chore(metrics): measure realized delta-scoping savings and findings parity (#2164)

[bdfinst]

Part of #2200. Blocked on slice 0.

Question

Does the verdict ledger (#2166) plus delta-scoped dispatch (#2167, PR #2199) actually skip re-review at the /build Step 6 backstop and on repeat /code-review runs, without losing any findings?

Measure

• Skip rate: skipped (lens, file) pairs ÷ candidate pairs, per backstop and per repeat run, taken from verdict_scope.py output (or the slice-0 log). Compare against chore(code-review): measure re-review duplication across build checkpoints and repeat runs #2165's theoretical ceiling of 20.69%.
• Tokens avoided: skipped pairs priced with cost_meter's per-lens average. Use the byte-estimate method; chore(code-review): measure re-review duplication across build checkpoints and repeat runs #2165 documented that the "measured" leg over-counts.
• Findings parity (the safety gate): replay 3 real rounds, each on the same diff and commit, once with the ledger on and once with it disabled (move the file aside), and diff the actionable findings (finding_signature.is_actionable) against the ledger-on run.
• Ledger health: rows written per dispatch, plus the number of missing-scope-marker events.

Acceptance criteria

• Numbers are posted to chore(metrics): evaluate realized benefits of epics #2164 and #2172 #2200 against the slice-0 thresholds.
• Any actionable finding that the ledger-off run finds and the ledger-on run misses is a regression. File it as a bug; do not average it away.
• If the skip rate is near zero, state the cause (checkpoint dispatches not writing rows, path canonicalization, or content really changing) rather than just reporting the number.

---

Generated by Claude Code