fix: stop evicting in-flight curation decisions on an unrelated publish

[Barsoomx]

What

Shortlist revalidation compares the shortlist the verdict actually depended on — its entries
and comparison_complete — instead of a hash that also includes a project-wide memory count.

_manifest_hash and everything it hashes are unchanged.

Why

Measured on the stand: in one 20-minute window, 46 successful curation derivations against 45
stale_decision. Roughly half of all judged candidates were thrown away after the provider
call was paid for. At the observed rate that was the single largest remaining source of
discarded paid inference — ~$1.8/hour with a 74% discard rate.

_manifest_hash (curation_shortlist.py:181) hashed three things:

authorized_corpus_count   <-- a bare COUNT of every authorized memory in the project
comparison_complete
entries[]

_revalidate_shortlist rebuilt the shortlist inside the settle transaction and required the
whole hash to match. So any newly published memory anywhere in the project bumped the count
and evicted every curation decision then in flight — including decisions about candidates the
new memory has nothing to do with and would never have appeared alongside. With concurrent
workers on one project, a single publish wasted the calls racing beside it.

authorized_corpus_count is not an input to any decision. Every rung of derive_decision reads
candidate tier, per-target facts and comparison_complete; _apply_evidence_policy reads the
same. The count exists for telemetry and for the stored manifest identity.

Soundness

The shortlist is top-K (MAX_ENTRIES = 12). A new memory that would have ranked into it changes
entries and is still caught. A memory that ranks below the cut could not have appeared in the
shortlist the judge saw, so it cannot change the relation set that was judged. Dropping the bare
count therefore removes evictions that could not have affected the verdict, and only those.

Fences deliberately left intact

• _manifest_hash itself — it is stored on CurationDecision.comparison_manifest_hash, carried
  on CurationJudgeResult, and compared in _validate_verdict. Changing it would invalidate
  stored records and that parity check. The new shortlist_revalidation_hash is separate.
• _shortlist_memory_fence — target advanced still means stale_decision.
• _candidate_fence, _acquire_targetless_publication_lock, and the _validate_verdict
  manifest parity check are untouched.
• revalidate_curation_shortlist still returns False on CurationShortlistError.

Tests

• revalidation succeeds when an unrelated authorized memory appears after the freeze (the defect)
• revalidation fails when a selected entry's transition advances
• revalidation fails when a selected entry's body changes without a transition
• revalidation fails when a new memory displaces a shortlist entry
• revalidation fails when comparison_complete regresses
• settle path: a publish commits while an unrelated memory appears mid-flight
• settle path: a publish is still stale_decision when the appearing memory would have been
  shortlisted
• a verdict judged against a different manifest is still stale_decision

Verified: pytest -m "not transactional" → 3490 passed; pytest -m transactional → 89 passed;
engram_curator_eval → all thresholds pass; ruff check and ruff format --check clean.

Operational note

Curation is currently paused on the stand via ENGRAM_CANDIDATE_DECISION_ENABLED=0 (B-012),
because it was burning ~$43/day at a 74% discard rate. The flag lives only in the server's
.env, which a release overwrites from the local copy, so deploying this change lifts the pause
automatically and the blocked works clear on the configuration-fingerprint change.

[Barsoomx]

✅ Snyk checks have passed. No issues have been found so far.

Status	Scan Engine	Critical	High	Medium	Low	Total (0)
✅	Open Source Security	0	0	0	0	0 issues

💻 Catch issues earlier using the plugins for VS Code, JetBrains IDEs, Visual Studio, and Eclipse.

[codecov-commenter]

Bundle Report

Bundle size has no change ✅

Affected Assets, Files, and Routes:

view changes for bundle: engram-frontend-client-array-push

Assets Changed:

Asset Name	Size Change	Total Size	Change (%)
server/middleware-*.js	2.72kB	3.21kB	548.89% ⚠️
server/middleware-*.js	-2.72kB	495 bytes	-84.59%
static/qGnLcJ8hWrta19wOF19jp/_buildManifest.js (New)	1.51kB	1.51kB	100.0% 🚀
static/qGnLcJ8hWrta19wOF19jp/_ssgManifest.js (New)	77 bytes	77 bytes	100.0% 🚀
static/LjnlkYCfPCZsainkYH8p_/_buildManifest.js (Deleted)	-1.51kB	0 bytes	-100.0% 🗑️
static/LjnlkYCfPCZsainkYH8p_/_ssgManifest.js (Deleted)	-77 bytes	0 bytes	-100.0% 🗑️

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 91.10%. Comparing base (ae40b22) to head (f916cf7).
⚠️ Report is 1 commits behind head on master.
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #306   +/-   ##
=======================================
  Coverage   91.09%   91.10%           
=======================================
  Files         223      223           
  Lines       23344    23350    +6     
=======================================
+ Hits        21266    21273    +7     
+ Misses       2078     2077    -1     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.