Cache unchanged AI review verdicts

[pierrick-fonquerne]

What changed

• store a versioned, invisible and 16 KB-bounded verdict cache in the team review comment
• skip every Mistral call when the diff and reviewer workflow revision are unchanged
• reuse three-lens verdicts for unchanged findings whose exact identity and lens context still match
• ignore corrupt or obsolete cache payloads and avoid caching incomplete model runs
• trust cache markers only when they come from the configured workflow-bot author

Why

Team review reruns currently repeat all specialist, synthesis and lens calls even when the review input is unchanged. This adds latency and API cost, and needlessly recomputes stable per-finding verification.

Impact

Identical reruns return after the GitHub diff/comment lookup. Partial changes can reuse safe per-finding verdicts while changed context is reverified. Existing visible review formatting is unchanged because the cache is stored in an HTML comment.

Validation

• cargo +1.88.0 fmt --all -- --check
• cargo +1.88.0 clippy --workspace --all-targets --all-features -- -D warnings
• cargo +1.88.0 test --workspace --all-features --no-fail-fast
• cargo +1.88.0 check --workspace --all-features --all-targets

Closes #11

[pierrick-fonquerne]

Implementation review complete: no blocking findings.

Validation evidence:

• all repository CI checks pass, including Rust 1.88 MSRV
• real smoke: first run wrote the trusted hidden cache after agents/synthesis/lenses
• identical rerun kept comment ID 5269995677, logged Identical diff and reviewer revision, and made no specialist calls
• corrupt/version/context/identity/size invalidation paths are covered by tests
• transient agent, synthesis, or insufficient-lens results are never cached

Temporary smoke PR: nubster-opensources/lightshuttle#317 (closed without merge).