence, and the edges with their payloads.
• R-5 (was Q-H). Watching a selection is its own cut, mapped in section 18 of this document, after this one and depending on it. Nothing of it is smuggled into the selection cut and nothing here is designed around it speculatively; the seam already drawn — one selector type serving both a query and a subscription — is all this cut owes it.

Still open:

• Q-A. Leaf Short titles for question and ruling (Huginn/Epiphany). Unchanged: A, not now (recommended); B, a leaf follow-up before Cut 13; C, a text projection (Cut 11's question). Depends: the organ's header row for those two kinds.

• Q-B. Huginn's receipt ordinal. A, record it (recommended; exact order, no cursor_stale on an append-only mind); B, do not: Huginn refuses cursor_stale on advance like the reference does — now a legitimate option under one contract, costing a re-walk on a busy mind and id-order ties within a second. Depends: Huginn's consumer cut only.

• Q-C. Delete open_items and history. A, delete (recommended); B, thin presets in Huginn; C, wire operations (refused). Depends: Cut 13's tool list.

• Q-D. The representation. A, declared-alias any-of over strings in a new v1 selection (recommended; D1, D2); B, Qdrant's Filter (rejected, section 3); C, typed enums on the wire (impossible in the substrate: it cannot know a consumer's enums; the organ keeps the typed refusal at its door). Depends: everything in sections 2-10.

• Q-E. Withdrawn: the seam question is answered by the placement ruling.

• Q-J. Number encoding on the wire. RULED B, decimal strings, 2026-09-17. The operator: floats are fragile, and strings are the most reliable way to store a number. Self's recommendation of A was argued backwards and the map said so wrongly. B was framed as "a second value encoding beside any_of's strings"; it is the opposite. Every value in this vocabulary is already a string, so decimal strings keep the wire to one value encoding, and it is float64 that would add a second. What B really costs is a comparison rule, and that is where the danger sits: a decimal string compared lexicographically answers that ten is less than nine. So the cut must specify a canonical decimal form and a numeric comparison, and pin both with mutants — a lexicographic comparison must die, and a non-canonical spelling of an equal value must not change an answer. Name the edge cases the canonical form settles: leading and trailing zeros, an explicit plus, exponent notation, and negative zero. No exactness limit needs stating, because there is no longer one.

  • Superseded, kept as history: A, IEEE float64, with the 2^53 limit stated in the schema and in this map (recommended: one encoding, no parse rule per runtime, and no catalog in hand declares an integer member wide enough to lose). B, decimal strings — exact, but a second value encoding beside any_of's strings and a per-runtime parse rule, and two ways to spell a number is how a vocabulary starts growing terms. C, add integer: int64 beside number now, sharing op — additive and untagged-free, and can be added later without breaking A, which is why adding it now is premature. Depends: the schema files and both runtimes' predicate types.

• Q-K. Family names for abstract bases on the wire. PARKED, 2026-09-17, and Self's reason was wrong. The operator: this was possible when CultCache was a single file, so it is odd that it is hard now; not load bearing either way. That is correct. In-process a family is one assignability check, because the runtime holds the hierarchy. Over the wire there is no type system, only schema names, and an abstract type carries no [CultDocument] so it never gets one. The hardness is not modelling, it is that nothing publishes the hierarchy where a remote caller can see it. The honest cost is therefore a catalog addition, each leaf's entry naming its base chain, after which a family is the leaves whose chain contains the name asked for. Small work — but it moves the catalog's shape, and this cut's whole claim to being cheap is that it moves neither the catalog's shape nor its hashes. So it is deferred to a cut that is already touching the catalog, not deferred because there is nothing to name.

  Trigger named, 2026-09-17: the source-generator campaign needs it first. The operator's position is that any project using CultNet wants the types it touches native, and that where the hierarchy is not being declared in parallel by hand, a generator should build it. That generator runs against published schemas, with C# authoring and the schema as the contract, which is what the reference-runtime rule already says.

  The obstacle is the same missing fact. CultCache.cs:847,887,966 walks the base types while collecting members and then discards the hierarchy, and ToCatalogEntry (:152-177) publishes a schema id, the canonical schema, the compatible ids and a flat member list carrying slot, name, type, reference-ness, many-ness, target and alias. Nothing records what a type inherits from. So a generator fed published schemas alone emits one flat class per leaf with the inherited members copied into each — exactly the hierarchy the operator wants built for them, and unbuildable from what is published today.

  So the base chain in the catalog is a prerequisite of the generator campaign, and family names fall out of it for free once it exists. Until then a family expands at the call site into its leaf schema names, which the generator can emit as a constant, costing the protocol nothing. The honest limit of call-site expansion, to be stated wherever it lands: it is closed-world per compilation, so a leaf declared where the caller does not compile leaves the list silently short, and an open-world caller must ask the catalog at run time instead.

  The generator is its own campaign, mapped after this cut lands (operator, 2026-09-17: "definitely later"), because its value is emitting against a shape that is settled and this cut is what settles it. It is also where the query sugar comes from: typed per-kind builders make filtering a document by another kind's field a compile error rather than an empty page. No source generator exists in CultLib today — no Roslyn generator anywhere in the repository, and src/GameCult.Caching/emitted/ is empty.

  Superseded reasoning, kept as history: an abstract type has no [CultDocument], no schema name and no catalog entry, so there is nothing to name; a caller who wants "every crafted item" enumerates the leaves from the schema catalog it already reads. B, a [CultDocumentFamily] declaration plus a catalog concept plus a matching rule in every runtime, in a later cut. Note this is only about the schemas list: the hop already resolves an abstract target to its leaves (D9), so the hostile example works under A.

• Q-L. Refusing an out-of-target reference at write time. RULED A, 2026-09-17, with the follow-up upgraded from optional to intended. The operator: notice on read for now, but on write eventually; today this is left to consumers to enforce, as Aetheria's Database Tools did and as CultCache Studio is expected to, and it is easy to insert garbage at the source level which a cheap check would prevent.

  Two things make the follow-up worth more than "nice to have." First, enforcement living in consumers is duplicated authority by construction: every tool that writes records re-implements the same check, and a tool that forgets is indistinguishable from one that does not, until a read refuses. Second, the check really is cheap — the declaration already names the target (TargetSchemaName on the member catalog entry), and the write path already has the record in hand.

  What Self could establish about the consumer story, 2026-09-17: it rests on less than it sounds. There is no CultCache Studio repository under F:\Projects, and no project file by that name anywhere; every reference to it is forward-looking, in Aetheria's migration documents, where it is described as absorbing that project's Database Tools. Aetheria's actual data tooling today is the tools/AetherDb console project. So the enforcement being relied on is one tool, in one project, for one catalog, and there is no second implementation to be inconsistent with yet — which is the cheapest moment to move the rule to its owner rather than the most expensive.

  Original options: A, follow-up (FU-Ref) (recommended): this cut refuses it at read (reference_outside_target, S18), which is where the corruption is noticed; refusing at write touches the cache's put path and every runtime's apply path and is a change of a different size. B, in this cut. Under A, the read-side refusal is the only thing standing between a corrupt edge and a wrong answer, which is why it refuses instead of skipping.

• Q-M. The (schema, key) cross product. RULED A, 2026-09-17, in the operator's words: once per schema. The cross product stays, the limit is stated, and a caller needing exact pairs asks once per schema. No third addressing list and no rule for combining three of them. Original options: A, state the limit (recommended): schemas × keys stays a cross product; a caller needing exact pairs asks once per schema or verifies what it declared, as CultNetRawSnapshotQuery already does. B, allow a refs: RecordRef[] list beside keys — expressible without a term language and genuinely useful, at the cost of a third addressing list and a rule for how three lists combine. C, let a keys entry carry a schema — a path language in miniature, refused. Depends: section 13's table and snapshot_query.rs's client.

Q-A, Q-B and Q-C are no longer open. The operator ruled all three on 2026-09-17 and they are recorded in the Huginn campaign's map, F:\Projects\Epiphany\notes\eureka-pipeline-state-cut.md: receipt ordinals yes; open_items and history deleted rather than kept as presets; and the leaf gains titles on question and ruling, riding the same store version bump as the ordinal so there is one migration rather than two. They are named below only because work here still depends on them.

What depends on each: S1-S8 on Q-D A; S5 on the shard-log assumption (section 1) and on Q-B for Huginn; S9 on D6; S16 and the schema files on Q-J; S18 on Q-L A; the Huginn cut on Q-B and Q-C. Settled by R-1 to R-5: S10, S12, the fixture, FU-v0's trigger, FU-Py's five lines, Huginn's window deletion, and the watch cut's existence.

17. What was probed, read, and not settled

Probed: the untagged round trip (section 2). Read: every line anchored above; the CultMesh audit's 92 sites by an Explore pass whose classification I checked at the sites that carry the ledger (the five clones, the four matchers, the fallback, the two engines). Not run: dotnet or cargo on CultLib — the tree is dirty from the QUIC campaign and no claim here depends on a build; Hands takes the baselines named in Pins. Not settled: whether every committed write has a shard-log sequence (section 1, Hands confirms at the append site); the exact split of mutation entries against the QUIC campaign's runner (section 10, Self at landing); whether MessagePack-CSharp needs a union attribute for the page's headers | documents (it does not if both are optional arrays with exactly one present, which is the spelling chosen so no union is needed — Hands confirms the schema says "exactly one" with oneOf over required).

