Add empirical review-findings writeup (evals/README.md)

[kimdzhekhon]

Summary

Adds evals/README.md, documenting what actually happened across this repo's review/execution work so far, in the style of Code-Review-Loop's own evals/README.md: a summary table, a "what this bought" bullet list, then per-round detail.

• Round 1 (static review): filed Secret suffix leaks into masked_preview/context_line for generic_high_entropy_assignment rule #2 (secret-suffix leak into masked_preview/context_line, a real violation of the tool's own "never print raw secret" guarantee), Finding ids collide across rounds — STILL_OPEN carry-forward can overwrite an unrelated finding's resolution #3 (finding-id collision across rounds), [Low confidence, needs verification] dedup fingerprint's rule_id may block cross-source dedup between builtin/gitleaks/trufflehog #4 (dedup fingerprint concern, explicitly filed as low-confidence/unverified since gitleaks/trufflehog weren't available to test it).
• Round 2 (deepened static review): filed mask() reveals up to 67% of a secret in plaintext for lengths 9-19 (violates design-spec §3) #5 (mask() over-reveals up to 67% of short secrets), fingerprint()'s rule_id lets the same secret produce duplicate un-deduped candidates (confirms #4) #6 (fingerprint dedup gap reproduced entirely within the builtin scanner — this is what turned [Low confidence, needs verification] dedup fingerprint's rule_id may block cross-source dedup between builtin/gitleaks/trufflehog #4 from a hypothesis into a confirmed bug; the same commit closes both), files_scanned counts files inside skipped directories (.git/node_modules/target/...), wildly inflating reported coverage #7 (files_scanned counts files inside skipped directories, inflating reported coverage by orders of magnitude in real repos).
• Round 3 (real CLI execution): actually ran claude -p --model haiku end to end (2 scans + 1 describe, real cost $0.7846) against a small AWS-key/GitHub-token test fixture. Found README masking formula is stale after #5's fix — still documents the buggy clamp(3, 6) #8 (README's masking formula was stale after mask() reveals up to 67% of a secret in plaintext for lengths 9-19 (violates design-spec §3) #5's fix — code was already correct, only the doc lagged) by cross-checking real output against the README, not by reading the diff. Also hand-verified: files_scanned=4 (correct), every masked_preview matches the post-mask() reveals up to 67% of a secret in plaintext for lengths 9-19 (violates design-spec §3) #5 formula, no scanned secret appears unmasked anywhere in report.md/state.json, and both BLOCK verdicts are consistent with the documented exit-code contract (BLOCK=3).

All 7 issues (#2–#8) are closed, each by a commit that fixed the actual root cause.

Test plan

• Every claim cross-checked against real issue bodies (gh issue view), fix commits (git log), and real run artifacts (runs/manual-test/, runs/exit-check/) — not written from memory.
• Doc-only change, no code touched.