feat(brains): publish the generated flow skills and recut brains-build [BRNS-CORE-011] [BRNS-CORE-012]

[olegshmuelov]

What this delivers

Publishes the three flow skills (brains-board / brains-automation / brains-workflow) and the recut brains-build, rendered from ssvlabs/brains 864c45fa (stamped as the manifest's source_commit). Version 2.7.1 → 2.8.0 — hosts update on version, so the bump is the delivery. brains-write is re-rendered with quoted frontmatter (monorepo #1245).

Closes BRNS-CORE-011 and BRNS-CORE-012.

Generated, not authored here

The five skill artifacts are byte-identical rendered copies of the monorepo's capability catalog; the manifest digest-pins each one. Do not hand-edit them — fixes belong upstream and arrive by regeneration.

Contract-test changes

Skill-set pin 7 → 10; trigger-owner disjointness asserted in both directions from the authored trigger arrays; the secret-hygiene rails asserted positively (three rule sentences); the leak denylist narrowed to true internal identifiers — automation_secret and telegram_push are registered tool names the automation skill must teach, not leaks.

Verification

Contract suite green with a real claude plugin validate --strict; delivery guard passes 2.7.0 → 2.8.0; mutation proofs (rail deletion, trigger collision, unbumped version) each fail the intended assertion. Codex review of this diff: clean on the delivery itself; four findings in the generated procedure content filed upstream as BRNS-CORE-057 — that content is already live in production on the web-agent surface, so it does not gate this delivery.

[nir-ssvlabs]

LGTM on the delivery itself — but there's a merge-ordering hazard with #22 that isn't visible from inside either PR, so please read the last section before merging.

Verified rather than assumed (the invariant here is "these bytes are the monorepo's, unmodified"):

• All five artifact digests match their bytes. I recomputed sha256 over each artifact_path and compared to the declared artifact_sha256 — 5/5 match, including the two re-rendered files (brains-write, brains-build), so nothing was hand-edited and the digest wasn't refreshed to cover an edit. source_commit 864c45fa is a real commit on the monorepo.
• The version-only manifest diff is correct, not an omission. Skills are directory-discovered on both clients — the Claude manifest has no skills key and the Codex manifest points at ./skills/ — so three new skill directories ship without being enumerated. Worth stating because "3 skills added, manifests only changed a version string" is exactly the shape that would be a reaches-nobody bug if either client required a manifest list.
• 2.7.0 → 2.8.0 on both manifests, aligned and forward, and minor-shaped matches the convention #22 documents (feat(…) → minor).
• Contract pin correctly enumerates all ten skills alphabetically, with the three new ones in the digest map.

⚠️ Merge order matters, and getting it wrong silently breaks #22 rather than this PR. Both PRs branch from main at 2.7.0: this one goes to 2.8.0, and #22 (the claude.ai capture-claim fix) goes to 2.7.1.

• #22 first, then this — 2.7.0 → 2.7.1 → 2.8.0, both strictly forward, both guards pass.
• This first, then #22 — #22 becomes 2.8.0 → 2.7.1, a downgrade. Its delivery guard fails by design (that guard compares SemVer precedence precisely so a downgrade can't slip through), and if anyone waves it through, #22 publishes a version below the live one — so hosts stay on 2.8.0 and the capture-claim fix reaches zero users while looking merged. That's the exact failure mode the guard was written for.

So: merge #22 before this one, or if this lands first, #22 needs a re-bump to 2.8.1 before it can go. No change needed here either way — this PR is correct standalone; the constraint just isn't observable from either PR's own checks, since each is green against a main that still reads 2.7.0.

Approving.