WS-ART-001-03B3B3B: add bounded DOCX extraction

.agent-loop/initiatives/WS-ART-001-immutable-artifact-storage/STATUS.md

@@ -77,9 +77,12 @@ merge; repository-authored evidence alone is not authority.
 and adds bounded passive-PDF text extraction inside the existing isolated
 child.
 
-03B3B3A is the active successor. It installs only the approved `defusedxml`
-wheel and adds the shared bounded OPC/OOXML container security capability. It
-does not activate a DOCX, PPTX, or XLSX extractor, AUTH, or sufficiency.
+03B3B3A merged through PR #233. It installs only the approved `defusedxml`
+wheel and adds the shared bounded OPC/OOXML container security capability.
+03B3B3B is the active successor. Its bounded DOCX extraction and durable
+omission-fact implementation passed focused tests and required internal review;
+hosted CI, CodeRabbit, and human merge remain pending. PPTX, XLSX, image, AUTH,
+and sufficiency work remain inactive.
 
 AUTH `WS-XINT-002-04B` follows the complete hidden split-03B series and
 activates only fixed-service binding and guide read. ART-03C then removes the

.agent-loop/initiatives/WS-ART-001-immutable-artifact-storage/chunks/WS-ART-001-03B3B3B-docx-extractor.md

@@ -30,9 +30,15 @@ P2
 ```text
 backend/app/modules/artifacts/guide_docx.py
 backend/app/modules/artifacts/guide_extraction.py
+backend/app/modules/artifacts/guide_extraction_service.py
 backend/app/modules/artifacts/guide_extraction_worker.py
+backend/scripts/run_test_lanes.py
+backend/tests/test_artifact_architecture.py
 backend/tests/test_guide_docx.py
+backend/tests/test_guide_extraction.py
+backend/tests/test_guide_bindings.py
 backend/tests/fixtures/guide_docx/**
+backend/tests/fixtures/guide_extraction_probe_worker.py
 docs/spec_artifact_storage_service.md
 .agent-loop/initiatives/WS-ART-001-immutable-artifact-storage/**
 ```
@@ -50,16 +56,39 @@ PPTX/XLSX behavior, package additions beyond approved lock, framework/AUTH/Celer
   comments, tracked-deletion text, and embedded objects are omitted and
   recorded in omission facts. Streaming stops with `limit_exceeded` before
   exceeding D42's 4 MiB canonical-output limit; no partial result is usable.
+- Canonical DOCX output is compact sorted JSON with one `blocks` array. A
+  paragraph block is `{"type":"paragraph","text":"..."}`. A table block is
+  `{"type":"table","text":"..."}` with tab between cells, newline between
+  rows, newline between multiple paragraphs in a cell, and a newline on each
+  side of a nested table when adjacent cell content exists. Empty paragraphs,
+  rows, and cells remain explicit. `w:t` and visible field-result text are
+  retained; `w:tab` becomes tab and `w:br|w:cr` becomes newline. Hyperlink
+  display text stays in document order. Field instructions, hidden text,
+  comment markers, tracked deletions, drawings, pictures, and other passive
+  non-text body objects are omitted. Active embedded package content remains a
+  malformed OOXML rejection before DOCX extraction. Traversal deeper than 64
+  nested document/container levels returns stable `malformed` evidence.
+- Successful DOCX omission facts have the fixed bounded boolean schema
+  `truncated`, `omitted`, `headers`, `footers`, `comments`,
+  `tracked_deletions`, `embedded_objects`, `hidden_text`, and
+  `field_instructions`. The existing isolated-child result protocol carries these
+  facts; other formats retain the exact default
+  `{"truncated":false,"omitted":false}`. Persistence binds and replay-checks
+  omission facts with the canonical output. DOCX uses exact policy identity
+  `guide-extraction-v3`, so obsolete unsupported evidence cannot replay.
 - Import the parser only in the isolated child and prove deterministic output,
   malformed/unsafe input, exact output boundary, crash, timeout, cancellation,
   cleanup, approval-gate, and coverage behavior.
+- Assign the focused DOCX test module to an existing canonical hosted semantic
+  lane without changing lane or coverage policy.
 