Read on the second pass, and what it changed. Each of these was read at the source, and three of them contradicted something this map or its predecessor's notes asserted:

Read	Held	Changed
CultCache.cs:474, :151-176, :598-629, CultSchemaTypeNames.cs:9-31	Yes. A member's type name is derived from the CLR type, persisted in the catalog entry and compared inside the canonical shape. Numeric addressing needs no new declaration, no hash movement and no migration — the load-bearing claim of this whole widening	—
CultCache.cs:112-150	—	No. Members and IndexAccessors are internal with no friend assembly, so the evaluator cannot read a declared value from outside GameCult.Caching at all. "No cache change is needed for addressing" was false; the cache add is ~150 lines, not ~60, and the ledger moved from ≈ +370 to ≈ +585
CultCache.cs:931-943, :462-468	—	No. A hiding member is already refused (HiddenMemberMessage), and a duplicate alias already throws from ToDictionary. D10 shrank from "a missing refusal" to "a refusal that names nothing", and its test only means anything if it asserts the message
packages/cultnet-rs/src/snapshot_query.rs	—	No. Rust has a fifth selector engine (serve_read_only_raw_snapshot:112-125) that "Rust: none" concealed, and its client (:195-217) shows the cross-product limit that v1 does not fix (Q-M)
CultDocumentContracts.cs:92-99, CultCache.cs:1100-1113, :1495-1509, :1818-1838, :254-265	Yes. ICultRecordRef is the enumerable handle D11 needs; assignability is already how the cache resolves a base type to concrete documents (D9); AllDescriptors is public	—

Still not settled, and who settles it:

• Whether every committed write takes a shard-log sequence (unchanged; Hands, at the append site).
• Whether CultDocumentDescriptor's public member view can be a readonly struct without forcing an allocation per row per predicate in the evaluator's inner loop. Nothing here depends on the answer being fast, but a Select-per-row over a fresh list would be a silly cost to build in; Hands measures if a fixture of a few thousand rows is slow, and not before.
• The exact split of mutation entries against the QUIC campaign's runner (unchanged; Self at landing).
• Whether serve_read_only_raw_snapshot's existing tests survive the lowering unchanged; they should, since the behaviour is identical, but they were not read line by line.
• Q-J, Q-K, Q-L, Q-M: the operator.

Not probed, and deliberately: nothing was built or run against CultLib. The tree is hosting a live QUIC campaign, and no claim in this map needs a build to be true — every one of them is a source read, and the three that were wrong were wrong because the first pass did not read far enough, not because a build would have caught them.

---

Cut 2: watching a selection

Ruled 2026-09-17 (R-5): the operator wants this and wants it soon, so it is mapped here rather than left as a follow-up trigger. It is its own cut, landing after Cut 1 and depending on it. Nothing below is smuggled into Cut 1, and Cut 1 is not shaped around it: the seam Cut 1 draws — one selector type serving a query and a subscription alike — is the whole of what Cut 1 owes this one. Cut 1 can go to Hands without waiting for a ruling on anything here.

Same discipline as Cut 1: facts by source read, deletes first, an authority map, verification with a mutation per rule including a loosening, and a subtraction estimate with real numbers.

18.1 Body facts

The local watch surface today (src/GameCult.Networking/CultNetDatabase.cs):

• Watch<T>() :1051-1057 filters the change stream by CLR type; WatchAllChanges() :1063-1067; WatchRecord<T>(key) :1072-1075; WatchGlobal<T>() :1080-1083; WatchByName<T>(name) :1088-1094; WatchByIndex<T>(alias, value) :1100-1107.
• WatchByName and WatchByIndex do not do what they are named. Both read the cache's current single winner for the lookup and compare it to the change's document by ReferenceEquals. Combined with section 1's finding that the cache's index maps hold one key per value, this means WatchByIndex<Weapon>("kind", "cannon") observes changes to whichever single Weapon most recently claimed cannon, not to the weapons whose kind is cannon. It also re-runs the lookup at change time, so a change that clears the value is evaluated against the state after the clear. The same two lines exist again in CultMesh as MatchesName / MatchesIndex (CultMesh.cs:3312-3333).
• The rule is set-dependent even without a hop, and nothing tracks membership: there is no "was in the set, now is not" anywhere in the local path. Reconcile (CultNetDatabaseSubscriptionServer.cs:251-292) is the only membership tracker in the reference and it belongs to a server, over a delivered set, per peer.

The CultMesh collection surface (src/GameCult.Mesh/CultMesh.cs): nine Collection* overloads at :2278-2480 — three shapes (all / by name / by index) by three sources (CultCache :2278, :2302, :2331; CultNetDatabase :2362, :2386, :2413; CultMeshNode :2442, :2456, :2471, which delegate to the database trio in three lines each). CollectionByIndex over a cache resolves its contents as Optional(cache.GetByIndex<TDocument>(alias, value)) — a "collection" of at most one row, whose live feed then filters by MatchesIndex. Three shapes exist because the shape is the query language; a selection replaces the shape axis entirely and leaves the source axis alone.

What Cut 1 leaves in place for this cut: CultNetSelection and its validation, CultNetSelectionEvaluator.Matches(selection, descriptor, key, document) for a single row, the cache's public read surface, and D6's rule that a hop-bearing selection is set-dependent and must reconcile rather than decide per change.

18.2 The shape

CultNetDatabase.Watch(CultNetSelection selection)
  -> Observable<CultNetSelectionChange>

CultNetSelectionChange {
  kind:      Entered | Left | Updated     // membership, not storage
  schemaId, recordKey
  document:  object?                      // present on Entered/Updated under `document` projection
  previous:  object?                      // present on Left/Updated
  cause:     Self | Edge                  // this row changed, or a row it cites / is cited by did
}

Rules, each one owner:

• Membership, not mutation. The stream reports a row entering the selection, leaving it, or changing while inside it. A write to a row that was outside and stays outside emits nothing. A write that changes a value the selection filters on emits Entered or Left, not Updated — which is precisely what WatchByIndex cannot express today.
• cause: Edge is how a hop-bearing selection stays honest: when a change to row B moves row A in or out, the subscriber is told about A, with the cause marked, and never about B unless B is itself in the set.
• No order, no cursor, no asOf, no limit. A live feed is not a page. A selection carrying limit, cursor or descending is refused at the door (selection_invalid { field }) rather than silently ignored: "the latest ten, live" is a different question with a different answer (it needs an eviction rule), and this cut does not answer it. projection is honoured; schemas, keys, fields, cites and cited are honoured.
• The initial set is not a snapshot message. Watch is local and in-process; it emits Entered for every row already matching, at subscribe time, under the cache's lock, before any live change. There is no asOf because there is no wire and no second reader.
• A hop-bearing selection recomputes; a plain one decides per row. Without cites/cited, membership is Matches(selection, row) and costs one row's read. With a hop, a change anywhere can move anything, so the watch re-evaluates the selection over the cache and diffs against the membership set it holds. That is O(rows) per change and it is stated plainly rather than hidden (Q-N).

18.3 Deletes first

Path:lines	Lines	What
CultNetDatabase.cs:1080-1107	28	WatchGlobal, WatchByName, WatchByIndex — the three that encode a query in a method name, two of them wrongly. WatchGlobal lowers to keys: ["global:{schemaId}"], the other two to a fields predicate
CultMesh.cs:3312-3333	22	MatchesName, MatchesIndex — the same ReferenceEquals-against-the-winner rule, second copy
CultMesh.cs:2278-2480	203	the nine Collection* overloads; after: three, one per source (CultCache, CultNetDatabase, CultMeshNode), each taking a CultNetSelection, about 75 lines
CultMeshSnapshots.cs and CultMesh.cs local-watch predicates, audit sites #2, #4-9, #16-21	~240	the remaining local watch loops and their inline predicates, about 40 after
Total	~493	~115 after; net about −378

Watch<T>(), WatchAllChanges() and WatchRecord<T>(key) stay, as Cut 1's v0 lowering stays: a CLR-type filter and a single-key watch are conveniences that lower to a selection and have no rule of their own. If they grow a rule again, W6 catches it.

18.4 Adds and authority

Add	Owner	Live consumer	Protected invariant	Replaces
CultNetSelectionWatch in GameCult.Networking (~130 lines): the membership set, the per-change decision, the hop recompute, CultNetSelectionChange	CultNetDatabase	Mesh's three Collection overloads; any local consumer that watches a question	membership is tracked in one place, and a row that leaves the set is reported leaving	WatchByName / WatchByIndex / MatchesName / MatchesIndex, four spellings of one wrong rule
CultNetDatabase.Watch(CultNetSelection) (~15 lines)	CultNetDatabase	the above	one entry point	three named-query methods
Rust: nothing in this cut	—	—	—	no Rust caller watches locally; when one does, it gets the same shape over Row/RowSet

• Owner. CultNetSelectionWatch owns membership and only membership. CultNetSelectionEvaluator.Matches owns whether a row is in the set — the watch never re-implements a predicate. CultNetDatabase owns the change stream. The cache owns values, as in Cut 1.
• Inputs. The change stream, the cache's rows through the public read surface, the validated selection. Not inputs: a clock, a transport, the shard log's sequences (there is no order here), a peer's authorization.
• Outputs. Entered / Left / Updated with a cause, and one typed refusal for a selection carrying page-shaped fields.
• Derived state. The membership set, and the incoming-edge index when the selection hops. Demotions: Watch<T> and WatchRecord<T> are no longer queries, they are lowerings; Mesh's collection shape is no longer a language, it is a Selection; MatchesIndex is not an owner of anything.
• Forbidden writers. Nothing outside CultNetSelectionWatch may decide membership; no watch path may call GetByIndex or GetByName to decide whether a change is interesting (that is the bug being deleted, in both its copies); Mesh may not filter a feed the watch answered; the watch may not emit Updated for a row that crossed the predicate boundary; the watch may not accept limit, cursor or descending.
• Shared paths. Matches under both the query and the watch — a row that a page would return and a row the watch reports entering are decided by the same function, or the two truths drift. The subscription server's Reconcile keeps its own per-peer delivered set (it answers a different question: what this peer has been told) and does not become the local membership tracker.
• Deletion line. Commit 1: CultNetSelectionWatch and Watch(Selection), with the three named-query methods deleted and their tests rewritten. Commit 2: Mesh's nine overloads collapsed to three, both matchers deleted. Nothing lands with WatchByIndex alive beside it.

