W6: schedule six foundation layers with corrected citations and frozen roles - #1021

Scope

Record six foundation W6 layers as SCHEDULED after #1007 LANDED and apply the CC's terminal adjudication of 599d9bb. Add all three omitted retrieval roles, separate native page/layout references from flat conversion, remove licence exclusion, retain an individual comparison reason for each non-memory survivor and bind retained reopen conditions into readiness. All six comparisons are re-frozen before any run. The census remains 33, with pi separately scoped changed evidence.

• Base after rebase: 2a330b0a0be0bb64d1b86b89ee0c5b3bee1156a1, including Regenerate the final-stack manifest after #1018 changed one of its sources #1025 LANDED 3872b96c.
• Original consumed source base: 77185991, retained unchanged in the evidence record.
• Lane: lane:foundation.
• Owned paths: docs/w6-layer-verdict-delta-20261011.md, evidence/artifacts/w6-layer-verdict-20261011/, tests/test_w6_layer_verdict_records.py, and their sorted evidence registrations.

SOTA sources

• W6 governance: preserve survivors and resolve layers before research #1008 protocol at 569bbc59, docs/decisions/2026-10-11-layer-resolution-protocol.md, Decision steps 1–3 and Terminal states: missing votes do not adjudicate, all cumulative survivors need disposition, and SCHEDULED retains its id/hash/readiness/fixed rule.
• Generate the canonical final-stack manifest (W4 PR 2) #1007 at c2dee173, catalogs/landscape/final-stack.json and the supported scripts/final_stack.py run_pin / is_definitive readers. The original landed snapshot is hash-bound; it supplies next_review 2026-10-22 and the operational run-pin reconciliation.
• The 09-22 wave, 09-23/26/29 retained adjudications and scoped frozen-09-26 GPT fit completion at evidence/artifacts/landscape-sweep-20260926-gpt6-fit-revote-20261003/. This completion does not overwrite newer 09-29 proposal scopes.
• Every corrected B1 line is cited at its actual lock commit in citation-corrections.json, with raw source-line/file-hash witnesses. The complete 44-repository source inventory is primary-sources.json. Four invalid test refs and one invalid role ref are explicitly withdrawn as qualification gaps.
• openai/codex 092d3acd6bec3e3a14bdc7e7a2810ab628ab759d, README.md:1/:42 and justfile:87; giancarloerra/SocratiCode 2218f25153d0f3f4a76ee240a5643dbc873e80be, README.md:44 and CONTRIBUTING.md:22/:36.
• anthropics/claude-plugins-official f359d43b406057f1a8cdbc4ea93b827642c162e6, plugins/pyright-lsp/README.md:3; microsoft/markitdown fd239d5d2be43d9b68329730206b9312c7d5a388, README.md:10/:63: actual role/install statements replacing warning/bootstrap spans.
• catalogs/foundation/memory-stack-20260925.json#/non_repository_decisions and the retained 10-08 grand catalog define the hash-bound model universe. The Mac-only 10-02 deployment verdict is background, not a merit disposition for current-host memory candidates.
• CPython 3.13.16, Lib/unittest/loader.py TestLoader.discover, is the existing repository harness. The new record-contract regression fails before the corrections and passes afterward; no new runtime harness is introduced.
• agent-browser aff6125c023b810ea3f2e5deec5379e9a4270bdc, README.md:120/:206-219; MinerU b7f3dd0c5f6e6d66a298239ffd02b328b873918e, README.md:52/:153/:236-240; docs-mcp-server b8447383cfe5f9933da0a2f16a84c1a663e2d216, README.md:3/:16/:61-64: native advertised retrieval functions now enter the shared role.
• MarkItDown at fd239d5d2be43d9b68329730206b9312c7d5a388, packages/markitdown/src/markitdown/converters/_pdf_converter.py:570-589 and packages/markitdown/src/markitdown/_base_converter.py:8-25: native Markdown result has no required page/coordinates. It is tested in flat content conversion. Distinct page/block/geometric reference witnesses for Poppler and the other advertised arms, including Logics-Parsing page ids and Unlimited-OCR det markers, are bound to exact source files/lines in document-retrieval.json and primary-sources.json. The CC completion also adds token-savior reference navigation, next-plaid project indexing, xberg embedding/reranking, Qdrant multivector indexing and native MarkItDown OCR. MarkItDown OCR is qualified as native text extraction with the declared LLM client and image-source association, without fabricated layout metadata; xberg primitives and Qdrant external-vector inputs keep their scoped interface boundaries.
• MemPalace/MemPalace 1f0803a31958b30320418356b2db88c4ed9c27f9, README.md:311-316: native-client auto-save hooks establish native_memory_capture applicability; lightonai/next-plaid c6c091d7ee14c81206bb28af90d8ff506fd0235e, README.md:22: semantic code search establishes semantic_code_retrieval applicability. Exact pinned lines and file hashes are retained in the arms; execution remains required.

