docs(crew): refresh shipped state, add anti-drift rule, document the Gate/E2E-shard gap

[devin-ai-integration]

Summary

Docs only. No code, no CI config, no runtime behavior change.

docs/crew/STATE.md calls itself "the code truth for what is built today" and IDENTITY.md:39 sends every agent there, but it still claimed the 0.0.6 thread-worktree branch was unmerged and crew-v0.0.6 unpublished — four releases stale. Agents sequence work off that file, so the drift propagates into plans. This PR makes it true again, adds a rule so the next drift is a review finding, and writes down what a green NuncioCrew Gate does and does not prove about E2E.

Follows the approved plan for #117 (plans/20260810-docs-truth-gate-audit/, on PR #123's branch), phases 01 and 03.

STATE.md corrections — every claim re-checked against live repo state on origin/main @ 35af74019:

Was	Now	Source
"The branch is not merged, crew-v0.0.6 is not published, the 0.0.5 → 0.0.6 updater relaunch remains pending"	releases published through crew-v0.0.9 (2026-08-07, Latest); the 0.0.6 line merged and released	gh release list
"No crew-v0.0.6 tag or public 0.0.6 artifact has been created"	same correction in the test-gate section	gh release list
Buzz pin 0.5.3 @ 3a96acea…	0.5.7 / desktop-v0.5.7 @ f167818d…, pointing at upstream-buzz.json as the machine-readable source so the line cannot drift again	docs/crew/upstream-buzz.json
Settings displays v0.5.3 · Local	v0.5.7 · Local	desktop/package.json:4, desktop/src-tauri/tauri.conf.json:4
attention/recovery line absent	merged through #108 (6793c86da), #113 (304173e42) and #114 (35af74019)	git log origin/main
roles track absent	issue #116 is the head, PR #120 open and in flight — named as open, not shipped	gh pr list

The updater obligation was not silently dropped: no signed install-and-relaunch has been verified on any release pair in the repo's evidence (docs/crew/verification/ has no such record; RELEASING.md:160-170 describes it as a post-merge publication gate), so STATE.md now says that plainly instead of naming a superseded 0.0.5 → 0.0.6 pair.

Anti-drift rule — one checkbox in the implementation checklist agents already read, plus a decision entry. This PR is itself the rule's first exercise: it changes shipped state and updates STATE.md in the same PR.

Gate / E2E-shard audit — docs/crew/verification/0007-gate-e2e-shard-relationship.md (new) and a CI.md correction. The issue asked whether the Gate ignores the smoke shards or reported on a different run. It is the first, by design and contract-asserted — so this is not a CI bug report:

nuncio-crew-ci.yml:248-251   in-line rationale: advisory until flakes are triaged (#36/#37)
nuncio-crew-ci.yml:258       continue-on-error: true
nuncio-crew-ci.yml:320       needs: [changes, desktop-fast, desktop-rust, macos-arm, project-relay, buzz-acp]   ← smoke absent
check-nuncio-crew-ci-results.mjs:6-12   JOB_RELEVANCE has no smoke entry
nuncio-crew-ci-contract.test.mjs:150    assert.doesNotMatch(ci, /needs\.desktop-smoke-e2e\.result/)
nuncio-crew-ci-contract.test.mjs:155    assert.doesNotMatch(gateHelper, /desktop-smoke-e2e/)
ci.yml:295,306-307 (inherited, read-only)   upstream Buzz does the opposite

The actual defect was documentation: CI.md's job table never mentioned Desktop Smoke E2E at all, so a green gate read as "E2E passed". Fixed, mirroring the existing honest boundary at CI.md:66-70.

Related issue

Closes #117. Related: #109 and #110 (the two known-broken smoke lanes, not fixed here), #114 (source of the check-run evidence, already merged), PR #123 (the plan this follows), PR #112 (already closed by the founder — see DoD 3 below).

DECISIONS.md numbering: took D-031 and D-032, not D-028. origin/main's tail is D-027, but open PR #120 already uses D-028/D-029/D-030 and issue #122 proposes another D-028. D-031/D-032 are the next genuinely free numbers.

