Calibrate optimization candidate scoring with raw-free funnel evidence

[alexandrefimov]

Optimization Candidate scoring should be calibrated from structured analyzer facts and raw-free optimizer funnel evidence, not by loosening prompts, validators, or browser wording to increase draft count.

Scope:

• measure the optimizer funnel on fixture and sanitized real corpora with raw-free counters;
• separate expensive query, stats/query-context signal, recipe detected, safe to attempt, trusted draft, no-rewrite, recommendations-only, and no-draft reason;
• use structured analyzer facts where available, with rendered markdown only as compatibility fallback;
• keep no-recipe review tracks as telemetry and backlog routing, not recipe detection or proof of benefit;
• preserve strict SQL validation and the no-echo browser trust boundary.

Acceptance criteria:

• A raw-free calibration output records candidate, recipe-detected, draft-ready, trusted-draft, no-rewrite, recommendations-only, and failure/no-draft reason counts.
• Any scoring or cap change is backed by focused fixtures and regression tests.
• Browser/trusted-output tests prove no raw SQL, identifiers, predicates, local paths, artifact filenames, model names, or subprocess output leaks through optimizer surfaces.
• The calibration distinguishes recipe coverage limits from model/prompt failures.
• Documentation/changelog drift is checked and updated if scoring semantics or visible optimizer states change.

Maintainer notes:

This is maintainer-owned because it can change product ranking, optimizer trust semantics, and safety-sensitive visible states. Do not open a PR without agreeing on the calibration dataset shape and raw-free output contract first.