Refresh the common map for three incoming vacuolar records before PR #966 admission

[realmarcin]

PR #966 fails its required QC because GitHub tests the synthetic merge against e218c065cd07befb3179f8598cd3b741f8495a33, which includes three structure additions after the accepted 323-record snapshot 960633eb9a1a4296ee07813abab2bf9e2e3e93ee. The committed BGE bundle still correctly binds those original 323 inputs; publication refuses the enlarged corpus.

Evidence: failed QC job at draft head 3db5c715387d06ab7392a8d7476de7af83348e45. Six tests/test_scripts.py renderer tests fail with map is stale relative to current adapter input; 581 passed and 3 skipped. The other three PR checks passed. This is a source/artifact refresh blocker, not a reason to weaken the freshness gate or change its checkout to avoid the merged source.

The exact immutable base comparison contains 27 changed paths across #957, #961 and #963: three added structure YAMLs, history, README and generated pages/MiniLM artifacts. There are no edits/removals of prior structure YAMLs and no adapter/runtime/dependency changes. Git-verified additions passed the actual prepared semantic text builder:

• GO:0000220 — vacuolar proton-transporting V-type ATPase, V0 domain (data/structures/energy_complex/vacuolar_proton_transporting_v_type_atpase_v0_domain.yaml); semantic text SHA256 de4fa247fd2250347cb59fb30ef01e2e0a5e2ec135e2fb5cf5d64fc7152a482d.
• GO:0000221 — vacuolar proton-transporting V-type ATPase, V1 domain (data/structures/energy_complex/vacuolar_proton_transporting_v_type_atpase_v1_domain.yaml); semantic text SHA256 bbc63f7254c593e45c8ea0d90c279b5c8adefda95f6ac349e9ab4ebf48c92341.
• GO:0005775 — vacuolar lumen (data/structures/membrane_organelle/vacuolar_lumen.yaml); semantic text SHA256 3888519e6f7654c9bfe3334421a02baaca31800f388ad9381c72ac03718248de.

Acceptance:

• Integrate the exact reviewed e218c065 source baseline into the draft, preserving every incoming source/history change and all five new MiniLM source/site assets. Keep runtime 9ef77083… and the deployed c8b8f89e… governance pin.
• Export and verify the complete 326-record common input against Git source bytes. Confirm all prior 323 semantic rows are unchanged; bind the three new identifiers/text hashes explicitly.
• Preserve the old cache/bundles/receipts; refresh through a separate SQLite backup with the same pinned BGE profile and exact miss allowlist. At most three new vectors should be encoded; actual misses must be checked before model loading. Run the required real canary before the full projection.
• Publish the new 326-point PaCMAP bundle through the maintained renderer, check every common and MiniLM link, and retain the old accepted snapshot as historical provenance. Do not regenerate MiniLM or run models in CI.
• Re-run the affected renderer checks and authoritative QC, then require green checks on the final PR head/synthetic merge before queue admission.

Related: #910, CultureBotAI/culturebotai-claw#426 and CultureBotAI/culturebotai-claw#440. Exact base-delta, source-blob and CI receipts were recorded locally under reviews/runtime440-final-six/; no repository or model mutation was made during diagnosis.

[realmarcin]

The latest PR #966 run is another instance of this current-source blocker, not a failure of the reviewed 330-record snapshot.

• Published head 96dac5d9e5ff11f5fa7223fe884da8b1f8c67550 passed the complete local quality gate: 587 passed, 3 skipped; all other maintained gates passed.
• GitHub QC run 34982788442 actually checked out synthetic merge e4fe52aa4b234adea78323283f95ece92e1a1a62, merging that head into 87d5c0bf4a2c223317b7f9538ad580d6e1cee744.
• That runner's retained MiniLM check reports 332 records. Six renderer tests correctly reject the 330-record common bundle as stale; the other 581 tests pass, with 3 skipped. Identifier-liveness, label-correspondence and vendored-sync pass.
• Complete immutable Git-tree comparison confirms all 330 prior source blobs/modes remain identical. The only new common records are ciliary membrane (GO:0060170) and ciliary transition zone (GO:0035869). All five incoming MiniLM assets are bound exactly; the old 330 MiniLM record/vector values remain unchanged.

The repair is bounded to that exact CI-proven 87d5c0bf revision: preserve the accepted 330-record artifacts/cache, add only the two new BGE vectors through the reviewed canary, project/verify the 332-record bundle, regenerate the maintained site, and rerun the affected quality gate. The production freshness guard stays strict; there is no blind CI retry or model/profile change.

Local evidence is retained under workspace/embedding-remediation-20260915/builds/cellstructuremech-main-87d5c0bf/, with independent diagnosis in reviews/cell332-ci-diagnosis/REVIEW.json.