memory:promote — production-vs-fixture-precision-fix

[github-actions]

Summary

• Lesson distilled from: PR [mesh-scan] perf-regression: precision regressed (-13.3%) (#742) #748 — [mesh-scan] perf-regression: precision regressed (-13.3%) ([mesh-scan] perf-regression: precision regressed (-13.3%) #742)
• Lesson type: workflows/production-vs-fixture-precision-fix.md (new pattern)
• Source issue: [mesh-scan] perf-regression: precision regressed (-13.3%) #742 (perf-regression, mesh-scan generated)

What this lesson captures

Issue #742 (3x-disjoint cell) and Issue #744 (6x-confusable cell, see
[[no-hedge-single-peer-directive]]) are sibling -13.3% precision
regressions, not duplicates — and they were fixed at two genuinely
different loci:

• PR [mesh-scan] perf-regression: precision regressed (-13.3%) (#744) #750 fixed Issue [mesh-scan] perf-regression: precision regressed (-13.3%) #744 by strengthening the eval harness's synthetic
  caller prompt (eval/perf/harness.mjs buildRoutingMesh) — only affects
  that fixture's agent.
• PR [mesh-scan] perf-regression: precision regressed (-13.3%) (#742) #748 fixed Issue [mesh-scan] perf-regression: precision regressed (-13.3%) #742 by editing the real production tool
  descriptions in buildTools() (src/a2a/peer-bridge.js) — the
  list_peers/delegate_to_peer MCP descriptions every real deployed agent
  reads via the peer bridge, not just an eval fixture.

The lesson generalizes: trace a routing regression to its cause before
picking a fix locus — production tool descriptions when the bias is generic
(any worker can hedge/under-delegate), the eval fixture prompt when it's
specific to that eval's synthetic mesh.

It also flags two things PR #748 shipped with, per its own reviewer:

1. Opposing-instruction tension — the new list_peers text ("don't
   guess in favor of a loosely-matching peer") and the new
   delegate_to_peer text ("delegate even if you think you know the
   answer") pull in opposite directions for a loosely-matching task, risking
   swapping a precision fix for a recall regression (or vice versa) on a
   different eval cell.
2. No behavioral verification — the only new test asserts the
   description strings landed (assert.match), not that routing behavior
   actually changed. Confirming the regression is really fixed needs a real
   scripts/eval-perf.mjs re-run on the 3x-disjoint cell, checking precision
   and recall together.

Cross-linked [[no-hedge-single-peer-directive]] ↔
[[production-vs-fixture-precision-fix]] so either doc surfaces the other.

Validation

• node scripts/validate-quick-memory.mjs dev-mesh/curator/memory/quick.json → valid (no quick.json changes; workflow files only)
• quick.json entry count: 200/200 (unchanged — already at the hard cap; this lesson was routed to workflows/ instead for exactly that reason)

Review checklist

• Lesson is durable and general enough for future recall
• No code changes (memory root only)
• quick.json caps respected (200/200 entries, unchanged)
• Validator passed

[github-actions]

Triager classification — src/dev-mesh/classify.js precedence (infra > out-of-scope > flake > real_bug)

• label: infra_auth
• reason: no test/job ever executed. Both CI (run 28593349765) and dev-mesh-review completed with action_required and zero jobs — this is a GitHub Actions approval gate (triggering actor github-actions[bot], pull_request event withheld pending manual approval), not a code failure. There is no job log to read because nothing ran (ranAnyTest: false → infra_auth per the classifier, same branch as an auth/secret/network signature).
• action: escalate to a human — do not fix or re-kick. A re-run without approval will land back in action_required; someone with write access needs to approve the run from the Actions tab (or branch protection needs to stop requiring these checks for bot-authored PRs). This is already tracked as needs-human in needs-human: PR#772 stuck (not-clean:UNSTABLE) — non-required check unmet #778, so no new issue filed — flagging here as the classified checkpoint per the triage sweep.

No code change proposed; I don't edit code from this role.

[github-actions]

🤖 dev-mesh janitor: pushed an empty commit to nudge GitHub's mergeable recalc. If the state stays UNKNOWN after a few minutes, this PR likely needs a human.