ci(guard): enforce the committed-content self-containment guardrail + sweep breadcrumbs

[thejefflarson]

Summary

CLAUDE.md's "Committed content is self-contained — no external ticket IDs" section
(added in #301) was prose-only, not CI-enforced, so it kept drifting — recent PRs
(#305–#311 and older) kept adding JEF-nnn breadcrumbs into Rust/JS/YAML/shell
comments. This makes it real:

1. CI lint — engine/tests/self_containment_guard.rs, a cargo nextest guard
   alongside the existing file_size_guard.rs, walking the same directories
   directly (no git invocation). It fails the build on a JEF- prefix followed by
   a digit, or a bare linear.app substring, in tracked source
   comments/docstrings + docs (Rust/JS/JSX/Python/shell/YAML/Helm-tpl/CSS/Markdown
   under docs/, scripts/, charts/, engine/src, engine/examples,
   engine/tests, engine/web/src, engine/web/test, behavior/src, agent/,
   .github/workflows, plus CLAUDE.md, README.md, both Dockerfiles, and the
   web config/dist files).

   Boundary exemptions (per CLAUDE.md's documented scope):

   • Branch names, PR bodies, git history/commit messages never reach this
     filesystem walk — nothing to exempt, they're structurally out of scope.
   • CLAUDE.md's own guardrail text that NAMES the pattern (a `linear.app` URL)
     is allowlisted by exact substring — JEF-nnn placeholder text needs no
     allowlisting since the check requires a real digit after JEF-.
   • scripts/judge_bakeoff_cutchoice.py is excluded with a documented, narrow
     reason: it's under active edit on a concurrent branch landing alongside this
     one; a follow-up sweep clears it once that lands (DECISION NEEDED: flagging
     for the architect — this is the one deliberately-incomplete corner of the
     sweep).
   • The guard's own source file is exempted from itself (it has to name the
     literal JEF-/linear.app patterns to implement the check — the same
     self-reference CLAUDE.md's own guardrail text is allowed).

2. Sweep — every JEF-nnn/linear.app breadcrumb the guard would have
   flagged, across ~2300 occurrences in ~300 files, replaced with the ADR that
   records the decision, the module/mechanism name already sitting beside it, or
   the inline reasoning that was already there (the ticket number was almost
   always decoration on top of prose that already carried the why). A handful of
   spots needed an actual rewrite where the ticket number was structurally load
   bearing (a numbered citation, a "resolves NNN" tail, a bare compound
   JEF-A/B reference) — those got a real sentence, not just a deletion.

Testing

• no_ticket_breadcrumbs_in_tracked_source_or_docs FAILS on a synthetic
  JEF-9999 canary added to a scoped file, and PASSES on the swept tree (verified
  both directions before committing).
• npm --prefix engine/web run build && npm run lint && npm run typecheck && npm test: clean (build, 0 eslint errors, 0 tsc errors, 122/122 vitest passing — the sweep touched .jsx/.js test-description strings and dashboard.css comments, so I re-verified the web suite explicitly).
• cargo fmt --check, cargo clippy --all-targets -- -D warnings: clean, both for the engine workspace and the separate agent workspace (protector-agent + protector-agent-common; the eBPF crate needs the nightly+bpf toolchain and isn't locally buildable here — only its doc comments/vmlinux.rs annotations were swept, no logic touched).
• cargo nextest run --workspace: 1166 passed, 2 skipped (includes the new guard test and file_size_guard, confirming every edited file stays under the 1,000-line cap).
• cargo test for the agent workspace's buildable crates: 45 + 7 passed.
• helm lint charts/protector + helm template: clean (the sweep touched several chart template comments/values.yaml docs).
• Comment/doc-prose only — no code-logic change. Caught this the hard way: an early sweep pass mechanically stripped ticket citations from strings like describe("… (JEF-408)", () => { and clipped the adjacent () too, breaking 9 JS test files' syntax (eslint parse errors) and a handful of Helm template/workflow comments (dangling —/)/+-at-line-start artifacts, two of which tripped clippy's doc_lazy_continuation lint on accidental markdown list markers). All caught and hand-fixed before this diff; re-verified with a second full gate pass afterward.

Checks

• /soundcheck:pr-review: no Critical/High findings — pure comment/doc-prose sweep + one new filesystem-walking test with no untrusted input, network, shell exec, secrets, or auth/endpoint changes.
• /simplify (single-pass, Agent tool unavailable in this context): collapsed a one-element &[&str] exemption list + .any() iteration in the new guard down to a plain &str constant + direct .contains(). Left repo_root()/is_skipped_dir()'s duplication with file_size_guard.rs as-is — the existing test suite's own convention is per-file duplication (no shared tests/common module exists yet), and factoring one out would touch a file outside this diff's scope.

Scope notes / decisions

• Lint mechanism: a cargo nextest Rust test (matches the ticket's stated preference and the repo's existing file_size_guard.rs pattern) rather than a standalone CI grep step — keeps the check in the same "run the test suite" gate contributors already run locally.
• Scope beyond CLAUDE.md's literal list: I included .github/workflows/*.yml and both Dockerfiles even though CLAUDE.md's guardrail text doesn't name them explicitly (it says "ADRs, source comments/docstrings, CLAUDE.md, VISION.md, scripts, chart templates") — they're tracked, human-authored, tracker-blind prose in the same spirit, and both had breadcrumbs. Flagging as a documented scope call rather than asking, since under-covering the lint was the actual risk here.
• DECISION NEEDED: scripts/judge_bakeoff_cutchoice.py is still exempted (unswept) pending a concurrent branch some other in-flight work is editing — I was told to steer clear of that file. Once that branch lands, a small follow-up should sweep its 3 remaining JEF-nnn breadcrumbs and delete its EXEMPT_FILES entry in the guard.

Closes JEF-682

[thejefflarson]

Integration review (architect): verified this is a comment/docstring/doc-prose-only sweep with no code-logic change, re-verified the guard independently (fails on an injected JEF-9999 + linear.app canary, passes on the swept tree), and confirmed the scope extension to .github/workflows/*.yml + both Dockerfiles is sound (tracker-blind prose in the same spirit as CLAUDE.md's list).

Found and fixed one surviving over-eager-cleanup defect the sweep left: in agent/protector-agent-ebpf/src/vmlinux.rs the trailing comments // JEF-317, ON-NODE PENDING on the two offset_of! lines inside the const _: () = { ... } compile-time guard were mangled into /, ON-NODE PENDING — invalid Rust that would have broken the ebpf cross-compile (that crate is a separate workspace, so the nextest job wouldn't have caught it; only the ebpf job would). Restored the // markers.

DECISION (stale exemption): #314 is merged, so its temporary EXEMPT_FILES entry for scripts/judge_bakeoff_cutchoice.py was stale. Folded the 3 remaining breadcrumbs there (kept the CUT TRAP / REFUTE TRAP mechanism names; ADR-0034 already cited inline) and removed the exemption, so the guard now carries a single structural exemption (its own source file) plus the one EXEMPT_LINE_SUBSTRING for CLAUDE.md's guardrail-naming line. No fast-follow needed.

Branch updated to main (fast-forward, no force). Merging on green.