18.5 Verification

#	Test	Pins	Revert kills	Loosening kills
W1	Watch_ReportsARowLeavingWhenItsFilteredValueChanges	membership, not mutation — the bug being deleted	Left never emitted (today's behaviour restored: the change is simply not delivered)	Updated emitted instead of Left, so a subscriber holding a list keeps a row that no longer matches
W2	Watch_ReportsEveryRowSharingAnIndexValue_NotTheCacheIndexWinner	the deletion of the ReferenceEquals-against-the-winner rule	MatchesIndex restored	membership decided by the winner plus the changed row, which passes a two-row fixture and fails a three-row one — the fixture must carry three rows sharing the value
W3	Watch_EmitsTheInitialSetBeforeAnyLiveChange	the subscribe-time set	no initial emission	the initial set emitted after the subscription goes live, so a write racing the subscribe is delivered twice or not at all — needs a test that writes between subscribe and first emission
W4	Watch_ReportsAnEdgeCausedEntryWhenTheCiterChanges	cause: Edge, D6 locally	a hop-bearing watch decided per changed row (the citee never moves)	the citee reported with cause: Self, which loses the subscriber's ability to tell a row's own change from a set change
W5	Watch_RefusesAPageShapedSelection	no order, cursor or limit on a feed	the refusal dropped and the fields ignored	limit refused but descending ignored — one field's worth of silent ignoring, which is how a "harmless" ignored field becomes a bug report
W6	Watch_AndSelect_AgreeOnMembershipForTheSameSelection (property-shaped: for each fixture selection, the set the watch holds equals the ids a page returns)	the shared Matches	the watch given its own predicate	the watch's copy agreeing on the plain selections and differing on the hop-bearing ones — kills only if the fixture's selections include hops, which section 6's fixture does
W7	Mesh_CollectionTakesASelectionAndFiltersNothingItself (Mesh tests)	the collapse	an overload restored	the three surviving overloads re-filtering the feed after the watch answered

Negative greps: rg -n "WatchByIndex|WatchByName|MatchesIndex|MatchesName" src empty; rg -n "GetByIndex|GetByName" src/GameCult.Mesh src/GameCult.Networking empty; rg -n "CollectionByIndex|CollectionByName" src empty.

Mutations. Entries in this campaign's own file, target networking (killer dotnet test tests/GameCult.Networking.Tests) plus the Mesh tests for W7, one per rule above with its revert and its loosening.

18.6 Subtraction estimate

	Removed	Added
src/GameCult.Networking	28	~145
src/GameCult.Mesh	~465	~115
tests	~80 rewritten	~260 (W1-W7)
net source outside tests		about −233

This is the cut where CultMesh actually shrinks: about −350 in CultMesh.cs and the snapshot facade against about +117 in the database organ. It is also the cut that deletes a wrong answer rather than a duplicate one — four spellings of "watch the row that currently wins a unique-index lookup", replaced by one membership tracker over the selection the caller actually asked about. Cut 1 pays about +585 for the vocabulary; Cut 2 spends it.

18.7 Operator questions for this cut

• Q-N. The hop-bearing watch's recompute. A, recompute and diff on every change (recommended: obviously correct, O(rows) per change, and the local caches this runs over hold thousands of rows, not millions). B, maintain an incremental incoming-edge index invalidated per write — faster, and a second piece of derived state that can disagree with the cache. Take B only with a measured reason.
• Q-O. Does WatchRecord<T> survive? A, yes, as a lowering (recommended; it is one key and no rule). B, delete it too and make every caller say keys: [k] — cleaner, noisier at its call sites.
• Q-P. Ordering on a feed, later. Out of this cut by the rules above. If "the latest N, live" is wanted, it is a third cut with an eviction rule, not a field added here.

R-AA landed, 2026-09-22 (Sonnet), b7d77b7..99e366a on cultnet/selection-cut1. One new test file, 27 tests, 707 lines. No production code touched. Score 66% → 77.62%: killed 345→371, survived 85→57, no-coverage 75→50, over the same 428-mutant set both passes. Suite green on Yggdrasil, 271 passed / 2 skipped.

Every one of the nine named survivor groups is killed, with two corrections to the triage that matter more than the score.

• EdgesFor: three of five killed. STRUCK 2026-09-23 — what stood here was wrong. It claimed that within one hop direction (From, Role) determines a unique declared edge value, so no legal selection could make two edges share anchor, From and Role while To differs, and recorded To.SchemaId and To.Key as equivalent "so nobody pays for them a third time". A selection carrying both hops mixes their edges into one bucket, and Soul built the counterexample: a row self-referencing through the role it uses to cite a peer. To.Key is now killed in both runtimes; To.SchemaId is settled under R-BA. The text is struck rather than edited because it was read as a live instruction for two passes.
• A sixth mutant in the same method that nobody had named: OrderBy(pageRowIndex), which is R-B's primary key, ahead of (from, role, to). It survived the first pass and is now killed by a test that forces page order and (from, role, to) order to disagree. The rule R-B exists to protect had its first component unguarded, and the triage had not caught that either.
• The digest needed a golden value, not a comparison. Beyond the named branch mutants, about nineteen finer-grained ones survived — deleting individual AppendString calls. A same-versus-different digest test structurally cannot catch these: a mutant that swaps which case gets which flag still produces two different digests, merely mislabelled. Hands reconstructed a golden HMAC independently from the method's doc comments rather than from its private helpers, which is the right way to build one, and it killed all but three residual coincidences.

Line drift: every rule had moved 30–90 lines from Soul's 52082ec numbers, none to a different method. Locate by name, not by number, and do not trust the numbers in this section either.

Four things found along the way, flagged rather than silently fixed. Two are test gaps; two look like real defects:

• ReadLengthPrefixedFields, length <= 0: a legitimately zero-length field would be wrongly refused.
• ReferenceMembers's !member.IsReference guard: a data member could be treated as citeable.
• FindCursorPosition's loop bound: an out-of-bounds read when paging past the last row with a cursor exactly at the end.
• ComputeDigest's else-branch appends survive only because the golden test has both Cites and Cited non-null; a second golden with both absent closes them.

Self's rulings, 2026-09-22:

• R-AC. STRUCK 2026-09-23. It said the two EdgesFor To mutants were equivalent and "recorded, not to be chased". That instruction was false and was obeyed for two passes. Soul's counterexample kills To.Key in both runtimes; To.SchemaId is settled under R-BA. The sentence it closed with — a survivor correctly identified beats a test written to move a number — is still true, and is exactly why a survivor must be identified by a construction attempt rather than by an argument.
• R-AD. The three suspected defects above get behaviour tests first, since a test that fails today is the only thing that proves they are defects rather than readings. The out-of-bounds read in FindCursorPosition is the one to check first: it is reachable from an attacker-supplied cursor.
• R-AE. A second golden digest with both hops absent, closing the three residual mutants rather than leaving them as folklore.

Fix batch 4, Rust half, landed (Sonnet), cultnet/selection-cut1 now at d7f73af, rebased onto the C# side's 99e366a. Six commits: c91e937 (R-V and Rust's half of R-W), 97627dd (R-Y), 4bfd989 and 94c32ac (R-AB), 8005586 (S-12, prose), d7f73af (a vendored-schema resync).

217 Rust tests green on Yggdrasil, and re-confirmed locally with contracts/ entirely hidden, which simulates cargo-mutants' isolated copy. Mutants scoped to the diff: 23 tested, 23 caught, 0 missed. Nothing to triage.

S-2 is C#-only — Rust already checked the declared target before filtering by key, so the divergence was one-sided and the C# commits close it. No Rust change was needed and none was made. S-1 and S-3 are closed by R-V and R-W.

R-AB worked, and it took two fixes rather than one. The named escape was schema_discovery.rs's include_str!("../../../contracts/…"). Vendoring byte-identical copies under packages/cultnet-rs/contracts/ fixed that and left the crate still unmutable: tests/selection.rs's fixture loader read the same canonical tree three levels up, and 37 of 45 tests panicked under the isolated copy. The first full-crate run failed its unmutated baseline for exactly that reason. Vendoring the fixture too, plus a skip-when-absent guard on the test that reads C#'s own cs-written.json output — which cannot be vendored, being the other runtime's product rather than a checked-in input — made the crate mutable for the first time.

