Bound strategic cell agent context

notes/fresh-workspace-handoff.md

@@ -25,7 +25,7 @@ own volatile workspace identity and the release witness owns deployed identity.
 
 ## Current full-world pressure test
 
-Preserve immutable Runs 75 through 108. Run 108 reaches the exact 2,400-subject
+Preserve immutable Runs 75 through 109. Run 108 reaches the exact 2,400-subject
 scale target at complexity round 43; no complexity round may replay.
 
 Bounded parallel title-by-realm compaction and deterministic parallel child-ID
@@ -58,24 +58,38 @@ and reproduced the failure without writing another session checkpoint.
 
 Pushed source `8dd3147` keeps compaction journals structurally compact.
 Immutable Run 108 reached exactly 2,400 actionable subjects at round 43. Clock
-binding then advanced the campaign to revision 1551 and resolution epoch 1545
-before strategic wave 1. All five wave-1 pulses were rejected and the run
-failed terminally; every Nemesis attempt stopped before reasoning with
-`connector frame exceeded its bound`. No simulation cell committed and
-fictional time did not advance.
+binding then advanced the campaign to revision 1551 and resolution epoch 1545.
+Pushed source `b61e762` bounds Nemesis with paged anchors and responder dossiers
+while retaining complete deterministic eligibility privately.
 
-`propose_causal_follow_through` serializes the complete responder catalog into
-instructions, then serializes the initial 24 public anchors plus every current
-pressure and clock anchor with every eligible responder view embedded again.
-That multiplicative projection is the current defect. The tool already owns the
-complete deterministic eligibility catalog and paged world/subject history.
-Keep initial anchor summaries bounded; add paged current-anchor retrieval and
-paged eligible-responder retrieval for inspected anchors; keep full eligibility
-private for admission. Resolution remains cover owner and WorldKernel remains
-commit owner. Do not raise the connector bound or truncate canonical authority.
-The acceptance launcher retains one 35 MiB content-addressed runner per exact
-source commit and deletes disposable compiler output; do not restore a
-persistent Cargo target tree.
+Immutable Run 109 at
+`/var/lib/gamecult/ghostlight-dungeon/acceptance/full-world-delvehold-b61e762-109`
+proved that cut and dispatched the exact 240-cell wave. Pulse one completed 214
+cells; four bounded retries raised the durable checkpoint to 226 completed and
+14 failed cells. The final partial checkpoint SHA-256 is
+`f502ad560e0f23db226c326072e5e8a6d0ed9975f0e63bc67606d7d7cb9b2c27`.
+All 14 terminal failures are connector-frame bounds: six Projectors received
+hundreds of inactive constituent dossiers, while eight Interpreters received
+unbounded exact activity-target maps in context and generated schemas. The
+smallest failed Interpreter cells contain only two subjects, so repartitioning
+alone cannot fix the defect. No fictional wave committed and time did not
+advance.
+
+The active cut makes context a bounded query projection. Projector receives
+only exact decision-owner dossiers and at most 24 deterministically relevant
+activity targets. The existing Interpreter agent workbench starts from the
+same bounded target viewport and may inspect up to 12 exact name/ID matches;
+its dynamic schema expands only with inspected targets. Full cell membership
+and full permissions remain private deterministic state. Effect verification
+receives only targets cited by the candidate action. Resolution remains cover
+owner and WorldKernel remains commit owner. Do not raise the connector bound,
+truncate canonical authority, or split cells merely to paper over projection.
+
+After focused gates, commit and push, then clone Run 109 and resume only the 14
+failed cells from its exact checkpoint. Do not replay complexity, clock binding,
+Nemesis, or the 226 completed cell membranes. The acceptance launcher retains
+one content-addressed runner per exact source commit and deletes disposable
+remote compiler output.
 
