chore(release): 0.120.0

[drewstone]

Why this release exists

main declared 0.119.0. The registry already holds 0.119.0 — with different, consumer-visible peers:

	@tangle-network/agent-eval peer range
0.119.0 as published	>=0.139.2 <0.140.0
main before this PR	>=0.140.1 <0.141.0

#697 moved the cohort to agent-eval 0.140.1 / agent-knowledge 7.0.4 but bumped no version and added no changelog entry. publish.yml skips a version already on the registry, so re-tagging v0.119.0 cannot work. Until this bump, the fix could not reach a single consumer.

What the cohort move fixes

src/durable/spawn-journal.ts writes each supervision event wrapped in an envelope — {kind:'event', root, event}. Through agent-eval 0.140.0, parseSupervisorTree only understood the flat dialect, so it read kind as the literal string event, matched no tree event kind, and reported zero spawns for every journal this package produces. agent-eval 0.140.1 adds readJournalRow, which unwraps the envelope and tags the dialect runtime-envelope.

agent-knowledge moves 7.0.3 -> 7.0.4 in the same cohort because scripts/verify-packed-cohort.mjs asserts knowledge's own agent-eval dependency exactly equals the packed agent-eval version.

Contents

• pnpm run release:prepare 0.120.0 — package.json, the docs/canonical-api.md version banner, docs/api/primitive-catalog.md, and both testing fixtures.
• CHANGELOG.md — a 0.120.0 section covering the cohort move, the peer-range table, and why it needed its own version. The Unreleased release-check timing entry moves into it.
• Drive-by: deletes a stray >>>>>>> origin/main conflict marker committed into the 0.110.0 changelog section by feat(supervise): add recursive product control and durable identity #657. The repo's pre-push merge-conflict-markers check flags it on any branch that touches this file.

Verification

• pnpm run lint — exit 0
• pnpm run docs:freshness — OK — no drift detected, version pin + 3 substrate peers present and accurate
• pnpm run verify:bench (the step that stranded v0.118.0) — run locally; bench/src/swe-arena/factory.test.mts passes 22/22 in isolation. Two of its cases run real pytest subprocesses under a 5s timeout and time out only under parallel load on my machine. CI's agent-bench job is the arbiter.

No npm publish here — v0.120.0 gets tagged on the merged sha and CI publishes via trusted publishing.

[tangletools]

✅ Auto-approved drewstone PR — d2c72c18

This PR was opened by the trusted drewstone account.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: drewstone_author · 2026-08-01T23:28:25Z}