The vendoring is duplicate authority, and it drifted within hours. The rebase onto the C# side's landed R-X changed cultnet.error.schema.json, and the vendored copy did not see it. Hands' own drift test caught it and the copy was re-synced at d7f73af.

That is worth stating plainly rather than filing as a win. Twenty-two schemas now exist twice, and the guard against the copies diverging is a test rather than a structure. The cleaner shape — a build script staging the canonical contracts into OUT_DIR — fails for the same reason the include_str! did: cargo-mutants copies the package tree, so a build.rs reaching three levels up is just as unreachable. Publishing the contracts as their own crate would fix it properly and is out of this cut's scope. So the duplicate is accepted for now, on the record, with the drift test as its named guard. It goes in the follow-ups, not the win column.

Hands burned an unscoped full-crate mutants run — 2,329 mutants and a timeout cascade across files the cut never touched — before rescoping to the diff. It reported that unprompted and discarded the output. The ruling did say "the cut's diff"; the run cost most of a night on a shared host.

Soul, the merge gate, 2026-09-23

Opus, against d7f73af. Counts re-run by Soul itself: Networking 269 passed / 2 skipped, Mesh 255, cultnet-rs 224 across 12 targets. (R-AA reported "271 passed" — that is 269 + 2 skipped. The Rust half reported 217; the measured figure is 224. Both green, one overstated.)

Verdict: Cut 1 may not merge. Three reasons, any one sufficient.

SM-1, high. S-1 is not closed. One ruling became two different rules.

CultNetSelectionEvaluator.cs:358-386 resolves a reference's target by record key plus the declared leaf set — its own comment says "a reference's stored target is a bare record key, the wire carries no schema for it". packages/cultnet-rs/src/selection.rs:1765 and :1817 resolve it by an exact by_key.get(&(target.schema_id, target.record_key)) on the RecordRef's own schema id. R-V said both runtimes index rows and incoming edges by the pair. C# indexes rows by the pair and edges by the key; Rust indexes both by the pair.

Probed at the evaluator in both runtimes, same logical rows:

Shape	C#	Rust
Only row at key k is out-of-target; narrow ref declared to leaf_a points at k	REFUSED reference_outside_target	ACCEPTED, empty page
Two rows at key k; ref declared to leaf_a, stored target names leaf_b	ACCEPTED, resolves the leaf_a row, emits an edge	REFUSED reference_outside_target

They give opposite answers on the exact shared-key shape R-V was written for. A consumer paging the same data through the two runtimes gets a refusal from one and rows from the other. Parity is this cut's invariant.

No vector covers it: the shared fixture has zero duplicate record keys across its ten rows, exactly one reference, and that reference's role targets both leaves — so reference_outside_target is never reached cross-runtime at all. R-V's narrower promise does hold: within each runtime, both row orders agree.

SM-2, high. Both runtimes take a hostile cursor into an unhandled fault, on mirror-image cases.

• C#. ReadLengthPrefixedFields (CultNetSelectionEvaluator.cs:750-751) tests pos + length > bytes.Length in int. A prefix of 2147483647 overflows, the guard passes, and Encoding.UTF8.GetString throws ArgumentOutOfRangeException. HandleSnapshotRequestV1Async catches only the three typed selection exceptions and CultNetRudpSchemaServer.DispatchAsync has no catch at all, so an untyped exception reaches the server's poll loop. The peer gets no answer.
• Rust. read_length_prefixed_fields (selection.rs:2044-2046) parses the prefix as usize and computes pos + len; 18446744073709551615 overflows. Probed: panic.

Rust refuses the int::MAX case typed; C# refuses the usize::MAX case typed. Each runtime has exactly the hole the other closed — which is what happens when a rule is hardened twice independently and neither run fed it garbage. R-N's whole point was that every cursor refusal reaches the wire as cursor_invalid.

SM-3, high. R-Z did not land, and nobody said so — including me.

R-Z required the vectors to compare page bytes, plus asOf, next and a refusal's code, with projection vectors on both sides. git diff --stat 52082ec..d7f73af touches neither SelectionParityVectorTests.cs nor either vector file's vectors. The reader still compares row ids, matched, hasNext and edge fields, and expectedRefusalField — never the code. There is no page-bytes field in either JSON. rs-written.json still carries no projection vector, which is S-7's original wording verbatim. The C#-side projection vectors run through Select, which ignores Projection, so the comment claiming "page bytes under each projection" describes something the test does not do.

R-Z is absent from the batch's work and from its open list. Self recorded the batch as closing its rulings without checking that one, which is the tracking failure that let it pass silently.

SM-4, medium. Two ordering rules are pinned by their mutants, not their behaviour.

Twelve non-revert mutations, each changing a rule a different way from the Stryker mutant its R-AA test was written against. Nine died. Three survived:

• EdgesFor's secondary key. Permuting Role ahead of From.Key (:307-311), keeping every component, passes 269/269. R-AA's test kills the primary key's deletion; the composition of the rest is unpinned, and the fixture's single reference pins nothing either.
• The row tiebreak. Swapping SchemaId and Key.Value in the ascending tiebreak (:214) passes 269/269. The fixture does have one tied ordinal, but both rows are leaf_a, so the schemaId component of (ordinal, schemaId, recordKey) is exercised by nothing — not in unit tests, not in vectors. Rust's order_rows uses the same three components, so a silent permutation on either side goes unnoticed.
• !member.IsReference (:456) is a redundant second copy of the guard in CultCache.ResolveReferences/ReferencesOf, which already returns empty for a data member. Equivalent mutant, and duplicate authority — not a defect, but not innocent either.

SM-5, medium. The Rust half of the vocabulary has 59 live mutants and nobody has looked.

The diff-scoped run ("23 tested, 23 caught, 0 missed") measures only the lines batch 4 changed. Soul ran cargo-mutants over the whole of selection.rs: 351 mutants, 274 caught, 59 missed, 14 unviable, 4 timeouts, 36 minutes.

• canonical_number: about 35 missed plus all four timeouts. is_canonical — the Q-J door — has every branch mutation surviving. decompose, pad_right, canonicalize_decimal_digits, big_mul_small and big_to_most_significant_first can each be replaced with String::new() or "xyzzy" and the suite stays green. This is R-D, a byte-for-byte wire parity rule.
• validate: 4 missed, including a deleted ! and || → && — the door R-F put inside select.
• matches_schema_keys_fields: 3 missed, including two deleted ! and an == → != — which rows match.
• schema_alias::matches: == → != — R-E's one identity rule.

S-5 blocked this cut at 66% on the C# side and produced R-AA. The Rust side of the same rules has never been measured. The asymmetry is the finding.

Settled, not defects

• FindCursorPosition's loop bound is not a defect. It bounds at i < ordered.Count. Probed in both runtimes: paging to the exact end and replaying the last cursor gives an empty page, next = null, no exception, no panic. The Stryker survivor was a coverage gap. R-AD's headline suspicion was wrong, and settling it was worth the pass.
• length <= 0 does not exist in either runtime, and a zero-length field parses fine in both. The real bug in that function is SM-2's overflow.
• S-2 held, and Hands' "C#-only" claim is true — Soul checked rather than accepted it. Both Rust call sites run ensure_within_declared_target before any role or key filtering, with a test pinning it.
• S-3 held. A both-hops selection returns all its cites edges in both runtimes, each anchored by its own hop, identically.
• S-12 is half done. The fixture's prose was fixed; SelectionParityVectorTests.cs:200-208 still calls the name-alias case a "Confirmed cross-runtime defect" that the alias port already fixed. Low, but it is a falsehood in a file a future reader will trust.

Promises that held

S-4 (real CultNetErrorMessage bytes validated against the schema for all four codes), S-6 (a spliced digest and a reused cursor under a changed selection both refused), S-9, S-10, S-11 (the hop refusal is gone and the hop is served; still no dedicated regression test). R-AB's guard is real: all 23 vendored files are byte-identical right now, the hand list matches every include_str! 22/22, and flipping one byte makes the drift test fail by name. R-AA's tests are mostly behavioural: nine of twelve hand mutations died, including the digest under a double-append rather than a deletion, the comparator forced back to UTF-16 units, both role-naming sites under inverted precedence, and cursor positioning made inclusive.

Caching's three failures remain unproven, not passing — the stopgap's container runs as root and the tests need a write to a read-only directory to fail. Soul declined to call them environmental on its own authority.

What Soul could not run

• No page-byte parity comparison exists to run. C#'s page is a CultNetSnapshotResponseRawV1Message envelope; Rust's is a bare SelectionPage. There is no common wire type, so R-Z as written needs a shape decision before it can be implemented at all.
• No fresh Stryker run on the C# side; twelve hand mutations instead, which is what the brief asked for. The Rust figure is Soul's own measurement.
• Only selection.rs was mutated in Rust; contracts.rs, schema_discovery.rs and the rest of the crate are unmeasured.

The rig, handed over

C:\Users\Meta\eureka-soul-probes\cultnet-selection-cut1\ — deliberately outside the repo and outside the session scratchpad, so it does not evaporate the way the Idunn rig did. README.md says where each file drops in and what it proves: SoulProbeTests.cs, soul_probe.rs, soul-mutants.sh and twelve precomputed mutated files (the dotnet image has neither python3 nor perl to patch with).

