chore(rc.11): test-infra fixes captured during integration sweep

[branarakic]

Summary

Three commits pulled forward from the release/v10.0.0-rc.11 integration
branch — pure test-infrastructure fixes that surfaced while running the
RFC-38 sweep + targeted devnet scenarios against the merged rc.11 PR
universe. None of them touch production code; they make the existing
test surface honest about what it's checking.

Commits

1. ecf2bdd3 fix(rc.11): lu8 sweep script — replace unbound OK_FLAG with EXPLICIT_OK
   scripts/devnet-test-rfc38-lu8.sh summary block referenced an
   $OK_FLAG variable that was never defined; the script ran with
   set -u and crashed in the LAST line after all real assertions had
   already passed. Round-1 of the integration sweep counted that as
   1 FAIL out of 11. Fix: print $EXPLICIT_OK (the variable the
   scenario actually sets) instead. Round-2 sweep then went 11/11
   without changing any actual LU-8 behaviour.

2. c38b04df test(rc.11): two new devnet scenarios for async-promote + shutdown
   Two targeted scenarios that the existing RFC-38 sweep didn't cover:

   • devnet-test-rc11-promote-crash-recovery.sh —
     enqueues a 200-triple promote on node 5, tight-polls for
     state=running (worker claimed the lease), SIGKILLs the daemon,
     restarts node 5 via devnet.sh restart-node 5, asserts the
     post-restart job state is NOT in a forbidden RFC §6.2 state
     (no running → queued demote; no spurious abandonment of a
     still-valid lease). Validates the integration plumbing of the
     async-promote stack (PRs feat(publisher): TripleStoreAsyncPromoteQueue library (PR 1/4) #657 → feat(agent,daemon): wire async-promote queue + 5 HTTP routes (PR 2/4) #660 → feat(daemon): PR #3 — async-promote worker supervisor + lifecycle wiring #665 → feat(daemon): PR #4 — async-promote queue config knobs + E2E test + SKILL.md docs #667) survives a
     hard process kill. Requires the /api/assertion/<name>/promote-async
     route from PR feat(agent,daemon): wire async-promote queue + 5 HTTP routes (PR 2/4) #660 — will be exercisable as soon as that PR
     lands; sits dormant until then. Pass-class matrix in the script
     header documents every acceptable outcome.

   • devnet-test-rc11-shutdown-mid-publish.sh —
     node 5 (edge) launches 5 concurrent publishes against cores 1–4
     to saturate StorageACK reads, then SIGTERMs core node 2
     mid-flight. Measures shutdown wall-clock and greps daemon.log
     for new [shutdown-timeout] lines. Classifies the interaction
     between PR fix(cli/daemon): hard timeout on graceful shutdown to recover from cleanup deadlocks #655 (15s hard shutdown cap) and PR feat(core): plumb AbortSignal through DKGNode.stop() into protocol-router reads (PR-6) #669
     (AbortSignal through DKGNode.stop()) — green when shutdown
     completes <15s with zero new [shutdown-timeout] lines (PR feat(core): plumb AbortSignal through DKGNode.stop() into protocol-router reads (PR-6) #669
     is the active drain path; PR fix(cli/daemon): hard timeout on graceful shutdown to recover from cleanup deadlocks #655 stays passive). Uses
     node-based timing (now_ms()) so it runs identically on
     macOS BSD date and Linux GNU date. Requires PRs fix(cli/daemon): hard timeout on graceful shutdown to recover from cleanup deadlocks #655 + feat(core): plumb AbortSignal through DKGNode.stop() into protocol-router reads (PR-6) #669
     at runtime for the GREEN classification; will exit with a
     SOFT-PASS or FAIL classification on stock rc.10 so it can also
     serve as a pre-merge sanity check.

What's not in this PR

A third fix from the integration branch — 36fe4c91 fix(rc.11): align async-promote e2e test with PR #660's 200 ship contract — cannot
cleanly cherry-pick onto main because the file it touches
(packages/cli/test/async-promote-queue-e2e.test.ts) is introduced by
PR #667, which hasn't landed yet. Once #667 lands I'll open a tiny
follow-up PR (or @branarakic can absorb the 1-line change directly into
#667). The fix is:

-    expect(status).toBe(202);
+    // PR #660's wiring ships HTTP 200 on enqueue success, which the
+    // 24 sibling tests in promote-async-routes.test.ts encode as the
+    // de-facto API contract. Aligning with the ship contract until a
+    // future RC can do the 200 → 202 (HTTP-canonical for async
+    // accepted) breaking change for external clients.
+    expect(status).toBe(200);

Test plan

• scripts/devnet-test-rfc38-lu8.sh exits 0 with the same actual
  LU-8 assertions as before (no behavioural change).
• scripts/devnet-test-rfc38-all.sh was 11/11 PASS on the
  integration branch after this fix.
• devnet-test-rc11-promote-crash-recovery.sh: PASS on the
  integration branch (job state=succeeded post-restart, integration
  plumbing intact).
• devnet-test-rc11-shutdown-mid-publish.sh: GREEN on the
  integration branch (5459ms shutdown, zero new
  [shutdown-timeout] lines).

Why now

These fixes were authored during the rc.11 integration sweep
(release/v10.0.0-rc.11 branch — informational, not for merge). The
runbook docs/runbooks/RELEASE_RC11_INTEGRATION.md on that branch
captures the full sweep evidence. This PR pulls forward the
behaviour-neutral test-infra parts so they land on main and end up
in the rc.11 tag — no other rc.11 PR carries them.

Made with Cursor

[github-actions]

🔴 Bug: devnet.pid is the foreground supervisor, not the inner daemon worker. Sending SIGKILL only to this PID can leave daemon-foreground-worker alive, so the old process may keep the API port/store open and the restart path no longer reflects a real crash. Kill both devnet.pid and daemon.pid (or the whole process group) before asserting the node is down.

[github-actions]

🔴 Bug: this exits 0 on the exact path where the crash/restart scenario was never exercised (state=succeeded before the kill). In automated validation that becomes a false green for the recovery logic. Either retry until you actually observe running, or exit non-zero so callers can distinguish an inconclusive run from a pass.

[github-actions]

🔴 Bug: /api/shared-memory/publish does not accept a SPARQL-shaped selection object here; the route only supports "all" or a root-entity list. At runtime this request never creates publish load, and the surrounding || true hides the failure, so the script ends up timing an idle shutdown instead of shutdown under StorageACK pressure. Use an isolated CG/subgraph plus selection: "all", or resolve the root entities client-side and pass them as an array.