 ## Mission and current product shape
 

notes/ghostlight-current-system-map.md

@@ -178,7 +178,8 @@ exact campaign revision
   -> scheduler-owned Nemesis agent binds exact committed event, pressure, or clock anchors to eligible decision owners inside that cover
   -> deterministic agenda admission checks anchor visibility, simulation authority, player exclusion, and per-cell decision quota
   -> one private Projector/Persona/Interpreter membrane per cell
-  -> Interpreter agent submits or incrementally edits one private exact-decision draft
+  -> Projector receives exact decision-owner dossiers plus a bounded relevant-target viewport, never every inactive constituent dossier
+  -> Interpreter agent inspects bounded exact target matches, then submits or incrementally edits one private exact-decision draft
   -> deterministic draft compiler checks exact ownership, permissions, topology, and bounds
   -> independent per-action effect verifier checks the typed action against the Persona choice
   -> accepted unchanged actions retain their exact verifier binding across local draft repair
@@ -382,12 +383,17 @@ status reconstruction.
 
 Resolution-demand focal IDs are salience hints, not partition commands. They
 cannot create mandatory singleton cells or exceed the configured budget. Cell
-Projectors receive decision-relevant situation state; cell Interpreters receive
-exact permissions and the narrative products. Membership and revision bindings
-are derived by the runtime, so a model is never asked to copy an invariant that
-the planner already owns. Stable prompt prefixes are deliberately placed before
-dynamic state, and provider receipts expose per-attempt token/cache usage plus
-bounded local validation failures.
+Projectors receive decision-relevant situation state for the scheduler-selected
+decision owners, the count of other represented subjects, and at most 24 exact
+activity targets ranked by explicit relationship, mention in current goals or
+pressure, co-location, and stable ID. Inactive constituents remain represented
+by the exact cell cover but their full dossiers do not enter the model frame.
+Cell Interpreters receive exact permissions for active decision owners and the
+narrative products. Membership and revision bindings are derived by the
+runtime, so a model is never asked to copy an invariant that the planner already
+owns. Stable prompt prefixes are deliberately placed before dynamic state, and
+provider receipts expose per-attempt token/cache usage plus bounded local
+validation failures.
 
 `ModelPort` remains Ghostlight's single inference seam. The Codex-backed
 implementation is an ordinary typed consumer of the independent CodexConnector
@@ -433,7 +439,13 @@ subjects: complete submits bind decisions by owner-key, while incremental edits
 are rebound and validated under the selected owner by the workbench. Target IDs,
 canonical locations, pressure resolutions, movement destinations, state
 references, and public channels are enumerated from that subject's permitted
-slice. Movement and population-migration variants disappear
+slice. Activity targets are the exception to eager projection, not to exact
+authority: the initial schema exposes the bounded relevant viewport, and the
+same Interpreter workbench may query one active decision owner by a supplied
+name/ID fragment for at most 12 exact matches. A successful query expands only
+that workbench's dynamic schema. It cannot add a target to canonical permission,
+change the cell, or bypass the full deterministic validator. Movement and
+population-migration variants disappear
 when the subject has no exact destination; target-requiring activities cannot
 emit an empty target list. Semantic correction therefore judges meaning rather
 than repairing type combinations local code already knows are impossible.
@@ -442,12 +454,15 @@ Each permitted activity target is one runtime-derived descriptor keyed by its
 authoritative ID, with the target's exact name and current canonical locations.
 Reachable actor destinations map exact location IDs to names; population
 migration destinations carry both population and location identity. The
-Projector, Interpreter, and effect verifier can distinguish a canonical target
-actually named by the Persona from an unnamed local role, and can prove whether
-“go to Reed” means movement at all. Opaque permitted IDs are not enough reason
-to substitute a containing population, related institution, or unrelated
-destination. Descriptor prose is disambiguation context, not a new subject or
-fuzzy-match authority.
+Projector and Interpreter bounded views distinguish a canonical target actually
+named by the Persona from an unnamed local role, and can prove whether “go to
+Reed” means movement at all. The effect verifier receives only the exact target
+descriptors cited by its already structurally valid candidate action; the full
+permission map remains private to deterministic validation. Opaque permitted
+IDs are not enough reason to substitute a containing population, related
+institution, or unrelated destination. Descriptor prose and target queries are
+disambiguation context, not a new subject, permission writer, or fuzzy-match
+authority.
 
 Targetless local activity is explicit rather than encoded through a convenient
 unrelated subject. Preparation, investigation, communication, and obstruction