Self's rulings for fix batch 5, 2026-09-23

• R-AF (SM-1). One rule, and the wire decides which. Hands establishes empirically what a stored reference actually carries — a bare record key, or a key and a schema id. If the wire carries no schema, C#'s rule is right and Rust adopts it. If it carries both, Rust's is right and C# adopts it. I am not choosing by taste between two implementations when the Body already knows the answer; what I am ruling is that the answer is one rule, derived from the wire, in both runtimes. A parity vector must cover the shared-key shape in both directions — its absence is why this survived four passes.
• R-AG (SM-2). Every malformed cursor becomes cursor_invalid on the wire, in both runtimes, and nothing else escapes. Use checked arithmetic on both sides. C#'s dispatch path gains a catch so an untyped exception cannot reach the poll loop. Feed each runtime the other's proven overflow, plus a small set of deliberately hostile cursors, and pin them. The shape of this bug — each runtime holding exactly the hole the other closed — means neither side's tests are evidence about the other.
• R-AH (SM-3). R-Z needs its shape settled before it can be built. The envelope types differ, so the vectors compare the canonical MessagePack bytes of the page payload — rows, edges, asOf, next — which both runtimes can produce, rather than the transport envelope, which only one has. Refusals compare the code. Projection vectors exist on both sides, and Select must actually honour Projection or the vector is a lie. Flag for operator review: this is my ruling on a wire-shape question, and if the payload's canonical form is meant to be something else, it should be corrected before the vectors harden around it.
• R-AI (SM-5). Measure the Rust vocabulary and triage it. Run cargo-mutants across selection.rs and the rest of the crate's selection path, and close survivors by behaviour. canonical_number first: R-D is a byte-for-byte parity rule whose door currently survives every branch mutation, which means the canonical decimal form is pinned by nothing in Rust. Equivalents get a one-line reason.
• R-AJ (SM-4). Pin the ordering compositions, in both runtimes and in the vectors: EdgesFor's secondary key ordering, and the schemaId component of the row tiebreak, which needs a fixture with a tied ordinal across two schemas — the current fixture's tie is within one leaf.
• R-AK. Delete the evaluator's redundant !member.IsReference guard. CultCache owns that decision and already enforces it. Duplicate authority is the defect even when the duplicate is harmless today.
• S-12's remainder: correct the comment that still calls a fixed defect confirmed.

Fix batch 5 landed, 2026-09-23

Sonnet, five commits on sel-fix5. C# Networking 274 passed / 2 skipped (was 269/2), Mesh 255, Rust 50/50 with cargo-mutants --in-diff reporting 26 tested, 21 caught, 5 unviable, 0 missed.

R-AF: the wire carries no schema, so C#'s rule was the right one. Established by probe, not by reading either comment: a real Design reference field serialized through MessagePackSerializer dumps as A4-61-2D-65-71 — a bare five-byte msgpack string, "a-eq". The whole document is ["citer-1","a-eq"]. No schema id anywhere. CultRecordRefFormatter<T>.Serialize confirms it: writer.Write(value.Key.Value), unconditionally, for every CultRecordRef<T> in the codebase.

Rust adopted C#'s rule. Row::references() returns a bare target_record_key; by_key is grouped by record key alone; a new resolve_reference_target mirrors TryResolveReferenceTarget exactly — search every row sharing the key, return the first inside the declared leaf set, refuse deterministically by schema id if none is. RecordRef survives only where the wire genuinely carries both, which is the Citation.target query filter.

The vector that was missing now exists. The fixture gained a citer_narrow schema with a real SHA-256 id and rows for both shapes from Soul's table: a lone out-of-target row, which refuses, and a shared key with one in-target and one out-of-target row, which accepts and resolves the in-target one regardless of row order. Plus a new expectedEvaluationRefusal vector field — narrower than R-AH's job, just which typed refusal fires.

A second real defect, found while landing R-AF. Rust's matches_citation resolved and target-checked every declared reference on a row before filtering by the citation's requested role; C# filters by role first. Rust's role filter moved ahead of resolution to match. One existing Rust test was therefore pinning behaviour C# does not have, and was rewritten to mirror C#'s real shape. This was a live divergence nobody had named — neither Soul nor any ruling.

R-AG. C#'s pos + length overflowed in unchecked int; fixed by promoting to long. Rust's pos + len overflowed in usize; fixed with checked_add. CultNetRudpSchemaServer.DispatchAsync gained a backstop catch so no untyped exception reaches the poll loop. Each runtime was fed the other's proven overflow, plus three further hostile shapes, four new tests a side, each confirmed to fail without the fix.

R-AJ, and the finding inside it: the existing row-tiebreak test was degenerate. Hands hand-reverted the tiebreak and swapped it, and the test still passed 100% — purely by accident of which of SelLeafA/SelLeafB's real hash ids happened to sort first. It was rewritten to assign keys relative to the schema ids discovered at runtime, then re-probed and now fails correctly. The same shape was added in Rust, with a shared-fixture vector carrying a tied ordinal across two schemas — the original fixture only ever tied rows of the same schema. EdgesFor's secondary key got the same treatment.

R-AK. Verified before deleting: CultCache.ReferencesOf already searches only its own reference members for a role match, and index aliases are unique per descriptor, so a data member's role string cannot collide with a reference member's. The redundant guard is gone.

S-12 is closed; the stale "confirmed cross-runtime defect" comment is corrected against source.

Not done, and named rather than buried:

• No diff-scoped Stryker run. --since crashed on GitInfoProvider in the detached-clone container, so Hands fell back to whole-file globs: 312 tested, 262 killed, 47 survived, 3 timeouts, untriaged between the changed lines and the large pre-existing bulk of both files. Every rule Hands touched was separately confirmed by hand. The tooling failure was worked around, not root-caused, which means the next pass inherits it.
• The RUDP dispatch backstop has no dedicated integration test. It would need a fuller server harness. Flagged, not built.

R-AI landed, 2026-09-23 — canonical_number measured and closed

Sonnet, four commits rebased cleanly onto sel-fix5, pushed as sel-fix5-rsmut at fc51e9a; 50/50 Rust tests pass against the rebased tree. None of it touches the edge-resolution path — the whole diff is insertions inside canonical_number's own nested mod tests, with zero hits for references, by_key, matches_citation, resolve_reference_target or RecordRef, so nothing was measured against code batch 5 has since rewritten.

126 mutants: 89 caught / 29 missed / 7 timeout / 1 unviable → 114 caught / 4 missed / 6 timeout / 2 unviable.

is_canonical — the Q-J door, where every branch mutation had been surviving — is closed by direct grammar tests that had never been called before; the function was only ever reachable indirectly through a handful of validate_field spellings. compare, decompose, pad_right, canonicalize_decimal_digits, big_mul_small, big_to_most_significant_first, big_from_u64 and exact_decimal_from_mantissa each got direct tests, every one confirmed red under its own mutation by a real re-run rather than by inference.

Another test that passed by luck. render_f32's sign-bit mutant survived Hands' first negative-value tests because 1.0 and 2.0 happen to have bit 0 matching their real sign bit. Adding -1.0 and -2.0 killed it. Same shape as the degenerate row-tiebreak test in batch 5: a fixture whose values coincidentally satisfy the assertion.

Four survivors, all judged equivalent by hand-tracing rather than assumed: two in is_canonical (a leading-zero guard that is a pure early exit, since anything it would catch is caught later by the dot check; and an increment skipping a digit the caller already validated, which the following loop re-consumes either way), and | → ^ in both render_f64 and render_f32, where the fraction is masked to exactly 52 or 23 bits so the implicit leading bit is never set and the two operators are bit-identical.

Six timeouts are inherent infinite loops under mutation, not test gaps — a += → *= on a loop's own increment, two on big_from_u64's loop bound, and two exponent-sign corruptions that wrap a negative i32 into a near-usize::MAX bound. Any test reaching those lines hangs regardless of what it asserts; the timeout is cargo-mutants' signal for "behaviour changed", it simply cannot classify a hang as caught.

-j4 is not reliable for verdicts on this image, and that is retroactive

The identical commit gave different missed and timeout sets across three separate -j4 runs — render_f64's L284 alone flipped between missed and caught with no code change — while every -j1 run reproduced the same 4 missed / 6 timeout. The likely cause is the shared CARGO_TARGET_DIR under concurrent mutant builds, which is the same class as this pipeline's existing shared-target scar.

This makes every concurrent mutation figure in this campaign provisional, including Stryker runs made with --concurrency 4 and any cargo-mutants result above -j1. It does not invalidate them — a killed mutant that was genuinely killed stays killed — but a survivor list from a concurrent run cannot be trusted as complete or as accurate. Where a survivor list drove a ruling, the ruling stands on the hand-confirmed evidence beside it, not on the count.

Honest scheduling account

Six real runs launched, two duplicated and killed — one caught by Hands, one Self had to flag. Effective slot time on real measurement runs was about 47 minutes, spread across just over five hours of wall-clock, most of it back-and-forth: duplicate cleanup, the -j4 flakiness detour, and the rebase. Hands reported this unprompted and in detail, which is the only reason the -j4 finding exists at all.

Unmeasured, and named

contracts.rs (about 100 mutants) and schema_discovery.rs (about 50) were not run against sel-fix5 at all. One partial run, killed mid-flight, showed several MISSED across contracts.rs's validate_* family — decode_cultnet_message_from_slice, validate_message, validate_mutation_contract, validate_schema_descriptor, validate_document_record, validate_raw_document_record, validate_shard_descriptor. That is real signal on the door family, not a count. At -j1, 150 mutants is a several-hour commitment.