Evidence-class table

Local commands run

python3 -m unittest tests.test_w6_layer_verdict_records tests.test_resolution_protocol_docs -v
# before completion: exit 1 against exact 55c49da1 records, eight advertised-role failures and zero errors
# after fix: exit 0, 11/11 record contracts plus 5/5 protocol checks
python3 scripts/validate.py
# exit 0; 70 components, 11753 hashed files, 4 profiles, 239 receipts
python3 scripts/final_stack.py --check
# exit 0; 34 rows; report-only pin/status flags do not qualify FINAL
gitleaks git --no-banner --redact=100 --log-opts=55c49da1924216665b21bb538cfe43becf448fc7..HEAD .
# exit 0; no leaks on the new range

The integrity check reproduced every current prereg hash across file/row/terminal copies and verified that all 33 own-layer inputs occur in their frozen arm sets.

The previous B2-r assignments remain, with source-advertised agent-browser extraction, MinerU library retrieval and docs-mcp retrieval added. MarkItDown is scored on native Markdown content, with no page/coordinate requirement. Native source references have a distinct case and defined page/block/geometric dimensions, bound to the existing authored PDF and oracle. Each dimension is reported separately; page-only success does not establish geometry. A tool's internal layout model is not a result. Source pins, 46 arms and original fixture bytes remain bound; the document record adds one scoped reference case. All six superseded preregistration hashes remain recorded.

Each of the 25 non-memory survivors has its own retained comparison pointers and a reason the common bounded rule governs. Broader benchmark axes remain explicit untested boundaries. The 33 survivor records and pi retain operational, receipt, coverage and benchmark prerequisites inside readiness. Licence provenance stays metadata and never unqualifies an arm; spending/entitlement constraints remain distinct.

Decision record

docs/w6-layer-verdict-delta-20261011.md and the linked delta/citation witnesses. Each layer has one re-frozen id/hash/readiness/rule. The original 42 entries now become 46: remove duplicate SocratiCode and restore all five memory roles, including the native-file dependency. Components are not ranked as whole systems, and role/check applicability cannot change at run time.

A1 uses next_review from the landed view. Seven source/run version deltas have explicit reasons for the experimental/current-source or historical-control pin; no receipt transfers. FINAL requires supported adoption into matching runs[], upstream E2E and invoke at that operational pin. #1007 is satisfied; the remaining named runtime/model/instrumentation/runner inputs remain unmet. No date is a trigger.

Current prereg hashes, re-frozen before any run:

• native-clients.json: e70391bd323583f77a78e3327d29491a677e2fe7e5176c31d82e633fef124947
• code-navigation.json: e0ed8b0e911db34f91b57bea5b3d31a89c2a2f32287a177002654b38a59df40f
• document-retrieval.json: a06410fcea30b7638c56ac1df280c0a96f9912ec563459482351d167b61e4f11
• semantic-rag.json: 50ca296465f89d6a0aa643c67a063a3ded5eb38394462c5ec495972fa884b02c
• durable-memory.json: 6450b9467c9345a67dc4fb571429b1c055fa924f8b422c8b7ab48aebeaba5cee
• web-research.json: db1df25eed815d73ce09e518f0a01895b95732821e01bf22e95e8bad01a41819

Host evidence

No evidence/hosts files are changed. No native model/backend/broker acceptance is claimed.