 ## Verification commands
 
 ```bash
 (cd backend && uv run ruff check app tests)
-(cd backend && python scripts/check_guide_extractor_dependencies.py)
-(cd backend && uv run pytest -q tests/test_guide_ooxml.py tests/test_guide_docx.py tests/test_guide_extraction.py --cov=app.modules.artifacts --cov-report=term-missing --cov-fail-under=90)
+(cd backend && uv run python scripts/check_guide_extractor_dependencies.py)
+(cd backend && uv run pytest -q tests/test_guide_ooxml.py tests/test_guide_docx.py tests/test_guide_extraction.py tests/test_guide_bindings.py tests/test_artifact_architecture.py)
+(cd backend && uv run pytest -q tests/test_guide_docx.py --cov=app.modules.artifacts.guide_docx --cov-report=term-missing --cov-fail-under=90)
 (metadata_dir="$(mktemp -d)" && trap 'rm -rf "$metadata_dir"' EXIT && (cd backend && WORKSTREAM_TEST_ADMIN_DATABASE_URL=postgresql+asyncpg://workstream:workstream@localhost:5433/postgres .venv/bin/python scripts/run_isolated_tests.py --metadata-json "$metadata_dir/result.json" --timeout-seconds 12600 -- .venv/bin/python -m pytest -q --ignore=tests/test_isolated_database_runner.py --cov=app --cov-report=term-missing --cov-fail-under=78))
 python3 scripts/check_stale_artifact_contracts.py
 python3 scripts/check_markdown_links.py

.agent-loop/initiatives/WS-ART-001-immutable-artifact-storage/reviews/WS-ART-001-03B3B3B-external-review-response.md

@@ -0,0 +1,42 @@
+# External Review Response: WS-ART-001-03B3B3B
+
+## Comments addressed
+
+- GitHub Agent Gates flagged `worker protocol` as stale human-worker
+  vocabulary in the chunk contract and storage specification. Both references
+  now say `isolated-child result protocol`, and the exact stale authorization
+  documentation check passes locally.
+- The first Backend gate reported `current_node_inventory_mismatch`. The DOCX
+  database-test parameter used generated ZIP bytes as its pytest ID, so ZIP
+  timestamps changed the collected node between the lane and independent
+  inventory passes. Stable explicit `json` and `docx` IDs now make repeated
+  collections identical.
+- CodeRabbit reported four valid DOCX hardening/test gaps. The extractor now
+  returns stable `malformed/docx_nesting_limit` beyond 64 recursive levels,
+  preserves `w:sdt`-wrapped rows and cells, resolves validated OOXML part names
+  through their case-folded stored-name map, and proves the full nine-key
+  omission schema across the isolated-child JSON round trip.
+
+## Comments deferred
+
+None.
+
+## Human decisions needed
+
+None.
+
+## Commands rerun
+
+- `python3 scripts/check_stale_authorization_docs.py`
+- `python3 scripts/check_markdown_links.py`
+- `git diff --check`
+- repeated `pytest --collect-only` for the affected binding module
+- `ruff check tests/test_guide_bindings.py`
+- `ruff check app tests scripts`
+- focused DOCX/OOXML/extraction/architecture/lane suite (154 passed)
+- DOCX coverage suite (12 passed, 93.24 percent)
+
+## Remaining risks
+
+Hosted Backend/Agent Gates and CodeRabbit completion remain required on the
+repaired PR head before merge readiness.

.agent-loop/initiatives/WS-ART-001-immutable-artifact-storage/reviews/WS-ART-001-03B3B3B-internal-review-evidence.md

@@ -0,0 +1,65 @@
+# Internal Review Evidence: WS-ART-001-03B3B3B
+
+Reviewed against trusted main: `2f441e99f1af`
+
+Reviewed at: `2026-07-31`
+
+## Candidate
+
+Hidden deterministic DOCX extraction on the merged shared OOXML boundary. The
+candidate emits bounded canonical content and durable omission facts without
+activating AUTH, guide sufficiency, PPTX/XLSX, or contributor submissions.
+
+## Deterministic Evidence
+
+- Ruff, approved extractor-dependency gate, stale artifact contracts, stale
+  wording, Markdown links, lane integrity, and `git diff --check`: PASS;
+- focused DOCX, OOXML, extraction, architecture, and lane suite: 154 passed;
