Docs: record the repo-sentinel dogfood findings; C1f is on by default

[b-macker]

Summary

This is a documentation-only PR, following up #263, #264 and #265.

• docs/governance-campaign-findings.md: a new section records rounds 3–5 of the repo-sentinel dogfood, run on live Gemini. It uses the file's existing live-status labels, and every fix is labelled honestly: only coherence_adjustments has actually been seen working live.
  • coherence_adjustments (CDD: reconcile coherence from telemetry; analyze the response after a failed call #263): confirmed. 171 analyzed rows across 8 runs, none that fail to reconcile. That count comes from the dogfood's own reconciler; the reviewer rows of one run were re-derived by hand and match.
  • Analysis of the response after a failed call (CDD: reconcile coherence from telemetry; analyze the response after a failed call #263): stub-only. Round 5 did retry on 503s (73 retry or skip events), but no send was shown to exhaust its retries, which is the case the fix covers.
  • C1f: the defect is confirmed live (4 of 4 runs, identical to six decimals). The fix is now on by default (Default entropy_baseline_adaptive to true (C1f) #265), but no live run has used it yet, so it stays stub-only.
  • A result that did not reproduce: round 4 found the adversarial reviewer was caught earlier. In round 5 there were only two failed validations across eight runs, both in one run, so the fixtures can't yet tell injection from no injection.
  • Evidence destroyed before it was read: the project's run.sh deleted round 3's telemetry, and its findings cited extract files that were never written.
• docs/open-investigations.md C1f: status changes from "partially addressed; ships off" to "fixed; on by default since Default entropy_baseline_adaptive to true (C1f) #265".

Test Plan

No code changes, so no tests were run. The numbers above come from the round 5 handoff file, and each claim was checked against the raw telemetry in it.

🤖 Generated with Claude Code

https://claude.ai/code/session_01ELUfjXZvx8kzXo1UJjrAhC

---

Generated by Claude Code

[github-actions]

NAAb Governance Report

Metric	Count
Files checked	16
Passed	16
Failed	0

All governance checks passed!

_{Generated by NAAb Governance Engine v4.0}