Row 38 S3 round 2 gate. Darkwing.
Tree: detached worktree at c9ef7ee4, build-r2.patch applied, sha256sum -c build-r2-manifest.sha256 28/28 OK.
Suites run one after another, each teed to its own file.
Start 2026-10-09T00:18:27Z, done 00:23:15Z, load 7 to 15.

suite                                         result
node --test packages/tasks/tests/*.test.mjs   51/51
node --test packages/bus/tests/*.test.mjs     67/67
node --test packages/business/tests/*.test.mjs  60/60
node --test packages/discord/tests/*.test.mjs  173/173
scripts/test-auth.sh                          15/15
scripts/test-conductor.sh                     17/17
scripts/test-config.sh                        24/24
scripts/test-discord.sh                       64/64
scripts/test-extension-package.sh             18/18
scripts/test-foundation.sh                    44/44
scripts/test-queue.sh                         27/27 (canonical-root check skipped: scratch worktree)
scripts/test-release.sh                       11 pass, 3 fail
scripts/test-task.sh                          90 pass, 8 fail

Failures, all cases that call the live provider:
test-release: healthy activation succeeds; pointer written with valid fields;
  repeat activation succeeds (log grows).
test-task: live hello task succeeds with exact marker; result.json contents are
  correct; wrong expectExact (reason exit-nonzero); fork base: teach succeeds;
  fork child recalls ancestor context; forked child recalled ancestor code word;
  user recall run succeeds; recalled user name.

Cause. A sandboxed `release.sh activate` in the same tree at 00:25:13Z left a
run whose stderr.txt ends:
  429: {"code":"1308","message":"Usage limit reached for 5 hour. Your limit will reset at 2026-10-09 10:48:02"}
The zai account had hit its 5-hour usage limit. The health check and the live
test-task cases all run a real model turn, so all of them fail with
exit-nonzero. The mock-adapter cases in the same suite pass (mock, mission,
seat, policy, workspace and user-layer cases).

Base comparison. Clean worktree at c9ef7ee4, no patch, run twice
(00:23 and 00:26):
  test-release  11 pass, 3 fail, the same three cases
  test-task     67 pass, 31 fail, both runs
The base fails more because Docker refused to create its compose network:
  failed to create network base2_default: Error response from daemon:
  all predefined address pools have been fully subnetted
gate2_default already existed from the build, so the candidate's containers
started. The same mock task run in both trees at 00:26:23Z: base exit 1
(network error), candidate exit 0, response MOSAIC_HELLO_OK.
So the base run says nothing about the live cases, and I didn't remove
other networks to make room.

Neither suite loads packages/tasks or packages/bus. Round 1's gate had
test-release 14/14 and test-task 98/98 on the same scripts. Nothing under
scripts/ changed between bee89d10 and c9ef7ee4. I'll rerun both suites in the
candidate tree after the provider limit resets and add the result.
