feat(bin): promote a fleet learning to a binding check, add recurrence measurement

[pramendra]

Intent

Workstream 5 of the fleet engineering plan (data/fleet-engineering-plan.html), sequenced last because it depends on workstream 4 landing first. Build the mechanism that promotes a recorded learning (in data/learnings.md) into an actual binding check, refusal, or brief clause - never leaving it as prose alone - and measures recurrence: how many times the SAME defect class has shown up since the rule was written. A rule with recurring hits is a wish, per this fleet's own stated doctrine (data/learnings.md self-critically notes it reached 16 entries and 99% of its startup memory budget before being rewritten once - evidence that write-prose-and-hope-it-is-remembered does not scale). Done when: at least one existing learnings.md entry is demonstrably promoted to a real, testable check (not just referenced in a brief), and there is a mechanism - even a simple one - for tracking how many times that defect class has recurred since. Scope note: this depends on workstream 4's executable-contract mechanism; if workstream 4 has not landed when starting, say so plainly and either scope down to a design-only pass or pick a learning that can be checked without workstream 4's machinery. Acceptance criteria: the mechanism must be built and demonstrably work (show it, not just assert it); any check added must be proven able to fail by mutating what it guards and showing it go red; claim only what was measured, stating plainly what was NOT built and why scope was narrowed if it was. At task start, workstream 4 (fleet-ws4-executable-contracts) had not landed - it was live on a sibling worktree with zero commits ahead of main - so this task was scoped down accordingly: promote one learning checkable without workstream 4's machinery, rather than build a design-only pass or invent registry/drift-guard generality beyond what is needed for that one rule. Follow-up commit: the prior pipeline run's document-fix step accidentally committed .serena/project.yml and .serena/.gitignore (Serena MCP environment leakage into the worktree); this run additionally untracks those files and adds .serena/ to .gitignore so the PR does not ship that artifact.

What Changed

• Add bin/fm-learning-recurrence-lib.sh, a shared helper (introductions_for_path) that walks a guarded file's git history and counts commits whose diff adds a line matching a banned pattern, excluding comment lines and excluding rewrites of a pattern that was already present (only counts genuine (re)introductions).
• Add bin/fm-learning-recurrence.sh, a CLI wrapper with a --list/<rule-id> interface that reports introductions_total and recurrences_since_recorded for a promoted learning, splitting counts before/after the rule's recorded date; registers the first promoted rule, no-mistakes-cli-not-skill, guarding bin/fm-brief.sh against the literal /no-mistakes pattern.
• Promote that learning into a binding, mutation-proven regression test (tests/fm-brief.test.sh:test_ship_briefs_never_instruct_skill_invocation) that fails if any ship brief (no-mistakes, direct-PR, or local-only mode) instructs a worker to invoke /no-mistakes as a skill instead of routing through the no-mistakes CLI, and rewords bin/fm-brief.sh's generated brief text, AGENTS.md's Validate section, and .agents/skills/harness-adapters/SKILL.md accordingly.
• Add tests/fm-learning-recurrence.test.sh, which exercises introductions_for_path and the CLI wrapper against a fixture git repo (isolated from this repo's own history) to prove pre-rule introductions, fixes, post-rule recurrences, comment exclusion, and rewrite-of-still-present-pattern exclusion are all counted correctly.
• Document the promotion mechanism in docs/architecture.md, and untrack .serena/project.yml/.serena/.gitignore (accidental Serena MCP artifacts committed by a prior pipeline run) by adding .serena/ to .gitignore.

Risk Assessment

✅ Low: The change promotes exactly one learning (no-mistakes CLI vs. skill invocation) to a binding, mutation-provable test, adds a git-history-based recurrence counter with its own fixture-driven behavioral tests (including a CI-round fix for a real double-counting bug on reformatted lines), updates docs consistently, and separately cleans up an accidental Serena artifact leak; scope matches the stated workstream-4-unavailable narrowing and no risky or irreversible operations are introduced.

Testing

Ran the new fm-learning-recurrence test suite and the fm-brief suite (both fully pass), then manually mutated bin/fm-brief.sh to reintroduce the banned no-mistakes skill-invocation phrase and confirmed the promoted binding check goes red before reverting, demonstrating the learning is a real provably-failing check rather than prose, with the recurrence counter separately verified against a fixture repo and end-to-end against the real production rule and pattern.

Evidence: tests/fm-learning-recurrence.test.sh full pass

ok - fm-learning-recurrence-lib.sh: introductions_for_path counts additions only, excludes comments
ok - fm-learning-recurrence.sh: CLI wrapper measures the real rule and splits on its recorded date
ok - fm-learning-recurrence-lib.sh: rewriting an already-present pattern line does not inflate the recurrence count

Evidence: Mutation proof: reintroducing the no-mistakes skill-invocation instruction turns the binding check red

sed replaced the CLI-routing sentence in bin/fm-brief.sh with a literal /no-mistakes instruction; bash tests/fm-brief.test.sh then reported not ok on the affected test; mutation was reverted and git status --short confirmed clean

Pipeline

Updates from git push no-mistakes

✅ **intent** - passed

✅ No issues found.

✅ **Rebase** - passed

✅ No issues found.

✅ **Review** - passed

✅ No issues found.

✅ **Test** - passed

✅ No issues found.

• bash tests/fm-learning-recurrence.test.sh (3 of 3 pass: additions-only counting, recorded-date split against the real production rule, and rewrite-of-still-present-pattern exclusion)
• bash tests/fm-brief.test.sh (full suite including new test_ship_briefs_never_instruct_skill_invocation, all pass)
• Manual mutation: reintroduced the literal /no-mistakes skill-invocation phrase into bin/fm-brief.sh, re-ran tests/fm-brief.test.sh, confirmed it goes red (not ok), then reverted the mutation and confirmed git status --short is clean
• Confirmed .serena/project.yml and .serena/.gitignore are untracked from git history while .serena/ is now in .gitignore, matching the intended cleanup commit

✅ **Document** - passed

✅ No issues found.

✅ **Lint** - passed

✅ No issues found.

✅ **Push** - passed

✅ No issues found.

[greptile-apps]

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

_{Reviews (3): Last reviewed commit: "no-mistakes(document): Fix stale harness..." | Re-trigger Greptile}