+- focused DOCX adapter suite: 12 passed at 93.24 percent coverage;
+- DB-backed binding/replay tests and repository-wide coverage remain delegated
+  to the hosted sharded Backend gate; no local full-suite run was used.
+
+## Reviewer Results
+
+| Reviewer | Result | Blocking findings |
+|---|---|---|
+| architecture | PASS | none |
+| security/auth | PASS WITH LOW RISKS | none |
+| QA/test | PASS WITH LOW RISKS | none |
+| senior engineering | PASS WITH LOW RISKS | none |
+| product/ops | PASS WITH LOW RISKS | none |
+| reuse/dedup | PASS WITH LOW RISKS | none |
+| CI integrity | PASS | none |
+| test delta | PASS | none |
+| docs | PASS | none |
+
+## Material Repairs
+
+- made body traversal omission-aware so block-level deletions and move-from
+  content cannot enter canonical output;
+- covered direct, inherited character, inherited paragraph, and document-
+  default hidden-text semantics and recorded simple-field instructions;
+- restored the worker-owned shared OOXML loader instead of adding a second
+  production validation path;
+- made omission facts part of the bounded worker protocol, immutable persisted
+  evidence, and replay identity;
+- corrected invalid-protocol tests so every assertion reaches its intended
+  validation branch;
+- documented the exact canonical blocks and fixed omission schema.
+- bounded recursive traversal, preserved content-control-wrapped rows/cells,
+  and resolved case-variant validated OOXML part names after external review.
+
+## Accepted Low Risks
+
+- Explicit false values on `w:vanish` are conservatively treated as hidden;
+  exact WordprocessingML visibility fidelity can be hardened later.
+- The isolated child assembles a bounded block before the exact 4 MiB final
+  check; existing input, decompression, memory, CPU, wall-time, and parent
+  protocol limits prevent partial durable output.
+- Parent and child deliberately repeat their small omission schemas so the
+  isolated protocol is independently validated; future formats should avoid
+  allowing those literals to proliferate.
+
+Valid findings addressed: yes
+
+Open sub-agent sessions: none

.agent-loop/initiatives/WS-ART-001-immutable-artifact-storage/reviews/WS-ART-001-03B3B3B-pr-trust-bundle.md

@@ -0,0 +1,108 @@
+# WS-ART-001-03B3B3B PR Trust Bundle
+
+## Chunk
+
+`WS-ART-001-03B3B3B` — DOCX Extractor (L1).
+
+## Goal and human-approved intent
+
+Add deterministic, bounded DOCX text and table extraction after exact DOCX
+classification and the merged OOXML security boundary. Preserve original
+verified guide bytes as authoritative. This hidden chunk must not activate
+AUTH, invoke guide sufficiency, add other formats, or touch contributor ZIPs.
+
+## What changed and why
+
+- Added an isolated DOCX adapter that emits compact sorted JSON blocks in
+  document order with fixed paragraph, row, cell, and nested-table semantics.
+- Added a fixed omission schema for headers, footers, comments, deletions,
+  passive objects, hidden text, and field instructions.
+- Advanced DOCX evidence to `guide-extraction-v3` and bound omission facts to
+  immutable persistence and replay checks.
+- Reused the existing worker-owned OOXML validator loader and kept both the
+  validator and DOCX adapter imports confined to the isolated child.
+- Added focused tests, the existing hosted semantic lane assignment, and the
+  exact storage-service contract.
+
+## Design chosen and alternatives rejected
+
+The isolated worker composes the shared OOXML validator with the DOCX adapter
+before descriptor-only seccomp. The adapter receives an explicit validator,
+then parses only validated `word/document.xml` and bounded style facts. The
+parent accepts only the exact result and omission schemas. Rejected alternatives
