PR C: reconcile count surfaces and WAL isolation

[ohjonathan]

Summary

Reconciles the four D1b-counts findings from the independently verified PR B baseline f2ed48d, and deterministically retires the accepted portfolio WAL/SHM flake.

• centralizes concept vocabulary loading for map, activation/doctor, snapshots, and MCP;
• threads configured max_dependency_depth (including zero) through snapshots and MCP context_map;
• excludes the configured generated context map once in scoped document collection;
• makes maintain check_links use the same body-aware diagnostics as link-check;
• closes every portfolio SQLite connection on success, failure, and setup exceptions;
• routes activation through the semantic no-op map writer.

Package metadata remains 5.0.0. Tags, publication, merge, and release actions are maintainer-owned.

Closes no issue. Implements the four remaining count rows in #148 for review.

Before / after count matrix

Controlled fixtures isolate one counted concept at a time. N/A means that surface does not expose that diagnostic class.

Finding / fixture	map	activate	doctor	link-check	maintain	MCP snapshot / map
D1b-counts-1, one unknown concept	1 → 1 curation warning	0 → 1	0 → 1	N/A	N/A	0 → 1
D1b-counts-2, depth-8 chain, max 10	0 → 0 depth warnings	0 → 0	0 → 0	N/A	N/A	3 → 0
D1b-counts-3, 2 docs + generated map	2 docs / 0 map-orphans → same	2 / 0 → same	docs 3 → 2; activation 2 / 0 → same	3 / 1 → 2 / 0	2 / 0 → same	3 / 1 → 2 / 0
D1b-counts-4, broken body wikilink	N/A	N/A	N/A	1 → 1, exit 1	0 / success → 1 / failure	N/A

The concept reconciliation is validation-on. The default docs scope remains clean at 219 documents, 0 errors, and 0 warnings. Non-default library scope deliberately changes from 576 to 1,450 warnings because it now includes the same 874 vocabulary findings already reported by map; no golden or tripwire was silently rebaselined.

Verification

• focused PR C suite: 229 passed
• complete suite: 1,600 passed
• post-review affected suite: 114 passed
• exact doctor phase-4 tripwire: green
• WAL order proof: forward, reverse, and isolated target each passed 3 consecutive runs with ResourceWarning and unraisable warnings promoted to errors
• independent WAL review: 5 forward + 5 reverse runs, 510 tests, all green
• Ontos map/activate/doctor/link checks: green; activation remained 0/0 and semantic no-op reported refreshed=false
• llm-dev manifest conformance: 4/4 checks passed
• compileall, git diff --check, forbidden-scope check, and clean worktree: green

Residual risk / sequencing

The library-scope vocabulary increase is deliberate and documented. PR D remains responsible for D4a-config-1/3/5, R2-testpypi-provenance-1, and the 15/20 consolidation reconciliation. Breaking D5b-dead-code-3 remains deferred to v6.0.0.

Please keep this draft until maintainer and independent Claude review complete.

[ohjonathan]

Publication gate update for head 64285cd08a7c587df588ac6652af1eb2a95ba9f6:

• CI: 6/6 green — Python 3.9, 3.10, 3.11, 3.12, non-editable install, and v5 release gates.
• Local: 229 focused + 1,600 full-suite + 114 post-review tests green.
• Independent code review found no remaining blockers; independent WAL review passed five forward and five reverse order runs (510 tests) with resource warnings fatal.
• Worktree is clean and matches the pushed head.

Release actions remain maintainer-deferred. Stopping here for maintainer/Claude review before PR D.

[ohjonathan]

PR C review — Claude — approve to merge (no blockers, no majors)

Verified independently in a worktree at the PR head (64285cd). This was the subtlest sweep PR — the risk was reconciling counts upward into the clean activate/doctor path and reddening the passed == 12 activation tripwire. That didn't happen.

Single source of truth — not per-command patching

• New shared io/concepts.py (load_known_concepts) that map/activate/doctor and the MCP snapshot all call.
• The generated-map exclusion moved into the shared collect_scoped_documents, so every sibling drops it consistently (and maintain/activate no longer carry their own skip-patterns for it).
• io/snapshot.py now threads max_dependency_depth + known_concepts into its validator, so MCP/cache counts share the exact CLI settings.
• A new test_health_count_consistency.py (+176) locks the cross-sibling agreement in place.

The tripwire risk (D1b-1) was handled safely

run_activation now runs the same concept check as map, but on a clean repo it adds nothing. Verified live: map/activate/doctor all report 0 concept warnings and agree; activate/doctor exit 0; test_doctor_phase4 passes. The report's 0→1 / 874 figures are demonstration-fixture / library-scope numbers, not the clean default path.

The other three

• D1b-2 — snapshot honors the configured depth → MCP's 3 phantom depth warnings go to 0, matching the CLI. Downward, tripwire-safe.
• D1b-3 — generated map excluded everywhere (orphan 1→0), reconciled downward.
• D1b-4 — maintain flipped include_body=False→True, so it checks body refs like link-check. On this (link-clean) repo both agree at 0; on a repo with a broken body ref both now surface it.

WAL flake — genuinely fixed, not papered

An autouse fixture that scopes each index to tmp_path and asserts no -wal/-shm sidecars leak at teardown, plus sqlite3.connect(...) → closing(...) at the root cause. No xfail/skip/retry. Ran the full tests/mcp/ suite (286 tests) 3× — all green.

CI 6/6 (incl. 3.12); 140+ focused + the new consistency/snapshot/cache/scan-scope tests pass; ledger + #148 checkboxes reconciled and the WAL tail retired.

One out-of-scope observation (not a blocker)

ontos maintain exits 5 on the actual repo (promote_check: Document load failed, decision_history.md missing History Ledger table, archive failures). Confirmed pre-existing — base f2ed48d exits 5 with identical errors, so it is not a PR C regression. Worth its own fix, tracked separately.

Verdict: no blockers, no majors — ready to merge as a v5.0.1-track patch. Release actions remain maintainer-deferred.