Self's decision: queue it, do not run it now. The merge gate needs a Soul pass on batch 5 first, and the selection cut cannot merge on a mutation sweep of an adjacent file. Recorded as a follow-up with the partial signal attached so the next pass starts from it rather than rediscovering it.

Soul, the merge gate again, 2026-09-23

Opus, against fc51e9a. Counts re-measured: C# Networking 274 / 2 skipped, Mesh 255. Rust is 254 tests across 11 targets, not the 50 reported — that figure was tests/selection.rs alone, one target of eleven.

Verdict: Cut 1 may not merge. Two reasons, either sufficient.

SM-6, CONFIRMED, high. R-AK deleted a load-bearing guard, and wrote a false premise into the source to justify it.

This one is mine. I ruled the !member.IsReference guard was duplicate authority and ordered it deleted. Hands verified and got it wrong, then left a comment claiming "index aliases are unique per descriptor, so a data member's role string can never coincide with a reference member's."

Source contradicts that. D10 (CultCache.cs:1085-1097) groups only members whose ResolveIndexAlias is non-null — alias against alias. A plain data member with no alias contributes its MemberName to the same role namespace and is checked against nobody.

Probed: a document with a data member Owner and a reference member aliased Owner registers fine, and then every hop path throws ArgumentNullException: Value cannot be null (Parameter 'targetType'). ReferenceMembers now yields the data member, ReferencesOf matches it to the reference member's edges by role string, member.TargetType! is null, and ResolveTargetLeaves throws. Before the deletion the guard skipped that member and the selection returned one edge.

Restoring the guard leaves 274/274 green — the suite has no opinion either way. Any schema shaped that way makes every cites/cited selection over it fault, the peer gets no typed refusal, and the batch's own new dispatch backstop swallows it with no log.

CONFIRMED, medium. The dispatch backstop is untested and answers nothing.

catch (Exception) when (false) — so the backstop never matches — survives 274/274. HandleSnapshotRequestV1Async catches only the three typed selection exceptions, unlike its siblings HandleShardLogRequestAsync and HandlePutAsync, which catch Exception and answer a CultNetErrorMessage. The snapshot path is the one handler where an untyped fault becomes silence. R-AG's invariant was that malformed input reaches the wire typed and nothing else escapes; a silently dropped dispatch is not that. It is a compensator, and it is already hiding SM-6.

SM-4, CONFIRMED, medium. R-AJ closed about half of it, and Rust got the thin half.

mutation	C#	Rust
drop edge-sort To.SchemaId	survives	survives
drop To.Key	survives	survives
drop From.SchemaId	killed	survives
drop Role	killed	survives
drop row-order recordKey	killed	survives
swap row-order schema/key (control)	killed	killed

The controls dying means the rewritten tests are not degenerate in the old way — R-AJ's headline fix is real. But the new Rust edge test puts both edges on the same from.schema_id and never reads to, so it pins exactly one thing: from.record_key before role. Dropping role entirely — the component it was written for — survives.

CONFIRMED, medium. Shared-key resolution is order-dependent and unpinned in both runtimes.

Iterating candidates reversed (C#) and preferring the last in-target candidate (Rust) both survive. Nothing pins "the first row in the caller's order wins" when two in-target rows share a key; R-AF's new vector covers one-in and one-out only. Both runtimes agree in source today, and that agreement assumes both are handed rows in the same order, which nothing between a C# cache and a Rust row owner guarantees.

PLAUSIBLE. RowSet::target_leaves(role) takes no schema

(selection.rs:1333), where C# resolves the target set from the declaring member's own TargetType. Two schemas legitimately declaring one role name with different targets cannot be expressed in Rust at all. Low-medium.

PLAUSIBLE, low. A + prefix in a cursor length field

parses in Rust (str::parse::<usize>) and is refused in C# (NumberStyles.None). Gated behind the HMAC, so unreachable without forging it. Read from source, not probed.

Promises that held

R-AF's wire fact is confirmed, and harder than the probe claimed. It is not one serialization — the type has nowhere to put a schema id. CultRecordRef<T> has exactly one field, ICultRecordRef exposes only Key, the formatter writes only the key, and all three declared reference shapes yield CultRecordKey. A document carrying scalar, many-list and many-dictionary references at once encodes with no schema id anywhere. No other formatter, version or nesting can introduce one without changing the type. C#'s rule was right, Rust adopting it is correct, and it is not wrong in the other direction.

Both runtimes now agree at the evaluator on every shape the previous verdict split them on — seven matched probes built independently in each runtime, identical answers, including the shared-key matrix in either row order. The role-filter reordering, which no ruling covered and only Hands had checked, agrees with C#.

R-AG is behaviourally pinned: moving the cast inside the parens — still wrapping, still looking like a fix — kills a test. Dropping the redundant length < 0 survives and is genuinely equivalent.

R-AH: right scope, under-specified, still unbuilt

Comparing the payload rather than the envelope is the right call — the C# message carries a non-deterministic messageId and three shard fields Rust has no concept of, while the shared six are in the same declaration order on both sides with camelCase keys already.

But "the canonical MessagePack bytes" is not a specification. The same page encodes as 285 bytes compact and 402 named, and the compact form emits a 4-element array for a 6-field struct because skip_serializing_if drops absent optionals positionally — it is not decodable. The ruling must also fix map encoding with camelCase keys, declared field order, and whether an absent optional is omitted (Rust today) or nil (C# today). Those differ right now, and the same rule has to reach Edge, RawDocumentHeader and SelectionDocumentRecord.

The structural reason it keeps slipping: C# has no page type at the evaluator layer. Select returns rows, and Projection appears once in the evaluator, inside the cursor digest. Either Select gains projection, or the harness builds the page outside it and the vectors start testing CultNetDocumentRegistry too. Nobody has chosen, and my ruling did not.

At fc51e9a: pageBytes|payloadBytes|canonicalBytes is 0 hits in both vector files, rs-written.json still carries no projection vector, and door refusals still compare a field and never a code.

Can parity be claimed without it? No — not byte parity. What exists is semantic parity over an agreed field subset. Even the selection bytes differ between the two written files for the same vector — C# writes nine fields with nils, Rust omits absent ones — so there is no byte-level agreement anywhere in this cut.

What Soul could not run

No Stryker run; the 47 survivors stay untriaged, and two of Soul's own confirmed survivors are on lines batch 5 changed, so that list is not empty of real signal. The --since/GitInfoProvider crash is still not root-caused. contracts.rs and schema_discovery.rs unmeasured. The R-AH harness unbuilt, pending the shape decision. Caching's three failures still unproven. One run was lost to CRLF and still exited 0 — fixed in the stopgap at Eureka b12016a, with a verdict sentinel so a job that never ran cannot report success.

Self's rulings for fix batch 6, 2026-09-23

• R-AL (SM-6). Restore the guard, and fix the owner. The guard is not duplicate authority — filtering members by kind is the evaluator's own job, and I was wrong to call it a duplicate. Restore it, delete the false comment, and add the probed shape as a test that fails without it. Separately, D10 is the owner of role-namespace uniqueness and currently admits the collision: make registration refuse a data member whose name collides with a reference member's alias. Both, with tests. Standing: a "redundant" deletion must be probed at the shape that would break it, not argued from a property nobody checked. This is the second time a redundancy ruling of mine was wrong; the first was caught by Hands, this one was not.
• R-AM. The snapshot handler answers like its siblings — catch Exception, return a CultNetErrorMessage. The dispatch backstop logs, always. A backstop that swallows silently is a compensator, and this one hid a high-severity regression for a whole batch.
• R-AN. Finish R-AJ in Rust. Its edge test pins one component of a four-component rule. Every component of both orderings dies under its own mutation, in both runtimes.
• R-AO. Pin shared-key resolution order — two in-target rows sharing a key — in both runtimes and in a vector. "The first row in the caller's order wins" is currently a coincidence of source, not a rule.
• R-AP (R-AH). Decide the shape, then build it. The payload, as a map with camelCase keys, fields in declared order, and absent optionals written as nil rather than omitted — C#'s current behaviour, because a decodable fixed-shape encoding is worth more than brevity, and because omission is what makes the compact form undecodable. The same rule reaches Edge, RawDocumentHeader and SelectionDocumentRecord. Select gains projection, so the thing the vectors test is the thing that decides; the harness must not assemble a page the evaluator never produces.
• R-AQ. Until R-AP lands, this cut claims semantic parity over a named field subset, not byte parity — and the map says so in those words. Soul is right that byte parity cannot be claimed today, and it is better to name the smaller claim than to let the larger one stand unearned. Flagged for the operator: wire parity is this campaign's stated invariant, so if semantic parity is not enough for Cut 1 to merge, R-AP becomes a blocker rather than a follow-up.

Fix batch 6 landed, 2026-09-23, verified locally and not on Yggdrasil

Sonnet, six commits on sel-fix6 (pushed by Self after review). C# Networking 279 passed / 2 skipped (was 274), Mesh 255, Caching 193 unchanged. Rust 259 across 12 targets (was 254/11). +650 / −47 across 13 files, no new targets, packages or dependencies.

