Add the 2026-10-10 CCA performance report

[GiulioDER]

What and why

Adds docs/reports/2026-10-10-cca-performance-report.md, an analysis of origin/master at d2317de (release 0.10.1): what each pipeline layer costs, what evidence exists for its value, the tier selector routing most diffs to DEEP (44 of the last 60 commits here), safe changes S1 to S14, and A/B candidates A0 to A9, each with a falsifiable prediction.

The findings are unedited. A note dated 2026-10-11 at the top records that #60 fixes four of the gate defects in section 4 (close checks nothing, --status n/a passes, stale verdicts after REVISE, no-fix runs never close), and that they stay true of master until it merges.

The two read-only helper scripts behind the COUNTED numbers (count.sh, tier_replay.py) are committed verbatim beside the report. They lived only in a session scratchpad, which would have left those numbers unreproducible. Both read origin/master, so a re-run measures the branch as it is then; the note says to substitute the sha.

Checklist

• pytest -q: not run. The PR touches no code under cca_checks or tests.
• ruff check cca_checks tests is clean. tier_replay.py itself has four style findings (three E501, one E741); I left them, since the script is committed as the evidence that produced the numbers and CI does not lint docs/.
• n/a: no checker behaviour change
• n/a: no new auditor or claim type

🤖 Generated with Claude Code