+were upload-request parsing, direct provider access, a second OOXML validation
+path, raw binary agent input, partial output, or generic document authority.
+
+## Scope control and product behavior
+
+Only hidden guide-source extraction changes. There is no route, authorization
+activation, Celery continuation, sufficiency invocation, provider read, guide
+binding, submission, checker, review, contribution, payment, or reputation
+change. Unsupported and malformed content remains a bounded internal artifact
+outcome, not a guide-insufficiency decision.
+
+## Acceptance criteria proof
+
+- Paragraphs, hyperlinks, tabs, breaks, tables, cells, nested tables, and empty
+  structures have exact canonical output tests.
+- Headers, footers, comments, tracked deletion/move-from content, passive
+  objects, fields, direct/inherited style-hidden text, and document-default
+  hidden text are omitted and recorded.
+- Active content and malformed XML fail through stable bounded outcomes.
+- Exact output-limit behavior returns no partial usable result.
+- The real isolated runner proves v3 policy, omission transport, and scratch
+  cleanup; architecture tests prove worker-only adapter/validator imports.
+- Persistence tests bind canonical output and omission facts and reject obsolete
+  policy evidence as a replay target.
+
+## Tests and checks
+
+- Ruff and approved extractor-dependency gate — pass.
+- Focused DOCX/OOXML/extraction/architecture/lane suite — 154 passed.
+- DOCX module coverage — 93.24 percent (12 tests).
+- Stale artifact contracts, stale wording, Markdown links, lane integrity, and
+  `git diff --check` — pass.
+- Hosted Backend/Agent Gates retain DB-backed replay, repository-wide coverage,
+  and semantic-lane proof; no local full-suite run was used.
+
+## Test delta and CI integrity
+
+No test, assertion, lane, workflow, dependency rule, or coverage threshold was
+removed, skipped, or weakened. Invalid result-shape fixtures carry valid
+omission facts so they still reach their intended validation branches. The new
+DOCX module joins the existing `shared_foundations` lane.
+
+## Reviewer results
+
+Architecture, CI integrity, test delta, and docs pass. Security, QA, senior
+engineering, product/ops, and reuse/dedup pass with only documented low risks.
+All blocking deletion traversal, hidden-style cascade, isolated-runner, shared-
+loader reuse, protocol-test, and documentation findings were repaired and
+re-reviewed.
+
+## External review
+
+The first Agent Gates run found two stale human-worker vocabulary matches; both
+now use the exact isolated-child result-protocol wording and the local gate
+passes. The first Backend run exposed a nondeterministic raw-ZIP pytest ID;
+explicit stable format IDs now produce identical repeated collections.
+CodeRabbit's first attempt was rate-limited without code findings and must be
+retriggered. Its completed review then identified four valid gaps: recursive
+depth, content-control-wrapped rows/cells, case-folded part resolution, and the
+cross-process omission assertion. All four are repaired with focused regression
+proof. Hosted checks remain required on the repaired PR head.
+
+## Remaining risks and follow-up
+
+DOCX has a large visibility model; this chunk intentionally supports a bounded
+subset and treats `w:vanish` presence conservatively, including explicit false
+values. Future formats should reuse the same isolated result protocol without
+growing duplicated schemas. PPTX/XLSX, durable sufficiency continuation, and
+AUTH activation remain separate later chunks.
+
+## Human review focus and merge ownership
+
+Review exact canonical ordering/separators, omission semantics, injected shared
+OOXML validation, v3 replay identity, isolation, and absence of AUTH or
+sufficiency activation. A human owns merge approval; the agent will not merge
+this PR.