The verification gap comes first, because it is Self's. The brief said "use ygg-verify.sh" without its path, where every earlier brief gave the full path. Hands searched, could not find it, and verified the whole batch locally on the workstation — the exact work the load budget forbids there. It reported the gap plainly rather than claiming the bar was met, and left the branch unpushed on purpose so the work would not look merge-ready. Both were the right calls. Recorded as a scar at Eureka 301df32; this batch still owes its Yggdrasil run, and Soul's pass will provide it.

• R-AL. The guard is restored and the false comment deleted. The real fix went to the owner: D10b in CultCache.DiscoverMembers now groups every member by its reference-role namespace (IndexAlias ?? MemberName), not only aliased ones, and refuses registration when a data member's bare name collides with a reference member's alias. Registration of the probed shape now throws. Since D10b closes the only production door, the guard itself is pinned by a test that builds the colliding descriptor through reflection to bypass the registry deliberately — and removing the guard makes it yield both members and fail.

• R-AM. HandleSnapshotRequestV1Async now catches Exception and answers a CultNetErrorMessage, like its two siblings. The dispatch backstop logs, always. Writing the test found a second defect: the backstop logged DynamicInvoke's TargetInvocationException wrapper — "Exception has been thrown by the target of an invocation" — rather than the real fault, so it now unwraps. The first run failed on the wrapper text, which is precisely the bug that half of R-AM exists to fix.

• R-AN. Two isolating Rust tests added; the previous one tied From and To across both edges, so dropping role survived. The To.SchemaId/To.Key equivalence argument recorded here was FALSE, and is deleted rather than amended. It claimed that within a page-position bucket no construction gives two edges with tied From and Role and differing To. Soul built the counterexample and ran it in both runtimes: a selection carrying both hops mixes their edges into one bucket, so a row that self-references through the same role it uses to cite a peer yields a cites edge (to = the queried peer) and a cited edge (to = itself), tied on From and Role, differing on To. A self-referencing role is a parent or peer link — ordinary, not exotic. This is the second justification-from-an-unchecked-property in three batches, and R-AL's standing rule applies to it.

• R-AO. Shared-key order pinned in both runtimes, both row orders, plus a vector case regenerated into both written files.

• R-AP, half built and half stopped, correctly. The encoding is decided and C# already did it — [MessagePackObject]/[Key("name")] writes a camelCase map with nil for absent optionals by default. The gap was entirely Rust-side: skip_serializing_if = "Option::is_none" on four types omitted keys under to_vec_named. Removed from all four, #[serde(default)] kept for decode. Byte-identical now, against real C# MessagePack 3.1.7 hex: RawDocumentHeader and Edge, each broken once by hand to confirm the assertions are not vacuous. The byte counts recorded here were wrong — 86 and 99 — and are corrected to 228 and 246, measured on Yggdrasil against MessagePack C# 3.1.7. The committed hex constants matched the real bytes exactly all along; only this prose was wrong.

  "Select gains projection" was not built, and Hands stopped and said why rather than reporting the ruling closed. Header and document projection is owned by CultNetDocumentRegistry.SelectPage/SelectAll, a different class from the evaluator, which knows only generic rows and descriptors and has no legitimate route to a schema's wire-binding override. Moving that ownership is buildable — thread a caller-supplied serializer, as ordinalOf/shardIdOf already are — but changes both signatures, every caller, and the vector harness itself.

R-AQ, restated honestly. The claim has moved from "semantic parity, byte parity absent" to "nested wire-type byte parity confirmed and vector-tested; whole-message byte parity still blocked on unbuilt page-assembly ownership." No vector yet exercises the real message output byte for byte.

Self's rulings

• R-AR. The page-assembly ownership question is its own cut, not a follow-up. It changes two public signatures, every caller and the harness, which is a cut's worth of work and exactly the kind of thing that has slipped five passes by being carried as a loose end. Cut 1 does not wait for it.
• R-AS. Cut 1's honest claim is nested wire-type byte parity plus semantic parity over the named field subset. The map says that in those words, and the target document is reconciled to it. Flagged for the operator: if whole-message byte parity is required before this cut merges, R-AR becomes a blocker and Cut 1 stays open; if it can be its own cut, Cut 1 can close on the claim above once Soul clears it.

Soul, the merge gate, third pass, 2026-09-23

Opus. Every run went to Yggdrasil, supplying the run batch 6 owed. Reproduced: C# Networking 279 / 2 skipped, Mesh 255, Rust 259 across 12 binaries. Caching is 190 passed / 3 failed — the three assert a write is refused, which root in a container cannot produce. Unproven, not passing, and Soul declined to call them environmental on its own authority.

Verdict: Cut 1 may not merge.

SM-7, CONFIRMED, high. R-AN did not land, and its justification is false.

Three named survivors are still alive:

mutation	Rust	C#
drop both to.schemaId and to.recordKey from the edge comparator	survives	survives
drop recordKey from the row comparator	survives	killed

The row survivor is the fourth line of SM-4's own table from the previous gate, and batch 6 added no Rust row-order test at all.

Soul built the counterexample and ran it in both runtimes. The false argument reasoned about each hop separately; a selection carrying both hops mixes their edges into one Vec. A row that self-references through the same role it uses to cite a peer produces a cites edge (to = the queried peer) and a cited edge (to = itself) in the same page-position bucket, tied on From and Role, differing on To. Rust's fixture peer role and a C# [CultReference(typeof(Self), many: true)] both express it. Reversing the to.recordKey comparison is killed only by Soul's probe, so the branch is live and decides real output. A self-referencing role is a parent or peer link — ordinary.

Both runtimes agree today only because the comparator text is duplicated — the "coincidence of source, not a rule" that SM-4 raised and R-AN was written to close.

SM-8, CONFIRMED, medium. The cut's parity claim has no home.

Section 9 says the claim is written into docs/runtime-parity-scope.md, "not here alone". That file contains zero mentions of typed selection, and never did — not even the earlier overstated wording. R-AS said the target document "is reconciled to it". It is not.

SM-9, CONFIRMED, low. The map's byte counts are wrong; the constants are right.

Serialized on Yggdrasil against MessagePack C# 3.1.7: CultNetRawDocumentHeader is 228 bytes, CultNetEdge is 246 bytes. The committed hex in selection_wire_parity.rs matches exactly. The landing note recorded 86 and 99. The thing Self asked Soul to distrust was sound; the prose beside it was not.

Promises that held

R-AL is complete, including inheritance. Seven shapes built through Reflection.Emit and pushed at the real registry, all refused: data name against reference alias, reference name against reference alias, alias against alias, data alias against reference name, implicit many-reference through List<CultRecordRef<T>>, and across inheritance, where the message names both declaring types. The one shape that still registers is two data members colliding with no reference present — not a hole, because the index namespace is looked up by exact IndexAlias and never by the name fallback; only the reference namespace uses IndexAlias ?? MemberName. PersistedMember.IsReferenceCandidate is a character-exact mirror of FromMember's isReference, so D10b cannot disagree with the descriptor it guards. Narrowing D10b dies; weakening the restored guard dies.

R-AM reaches the wire, not just the log. Real loopback UDP through LiteNetLib with the production transport: the null-Selection fault arrives at the client and decodes as CultNetErrorMessage with the real message. Mutating the handler to log but answer nothing dies.

R-AO's vector genuinely discriminates. Under last-wins the resolved schema stops matching the citation's declared target and the page empties, so the vector has real power rather than merely existing.

R-AP's bytes are right, and more of them than claimed. SelectionDocumentRecord is byte-identical at 269 bytes, and — the one that matters — the whole cultnet.snapshot_response_raw.v1 envelope is byte-identical at 376 bytes, Rust's encoder against C#'s message. Nothing broke on decode.

Observations, not defects

SelectionPage has no C# counterpart to be byte-identical to; byte parity is undefined for it. The new generic catch sends raw .NET exception text to an unauthenticated peer — exactly what R-AM ordered, and what the siblings do, so consistent rather than wrong, but worth its own ruling one day. D10b is stricter than the restored guard needs: a data member aliased Owner beside a reference member named Owner used to register and worked, and is now refused — defensible, but a breaking registration change with no note.

Soul on the wording, and Self's correction

Soul agrees R-AR is its own cut. It does not agree with how I described the residual gap, and it is right on both counts:

• "Nested wire-type byte parity" overstates it. The branch proves two types, each at exactly one value with every optional absent. Nothing in the batch touches a header with tags populated or an edge with a payload — precisely where string[] versus Vec<String> and byte[] versus serde_bytes could diverge. The correct words are "byte parity for RawDocumentHeader and Edge at an all-optionals-absent value".
• "Blocked on unbuilt page-assembly ownership" is the wrong word, and the expensive one. It ran two independent things together: the two encoders producing identical bytes for the same page value, and the two evaluators producing the same page value from the same rows. The second is what the vectors already compare. The first needs no ownership move at all — Soul wrote it this afternoon and it passes at 376 bytes. What R-AR buys is a single test driving rows to bytes in one call rather than two halves meeting on paper. Calling it blocked made the cheapest available evidence look expensive, which is how it slipped six passes while sitting there for the taking.

The rig

C:\Users\Meta\eureka-soul-probes\cultnet-selection-cut1-fix6\, beside batch 5's: three C# probe classes, a Rust wire probe, the counterexample patch, three mutation drivers, seven precomputed C# mutants, and a README carrying the full mutation table and the two ygg-verify.sh rakes (now in the script's own header, Eureka db1c84d).