Review status

The CC adjudicated 599d9bbba34501687c358e42b29a3c9a485a56f3 after the families diverged. The terminal correction retains its B2-r-2/B1-std and folded licence/reason/readiness decisions. The ONE completion commit carries the eight remaining advertised pairs and corrected base under the CC adjudication of 55c49da; this is not a new round. Both families read this delta; after the adjudicated commit only a new reproduced P1 reopens.

Astra's completion delta at 05fe7d112058593986e04efcf788324c9ad027da is ACK with no blockers under the supplied scope lock. She verified the eight pairs, their scoped boundaries, corrected base, three re-frozen comparisons and regression results. Her calibration records that fix2 missed these roles because it lacked a per-arm capability census. The CC's Claude completion delta and gate remain pending.

Follow-ups

• Astra completion P3 — Missing packet copies of cited sources and historical evidence: retain the pinned upstream blobs/ref witnesses, consumed comparison/final-stack/reader snapshots, authority records and older fail-before receipts at their hashes. Local verification matched 55 inventory hashes and 160 locator/quoted-line checks using 79 inventory/witness blobs; portable packet copies remain absent. Historical deliberations not supplied are not newly verified.
• Astra completion P3 — Archive and native-environment test gaps: retain genuine isolated-worktree results with declared dependencies and native prerequisites for the exact gaps in TEST-GAPS.md. Broader attempts encounter archive Git boundaries/metadata, long scratch paths, denied sockets/setup and fixture/output failures; merged usage modules lack prometheus_client and native systemd setup reports a read-only working-directory failure. These attempts are neither W6 regression evidence nor successful merged/native acceptance. No dependency installation was performed by the review.
• Astra completion P3 — Persist the corrected main-base regression: current Git ancestry independently confirms both main_base copies and body B-1 at 2a330b0, but no persisted test binds that metadata to retained merge provenance or rejects the old 64594c7 value. Add that regression in an authorized follow-up; the completion's eight-role fail-before coverage does not cover the metadata correction.
• Astra completion P3 — Historical hostname and locator hygiene: the full manifest has 20 inherited hostname occurrences on 18 lines; the patch/body adds none. Use commit-qualified locators for future historical records, and keep any migration of frozen host identifiers, registered references and digests in a separate reviewed change. No active W6 citation failure was reproduced.
• Astra P3 research queue — Uniform web fixture delivery: before affected scoring, establish an upstream-supported delivery path reachable by both local and remote retrieval arms while preserving the sealed bytes. Loopback alone does not establish remote reachability; use the reviewed, hashed amendment path without new candidates.
• Execution readiness: retain every stopped attempt and missing source recipe. The owned per-trigger/proposal readiness inputs require upstream runtime/model/instrumentation and supported benchmark seals; this packet is bounded source/record evidence, not general SOTA merit or completed runtime qualification.
• Source coverage: cross-layer source locators, reviewer packet attachment completeness and longer-term retention of prior preregistration preimages remain explicit follow-ups. Current and superseded hashes are recorded; historical bytes are not silently rewritten.

Checklist

• No workflow, permission, paid-service or credential changes.
• No account/host identities or absolute home paths in the packet.
• Peer work and protected paper units/configuration/bundles/recorders preserved.
• Six SCHEDULED records retained; ONE terminal adjudication commit with all six hashes re-frozen.
• New-range leak scan and required integrity checks pass.
• Astra and CC Claude delta reads at the corrected head, through the co-op.

Follow-ups from the completion reads (CC, 2026-10-12T11:02:01Z)

These are P3 items. Each would need a re-freeze, so none is folded into this landing:

• Claude micro at 05fe7d11:
  • P3-1: stale or incomplete head text, including the Unlimited-OCR survivor reason at document-retrieval.json:934 and resolution-delta.json:891 and :903.
  • P3-2: Code Review Rules (concurs with the GPT P3 3).
  • P3-3: MarkItDown OCR before any run. This is not an exclusion ground.
• GPT family fix3 at 05fe7d11:
  • missing packet copies of cited sources and historical evidence;
  • archive and native-environment test gaps;
  • a persisted regression for the corrected main-base claim;
  • historical hostname and locator hygiene;
  • the settled shared-source delivery question for foundation/web-research.