diff --git a/docs/cultnet-selection-cut.md b/docs/cultnet-selection-cut.md
index a2758c36..f2e34d95 100644
--- a/docs/cultnet-selection-cut.md
+++ b/docs/cultnet-selection-cut.md
@@ -671,8 +671,183 @@ edit `src/` at the same time.
     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.
-  - **To check:** pass B reports three Caching failures as pre-existing, where
-    pass A reported one. Soul settles it against `main`.
+  - **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.
 
 **Ledger correction.** Commits 0 and 1 came in at about twice the §14
 estimate:
diff --git a/docs/typescript-quic-realtime-cut.md b/docs/typescript-quic-realtime-cut.md
index 8c32f679..b68a2dae 100644
--- a/docs/typescript-quic-realtime-cut.md
+++ b/docs/typescript-quic-realtime-cut.md
@@ -773,7 +773,110 @@ and P3. **Cut 3 still does not close.** The method moved the gaps outward.
    checks. Prove it with 5 rounds under 16 burners on Windows.
 6. **F6.** Fix the prose.
 
-To Hands (Sonnet) as the eighth fix batch.
+**Superseded the same day, 2026-09-22, by the operator's retirement of the
+mutation harness.** The operator's words: *"The manual mutation testing turned
+out to be a terrible idea, lots of guarding the shape of the code instead of
+the behavior. Better to have nothing than a harness that punishes
+refactoring."* `scripts/mutate-cultmesh.mjs` and every committed suite are
+deleted, with no fallback. Four of the six rulings above named harness entries
+(R1, W1–W3, P1, Z1, Z2); those clauses are dead text. What each ruling
+protected is not:
+
+1. **F1 stands as behaviour.** The scenarios gain a **release** configuration
+   on both targets, run against the release library. The release build is the
+   thing that ships, so a rule the release build never runs is not pinned.
+   Without the harness there is no R1 entry; the release path is proven by the
+   scenarios themselves going red when the seam's `#else` shortens the wait,
+   which Soul checks by hand on a scratch copy.
+2. **F2 stands.** The seam wraps the whole argument to `wait_for` — the
+   duration expression — and `waitseam` asserts it equals
+   `milliseconds(host argument)`. The assertion is the pin; W1–W3 are Soul's
+   hand probes, not committed entries.
+3. **F3 stands.** `pollhammer` also runs against a runtime holding live
+   objects: an open connection to an unroutable address, plus a listener and a
+   stream where the fixture can make them.
+4. **F4 stands.** Any `timeout_ms <= 0` returns within a small bound, timed
+   with a generous margin. A negative timeout is probed. The seam's sentinel
+   moves to a value that cannot collide, or becomes a separate recorded flag.
+5. **F5 stands**, and its 5 rounds under 16 burners run **on Yggdrasil**. The
+   burners that motivated this ruling are what froze the operator's
+   workstation on 2026-09-22; no CPU burner runs on Starfire again.
+6. **F6 widens.** The prose fix now includes removing every reference to the
+   deleted harness, not only the stale 1300 ms history and the scenario list.
+
+**There is no mutation evidence for the bridge and there will not be one
+committed.** It is C++, and no ecosystem mutation tool is wired here. The
+bridge's rules are pinned by behavioural scenarios at the layer that decides
+them; a mutant is a thing Soul tries by hand on a scratch copy and throws away.
+
+**The eighth fix batch landed** (Sonnet) at `5646ff7`, before the retirement,
+and the harness deletion followed at `796433f`. Both are on
+`hands/cultmath-erf`, not yet on `main`. What landed:
+
+- the seam wraps the whole duration handed to `wait_for`, and its "unset"
+  sentinel became a separate `debug_wait_recorded` flag, so a recorded `-1`
+  cannot read as nothing recorded;
+- `pollhammer` runs against a runtime holding a live connection to an
+  unroutable address (RFC 5737 TEST-NET-1), a stream on it, and a real
+  never-connected listener;
+- a `zerotimeout` scenario pins that `timeout_ms <= 0` returns within 50 ms,
+  probing 0, -1 and `INT32_MIN`;
+- `closerace`'s settle waits for `debug_peak_calls` to reach the poller count
+  instead of sleeping a guessed duration;
+- `kGenerousLateToleranceMs` narrows from 800 ms to 400 ms, so the wall-clock
+  layer catches arithmetic applied outside the seam's own argument —
+  `waitseam`'s equality check structurally cannot see that.
+
+The release configuration and the seven mutation entries (R1, W1–W3, P1,
+Z1–Z2) landed **inside the harness**, and went out with it two commits later.
+The scenario runner and every seam survived untouched: the README and the dev
+Dockerfile now document building with `-DCULTMESH_QUIC_BUILD_TESTS=ON
+-DCULTMESH_QUIC_DEBUG_ASSERTS=ON` and running each scenario as a direct
+argument to the binary. **The release configuration is therefore owed again as
+a plain build, not as a harness target**, and R1's substance — that the
+release macro is unexercised — is a Soul hand probe now.
+
+**The Linux leg ran on Yggdrasil, 2026-09-22** (Sonnet), against
+`hands/cultmath-erf` HEAD, in an image built from the committed
+`scripts/quic-native-linux-dev.Dockerfile` (pinned `debian:13`, tagged by
+content hash `cultlib-quic-native-dev:13d7fbd109fc`). No repo change was
+needed and none was made.
+
+- **Asserts build, all ten scenarios pass**, 20 rounds each: `closerace`,
+  `holdclose`, `holdtimeout`, `polltimeout`, `pollbusy`, `pollhammer`,
+  `zerotimeout`, `payloadfit`, `latecall`, `waitseam` (11 probes, 40 repeated
+  polls on one runtime).
+- **Release build, asserts off, all four required scenarios pass**:
+  `polltimeout`, `pollbusy`, `pollhammer`, `zerotimeout`.
+- Worst overshoot anywhere was 4 ms, against `kGenerousLateToleranceMs` of
+  400 ms. **F1 is closed on Linux**: the build that ships now runs the timeout
+  rule.
+
+**A defect in the README's own worked example, found by running it.** The
+example builds the test binary and runs it directly, and every scenario dies
+with `libmsquic.so.2: cannot open shared object file`. Both the bridge and the
+test executable carry `INSTALL_RPATH "$ORIGIN"` and expect the MsQuic runtime
+beside them in `bin/`, but nothing puts it there for a test build — only
+`build-quic-native.sh`'s packaged output directory gets that copy. Hands
+worked around it with a `cp` after each build and correctly did not treat it
+as a bridge defect. **This is the documented path, written in `796433f` when
+the harness was deleted, and it has never worked.** A documented path that
+fails on first use is worse than no documentation: it was published in the
+same commit that removed the working one.
+
+**Still outstanding for Cut 3:**
+
+- The **Windows release configuration**. It lived inside the harness and left
+  with it, exactly as Linux's did; Linux has now been re-run without the
+  harness and Windows has not.
+- **F5's stress under load cannot be run on Windows at all.** It wants five
+  rounds with sixteen CPU burners, and those burners are what froze the
+  operator's workstation on 2026-09-22. `closerace` is Windows-sensitive (it
+  went red under load there in four of five rounds), so the check matters most
+  on the platform where it is now forbidden. **This is an operator fork, not
+  something to route around:** either Windows stress finds a host that is not
+  Starfire, or the check is accepted as unproven on Windows and recorded that
+  way.
 
 **Cut 1 Soul findings, 2026-09-16.** Held: every test count, both negative
 greps, all twelve mutations rerun and killed, the control catching a