chore(metrics): pre-register benefit metrics and close instrument gaps

[bdfinst]

Part of #2200. Gates slices 1–4.

Goal

Fix every metric, threshold and data window in writing before any post-merge data is read, and make sure each metric has an instrument that actually records it.

Tasks

• The table must fix a numeric keep/tune/retire line for at least: the delta-scoping skip rate (chore(metrics): measure realized delta-scoping savings and findings parity (#2164) #2202); the findings-parity rule (any miss is a regression); the abort's deferred-lens yield; the test-review true-positive change; the --expand rate; injection overhead in tokens per dispatch; injected-skill uptake; the completion-guard divergence rate and the false-positive rate that allows blocking; and the contradicted-verdict rate per invocation for source-verification.
• For every before/after comparison (chore(metrics): measure cheap-blocker abort, countable pre-phase and tiered findings (#2164) #2203, chore(metrics): measure skill-context injection uptake and completion-guard divergence (#2172) #2204), record the other merges, model changes and prompt changes that land in the window as known confounds. Those deltas are reported as directional, not causal.
• Post a table to chore(metrics): evaluate realized benefits of epics #2164 and #2172 #2200 with one row per mechanism: metric, instrument, baseline, keep / tune / retire thresholds, and the data window. Suggested window: ≥10 real /build or /code-review sessions, or 14 days after the latest merge (feat(code-review): scope backstop and repeat runs to the unreviewed delta #2199), whichever comes later. If fewer than 10 sessions exist after 28 days, close the window anyway and record a null result with the actual N.
• Audit each instrument against a real session:
  • review-verdicts.jsonl: are rows written for /build sub-step 4/6 dispatches, or only for /code-review? Count the missing-scope-marker boundary events.
  • ledgerSkipped / fullySkippedLenses: is the count persisted anywhere, or only rendered in a report?
  • checkpoint_abort.py: are abort outcomes (aborted, deferredLenses, and whether the deferred lenses found anything) recorded to a stream, or only printed?
  • subagent_skill_context.py: is there any signal that an injected skill was actually loaded? (Skill tool calls in the subagent transcript, read via scripts/lib/session_log.)
  • subagent_completion_guard.py: which boundary-event classifications does it emit?
• For each gap, add the smallest possible log line (a JSONL row, no behavior change) and document it in knowledge/telemetry-schema.md.

Acceptance criteria

• The metrics table is posted to chore(metrics): evaluate realized benefits of epics #2164 and #2172 #2200 before any slice 1–4 number is read. The comment timestamp is the proof.
• Each metric names an instrument that is confirmed to emit on ≥ 3 real sessions on the local machine, with those session IDs listed in the chore(metrics): evaluate realized benefits of epics #2164 and #2172 #2200 comment.
• Any new log stream is documented in telemetry-schema.md, and the full pre-push pytest directory list is green.

---

Generated by Claude Code