[bug] wiki_refine_draft sets every refined draft's confidence to 0.85, lifting it past the curator's approve threshold #578

What happened

wiki_refine_draft (handler_refine, mcp_server/handlers/wiki_refine.py:300) overwrites the draft's confidence with a fixed 0.85 every time an LLM submits refined prose, and records the audit memo with the same 0.85 (line 315). No source backs the value: ADR-0467 preserves the file as it was during #514 and records no threshold choice.

The template synthesizer sets a draft's confidence to the mean confidence of its source claims, or 0.4 when no claim carries one (mcp_server/core/draft_synthesizer.py:271-277). The curator uses that number as a gate: below MIN_CONFIDENCE_APPROVE = 0.6 a draft gets a reason and can at best be held, below MIN_CONFIDENCE_HOLD = 0.4 it is rejected (mcp_server/core/draft_curator.py:15-16, 99-110, 138-141). wiki_curate then moves approved drafts toward wiki_compile.

So rewriting a draft's prose lifts it above the approve threshold whatever the evidence behind its claims. A draft synthesized at 0.4 from weak claims clears the confidence check of evaluate_draft once refined, although no claim changed.

Reproduction steps

1. Run wiki_extract then wiki_synthesize on any store; pick a pending draft and note its confidence.
2. Call wiki_refine_draft with that draft_id and a new lead.
3. Read the draft again: confidence is 0.85, and evaluate_draft no longer reports "below approve threshold" for it.

Expected behavior

Refinement changes the wording of a draft and leaves its confidence as the source claims set it. The audit memo records the draft's confidence as it stands, not a constant.

Evidence required to close

A test on the SQLite backend that refines a draft whose confidence is below MIN_CONFIDENCE_APPROVE and asserts that its confidence is unchanged, that the refined_llm memo carries that confidence, and that evaluate_draft still reports the approve-threshold reason.