No upstream files touched (D-001 thin-fork budget: 0 upstream lines). .github/workflows/ci.yml is inherited and read-only. D-025 generic-ACP check: N/A — no wire contract, event kind, or engine-specific behavior introduced.

Definition of Done → change

DoD	Satisfied by
1. STATE.md refreshed, via a PR that follows the new anti-drift rule	the STATE.md table above; this PR changes shipped state and updates STATE.md in the same PR
2. Anti-drift rule in the AGENT-WORKING-AGREEMENT.md implementation checklist	new checkbox at AGENT-WORKING-AGREEMENT.md:87, recorded as D-031
3. PR #112 resolved	Already resolved before this session started — closed 2026-08-10T06:40Z with a comment naming the superseding issues (#116–#122). Reviewed against FOUNDER-PRODUCT.md, D-025–D-027 and #116 and I agree with the closure: #112's value was a sequencing index, and sequencing is exactly what the north star lock (#115) and #116 invalidated, while its status half is already historical (#105/#108, #113 merged). I did not reopen or revise it. Its one durable artifact — the channels.spec.ts:500 root-cause bisect, which corrects #110's own stated cause — existed only in the closed PR's body, so I copied it onto issue #110 to preserve the evidence.
4. Gate vs E2E-shard relationship documented with evidence	verification/0007-gate-e2e-shard-relationship.md + CI.md. The founder chose to keep the shards advisory until #109 and #110 close; that posture and its revisit trigger are now recorded in CI.md and as D-032. The gate wiring itself is unchanged.
5. Stale worktrees pruned	Owner action required — not actionable from CI/agent machines, acknowledged by the founder, who will run it on the macOS checkout. See below.

DoD 5, honestly: the 38 worktrees the issue describes (.worktrees/…, /private/tmp/crew-*, ~26 leaked buzz-ae-e2e-* temp entries) live in the founder's macOS checkout. On this session's Linux clone git worktree list shows exactly one entry and git worktree prune -v is a no-op, so there is nothing here to prune and no commit can carry this. The safe command to run on the founder's machine is git worktree prune -v (it only drops registrations whose directory is already gone; it cannot delete a live checkout or any branch).

.worktrees/bring-hermes-chat-into-crew must not be removed. The issue calls it merged-branch residue; it is not — its branch carries 6 commits with no equivalent on main or on #114's head. git cherry is unusable as the safety check here because squash-merge rewrites patch-ids (it flags even 17b4353bc, which demonstrably merged as #113). The recurring buzz-ae-e2e-* leak comes from desktop/tests/e2e/helpers/twoRelayHarness.ts:36, which does not clean up on aborted runs — that file is upstream-shared, so fixing the recurrence is a separate issue, not this one.

Testing

Docs-only; no tests apply to the content. Ran on the branch: pnpm --filter buzz check, pnpm --filter buzz typecheck, git diff --check — all clean. Docs-only paths mean the desktop jobs skip and the Gate accepts the deliberate skips (CI.md:11-13).

Gate audit evidence — concrete check-run, PR #114's main run 31362178966 on 35af74019:

Job	Conclusion	Completed
NuncioCrew Gate	success	06:38:19
Desktop Smoke E2E (1)	failure	06:44:05
Desktop Smoke E2E (2)	success	06:42:27
Desktop Smoke E2E (3)	failure	06:44:49
Desktop Smoke E2E (4)	cancelled (30m timeout)	07:00:04

The Gate reported green 22 minutes before shard 4 even finished — it does not wait on the shards, let alone consume them. Across the 8 desktop-touching NuncioCrew CI runs on main since 2026-08-07: shard 1 failed 8/8, shard 4 cancelled at the timeout 8/8, shard 3 failed 2/8, shard 2 passed 8/8, and the Gate succeeded 10/10. Shard 1 is channels.spec.ts:500 › channel question card accepts an answer (#110, failing since 25263120e/#96); shard 4 is the timeout lane (#109). Both are preexisting on main and out of scope here. Full per-run table in the verification record.

Link to Devin session: https://app.devin.ai/sessions/49189481f3f6406fb6ce1bdfa883620a
Requested by: @oscarlehuu

---

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no potential bugs to report.

View in Devin Review to see 1 additional finding.