fix(wiki): refining a draft leaves its confidence as the claims set it (#578) - #580

ZETETIC-REVIEW: APPROVE

Summary

handler_refine (mcp_server/handlers/wiki_refine.py:297-317) no longer passes a fixed confidence=0.85 to update_draft, and the refined_llm memo now records float(draft["confidence"]) (the draft's own confidence, read before the update) instead of the hardcoded 0.85. Correct fix for #578, well-scoped, well-tested, well-documented.

Stakes classification (Move 7)

Medium: core curation-gate logic (affects evaluate_draft's threshold check), not auth/billing/crypto/schema/concurrency. Diff is small (3 lines in the handler; rest is tests, ADR, changelog, manifest) and single-author. Moves 1-4 applied in full; 5-6 at the touched call site.

Evidence gathered

1. Correctness of the fix

• update_draft(conn, draft_id, confidence=None, ...) (mcp_server/infrastructure/pg_store_wiki_drafts.py:98,122-124): confidence is only added to the SQL SET list if confidence is not None. Passing no confidence argument leaves the column untouched: verified for the shared code path used by both the SQLite and PostgreSQL stores (same function, %s/%(name)s placeholder handled elsewhere).
• wiki.drafts.confidence is NOT NULL DEFAULT 0.5 on both backends (mcp_server/infrastructure/sqlite_schema_wiki.py:76, mcp_server/infrastructure/pg_schema.py:231) with a CHECK on the drafts table in the sqlite/pg draft-list projection (lines 29-30/187-188 apply to a different table but the drafts table itself is NOT NULL-constrained): draft["confidence"] can never be None at the float(draft["confidence"]) call site (wiki_refine.py:315), so the cast is safe. Confirmed the 0.4 synthesizer fallback: mcp_server/core/draft_synthesizer.py:277 (draft_confidence = sum(confidences) / len(confidences) if confidences else 0.4).
• Ordering: draft is fetched via get_draft before update_draft runs (wiki_refine.py:266), so draft["confidence"] used for the memo (line 315) is the pre-update value: correct here only because update_draft no longer touches confidence at all. Minor readability nit below.

2. Test strength (reproduced myself, not assumed)

• Ran UV_NO_SYNC=1 uv run pytest tests_py/handlers/test_wiki_refine_confidence.py -v on the PR branch: all 3 pass.
• Swapped in the pre-fix handler (git show origin/main:mcp_server/handlers/wiki_refine.py > mcp_server/handlers/wiki_refine.py) and re-ran: all 3 fail: test_refine_keeps_the_draft_confidence and test_refine_memo_records_the_draft_confidence both assert 0.85 == 0.5 ± 5.0e-07; test_a_refined_weak_draft_still_misses_the_approve_threshold fails with AssertionError: assert 'approved' != 'approved'. Matches the PR's stated fail-before/pass-after claim exactly.
• Restored with git checkout HEAD -- mcp_server/handlers/wiki_refine.py; git status confirmed clean, nothing left behind.
• Tests are deterministic and independent: each uses a fresh tmp_path SQLite DB via monkeypatch.setenv + reset_shared_store()/get_memory_settings.cache_clear() in a function-scoped fixture: no shared state between tests, no network, no time-of-day dependence.
• handler_refine has no callers anywhere in the tree outside this new test file (grep -rn "handler_refine\b"), confirming the PR's "not registered as an MCP tool" claim is not just documentation: it is literally true of the whole call graph.

3. Accuracy of factual claims

• git log -S wiki_refine --oneline -- mcp_server/tool_registry_wiki.py 'mcp_server/tool_registry*.py' returns no commits, and grep -rn "wiki_refine_draft\|wiki_get_draft" mcp_server/tool_registry*.py returns nothing: the "never registered" claim in the ADR, PR body, and issue comment is verified, not asserted.
• Thresholds MIN_CONFIDENCE_APPROVE = 0.6 / MIN_CONFIDENCE_HOLD = 0.4 (mcp_server/core/draft_curator.py:15-16) match every reference to them in the ADR/issue/PR body.
• 0.4 synthesizer fallback confirmed at draft_synthesizer.py:277 as cited.
• ADR wiki page and its docs/adr/ mirror are byte-identical in content (diffed both), and wiki/manifest.json gained the matching ADR-1065 entry.

4. Project gates (run from the worktree, UV_NO_SYNC=1)

• uv run ruff check: "All checks passed!"
• uv run ruff format --check: "1455 files already formatted"
• uv run python scripts/check_craftsmanship.py --base origin/main: "Craftsmanship gate: OK"
• uv run pytest tests_py/handlers/ tests_py/core/ -k "wiki or draft or curat or adr or manifest": 723 passed, 4471 deselected: matches the PR body's number exactly.

Move 0: ledger/seen-defect check

No completion ledger is embedded (not required by this project's PR template) and no seen-defect rationalization pattern ("unrelated," "pre-existing," "out of scope" without a filed issue) appears in the diff, PR body, or issue thread. The one open item: Path B unregistered: is explicitly deferred to a filed issue (#579), which is exactly the acceptable form under §14.3. No refusal triggered.

Layer check (Move 1)

Single file changed in mcp_server/handlers/; no new imports added or changed. No boundary crossed.

SOLID (Move 2)

No new branching, no interface change, no new type dependency. N/A: this is a value-passing removal, not a structural change.

Wiring & contract drift (Move 3)

No new public symbols. No signature change to handler_refine (still (args) -> dict). update_draft's confidence parameter was already optional (confidence: float | None = None) before this PR: its removal from this call site is exactly the change described, not a drift.

Complexity (Move 5)

No smell: a one-line removal and a one-argument value change.

Security & hygiene (Move 6)

No security-relevant surface touched. Commit message is conventional (fix(wiki): ..., references #578).

Non-blocking nit

• mcp_server/handlers/wiki_refine.py:266,315: draft is captured before update_draft runs, and the memo's confidence is read off that stale-by-construction snapshot. It happens to be correct only because update_draft no longer writes confidence at all; a future change that re-adds a confidence write to update_draft in this call would silently desync the memo from the stored value. Consider re-fetching the draft (or reading update_draft's return value, if it carries one) after the update, or a one-line comment noting the invariant this relies on.

Verdict

APPROVE. Fix is correct on both backends, the fail-before/pass-after claim reproduces exactly, tests are deterministic and independent, all cited facts (thresholds, "never registered," 0.4 fallback) verified against the code and git history, and all four project gates are green.