What Soul could not run

No Stryker; the C# survivor list stays untriaged. No cargo-mutants sweep — thirteen hand-built non-revert mutations instead, all at -j1. contracts.rs and schema_discovery.rs unmeasured, with batch 5's partial validate_* signal still hanging. The --since/GitInfoProvider crash still not root-caused. Soul proved the to comparator branch is reachable; it did not prove the two runtimes would disagree on it, only that nothing stops them.

Self's rulings for fix batch 7, 2026-09-23

• R-AT (SM-7). Every component of both orderings dies under its own mutation, in both runtimes — including to.schemaId and to.recordKey, which now have a known construction to test against, and Rust's row comparator record_key, which SM-4 named a pass ago and batch 6 never touched. Take Soul's counterexample from the rig rather than rebuilding it.
• R-AU. The false equivalence argument is deleted from the map, done above, not amended. Standing rule reaffirmed: an argument that a branch is unreachable is a claim to falsify, not a reason to skip a test.
• R-AV (SM-8). docs/runtime-parity-scope.md carries the typed-selection row, in the corrected words. A claim whose named home does not mention it is not recorded.
• R-AW (SM-9). Correct the byte counts to 228 and 246.
• R-AX. Land the whole-message parity vector. Soul already has it green at 376 bytes. This is not blocked and never was — my wording made it look so. Land it as a vector on both sides, and add a header with tags and an edge with a payload, since those are the shapes where the collection and byte-array mappings could diverge and the all-absent value proves nothing about them.
• R-AY. Cut 1's claim, in these words: byte parity for the snapshot_response_raw.v1 envelope and its nested record types, verified at populated and absent optional values, plus semantic parity over the named field subset. R-AR remains its own cut and buys a single rows-to-bytes path, not the claim itself.
• D10b's strictness gets a note in the registration error and the map: a shape that used to register and work is now refused.

Fix batch 7 landed, 2026-09-23

Sonnet, seven commits on sel-fix7, all verified on Yggdrasil, one slot at a time across thirteen sequential jobs with no concurrent verdicts. C# Networking 280 passed (was 279 / 2 skipped), Mesh 255 unchanged, Rust 264 across 12 binaries (was 259). Caching stays 190 / 3, the three still unprovable under a root container and deliberately not "fixed".

No production logic changed except one error message and its comment. Everything else is tests and docs.

• R-AT. All three surviving comparator components now die, in both runtimes, at -j1: dropping both to components, reversing to.record_key, and dropping record_key from Rust's row comparator.

  And Hands found that Soul's own construction did not actually kill the "drop both" mutant. Taken verbatim from the rig, it survives in both runtimes — not because the branch is unreachable, but because select's push order happens to match alphabetical sort for that naming, and the two runtimes push in opposite orders (Rust emits cited before cites, C# the reverse). Rust needed the self-citer's key to sort after the target's, C# needed the reverse. A third fixture passing by coincidence, found while executing rather than by a later pass. Each runtime now carries its own naming with the reasoning written into the test.

• R-AV. docs/runtime-parity-scope.md gains the typed-selection row in R-AY's words, naming the four files that carry the byte and semantic vectors and calling out R-AR as separate and still open. The claim now has the home the map said it had.

• R-AW. Verified, not assumed: the committed hex constants are 228 and 246 bytes, matching a real MessagePack C# 3.1.7 run. The map's prose is corrected above; the constants were right all along.

• R-AX. The whole-message vector landed, byte-identical at 376 bytes, and the thin evidence is no longer thin:

  • RawDocumentHeader with tags = ["alpha","beta"]: 239 bytes, identical — string[] and Vec<String> agree.
  • Edge with payload = [1,2,3] and payload_encoding = "messagepack": 261 bytes, identical — byte[] and serde_bytes agree.

  No divergence. That was a real risk named in the ruling and it did not materialise.

• D10b's breaking change is noted in the error message and its owning comment: a data member aliased Owner beside a reference member named Owner used to register silently and fault at resolution; it is now refused at the door with no compatibility path.

Every condition Soul set for merging is now met. A final gate pass follows.

Soul, the merge gate, fourth pass, 2026-09-23

Opus. Four sequential Yggdrasil jobs, one slot, Rust at -j1. Merge to origin/main is conflict-free; main's 42 divergent commits touch no file this cut touches.

Verdict: Cut 1 may merge.

Baselines, with one correction. Rust 264 across 12 binaries, Mesh 255, Caching 190 / 3 — all as reported. C# Networking is 283 passed / 2 skipped, not the 280 reported. Under-reported this time rather than over; the count still went into the record wrong.

The five conditions

1. Met for everything that decides output, not met on the letter. Killed in both runtimes: dropping both to components, dropping to.record_key alone, reversing it, dropping record_key from the row comparator, plus from.record_key and role in Rust. The batch's new fixtures are the killers and they are not coincidental — Soul attacked them with single-component variants and they held.

   But to.schemaId survives both dropping and reversing, in both runtimes. Hands never claimed it dies alone; R-AT did, and R-AT was mine. Soul could not construct a selection where it decides anything — under D10b a role resolves to one declared target type and reference resolution yields at most one row per (role, key) — but says plainly that is a claim, not a proof, and the same shape of argument that has been wrong twice here. Hence the follow-up is "construct a killer or delete the component", not "record it as equivalent".

2. Partially met, and the miss is mine. The false argument is gone where batch 6 restated it and still stands, unstruck, at its original site — the EdgesFor paragraph and ruling R-AC ("Recorded, not to be chased"). Those read as live instructions. I deleted the copy and left the original, which is the same sweep failure this map has recorded before.

3. Met. runtime-parity-scope.md carries the row in R-AY's words, naming four files that all exist.

4. Met, and the constants were right all along.

5. Met, and genuinely two-sided. The three constants are byte-identical strings in both files; each runtime serializes and asserts independently, so neither reads the other's answer. All three C# tests pass and all three fail when Soul flipped one hex digit of the expected constant. 239, 261 and 376 are real MessagePack C# 3.1.7 output.

SM-10, CONFIRMED. The 246-byte all-absent Edge vector is Rust-only.

The 228-byte all-absent header is nested inside the 376-byte envelope, so C# regenerates it. The edge has no such cover. A C# edge encoder that resumed omitting a null payload/payloadEncoding would still pass the 261-byte populated vector, and nothing on the C# side would fail — which is exactly what R-AP existed to kill. It leaves runtime-parity-scope.md's "verified at populated and absent optional values" true for three cells of four.

Five lines of C# closes it.

Self's decision: close four items, then merge

Soul named this the follow-up it "would put a name on". A known hole that a five-line test closes is not a follow-up — merging over it would be shipping the exact asymmetry the cut exists to prevent, with a note promising to come back. Batch 8 closes items 1 to 4, then the cut merges.

• R-AZ (SM-10). Mirror the 246-byte all-absent Edge constant into SelectionWireParityTests.cs, so both runtimes assert all four cells.
• R-BA. Settle to.schemaId: construct a killer, or delete the component from both comparators. A component no input can distinguish is not free — it is a line the next reader must reason about, and the reasoning has been wrong twice.
• R-BB. Correct the landing note to 283 passed / 2 skipped.
• R-BC. Strike R-AC and the EdgesFor equivalence paragraph at their original site, not with a note beside them.

Not blocking, recorded as genuine follow-ups, none a known defect in the cut's behaviour and all outside the selection surface: the C# Stryker survivor list untriaged; contracts.rs and schema_discovery.rs unmeasured with batch 5's partial validate_* signal hanging; the --since/GitInfoProvider crash unrooted; Caching's three, which need a non-root container to prove and must not be "fixed".

Rig: C:\Users\Meta\eureka-soul-probes\cultnet-selection-cut1-fix7\.

Fix batch 8 landed, 2026-09-23 — the pre-merge batch

Sonnet, two commits on sel-fix8 (cda28fa, 492dd0e), verified on Yggdrasil at -j1. C# Networking 284 passed / 2 skipped (283 baseline plus the one new test), Rust 264 across 12 binaries, unchanged — the deletion is behaviour-neutral.

• R-AZ. The 246-byte all-absent Edge vector is mirrored into SelectionWireParityTests.cs, so both runtimes now assert all four cells. Proven non-vacuous: flipping byte 0 of the constant fails exactly that test and nothing else, on a throwaway commit that was never pushed.

• R-BA. Deleted, not constructed — and the argument is stronger than the one the ruling carried. Hands traced both resolvers: each resolves a reference purely from (citer's schema, role, recordKey) against the evaluation's own fixed candidate index, with no edge- or caller-specific input steering which candidate wins — Rust's target_leaves(role) does not even take a schema. So once From, Role and To.RecordKey tie in the comparator, the resolved row including its schema is forced identical by construction.

  That does not depend on schema-authoring discipline the way my D10b argument did; it depends only on the resolution function being a pure function of already-tied inputs. It is a structural proof rather than a semantic claim, which is the distinction that has bitten this cut twice. Hands checked it against the live counterexample too — the self-citing row breaks to.record_key, not to.schema_id.

  The tiebreak is gone from both comparators, with its stale documentation in a third file. No test relied on it.

All four of Soul's pre-merge items are closed, three by this batch and R-BC by Self, which struck the false equivalence argument at both original sites rather than annotating it.