docs(skills): cue-eval-review — recorded-session review & replay-eval methodology

[xerhab]

What

Adds a Claude skill at .claude/skills/cue-eval-review/SKILL.md documenting how cue quality is reviewed and evaluated, so future sessions follow the same methodology that produced the 2026-07 precision cycles (#80, #81, #86):

• Data export — the docker exec Tenir-Postgres psql queries for segments and cues (host has no local psql).
• Hand review — interleave production cues into the transcript by at_ms, bucket bad cues into named failure classes (cross-generation confabulation, false corrections, paraphrase duplicates with the measured similarity bands, stale-topic, mishears), review-only until asked to fix. Includes the NFKC-normalization gotcha for regex probes.
• Replay eval — how to use scripts/cue_eval/ honestly: freeze the conversation set with controls, re-baseline on the current export, iterate subset→full, hand-read every iteration, match production decoding, treat stochastic effects as halved-not-fixed, and use the self-judge only comparatively.
• Threshold calibration — simulate gates over real production cues, hand-classify the dropped pairs, record the measured bands next to the constant.
• Ship & document — repo test/PR conventions plus appending an honest results section (including failed attempts and residuals) to the RESULTS doc.

Privacy note baked into the skill: deployment transcripts stay in analysis for the user and never get pasted into shipped code, prompts, tests, or docs.

Verification

Docs-only (no code paths touched). CLI flags and commands cited in the skill were checked against the current replay.py/judge.py argument parsers and the cues/segments table schemas in the deployment DB.