dev CI red: managed_gc_scope_authority_unavailable in session-directory tests after #6337

[probepark]

Dev CI is red on dev 42d7e061: https://github.com/Yeachan-Heo/gajae-code/actions/runs/37230554481 (last green dev 22e47b0f: https://github.com/Yeachan-Heo/gajae-code/actions/runs/37227241418).

Attributed to #6337 (merged 2026-10-04T19:36Z). Full analysis with stack traces and blob comparison: #6337 (comment)

Failing tests:

• shard-4-of-8, packages/coding-agent/test/session-manager/session-directory.test.ts:
  • managed session write protocol > completes after a crash following transcript unlink with a durable retained artifact root (throws managed_gc_scope_authority_unavailable)
  • managed session write protocol > advances a persisted root-only artifact receipt after a fresh-scope replay
  • managed session write protocol > still acquires the lock when a completed cleanup receipt no longer binds its target identity
• shard-7-of-8, packages/coding-agent/test/session-resident-lifecycle.test.ts:
  • resident cache prune retention, lifecycle cleanup, and JSONL parity > cleans resident cache on session deletion, session-file switch, and close (SessionMigrationBusyError)

Between green and red, only packages/coding-agent/src/session/internal/managed-session-scope.ts changed among the code under test (blob 54ff1965 -> 1a793b49). The test files are identical. #6337's PR CI ran only the affected test files and did not include session-directory.test.ts.

Expected: both test files pass on dev again.

[probepark]

Triage: B (in progress). Kind: bug. Priority: P1. Dev CI has been red since 42d7e061, and coding-agent shards 4 and 7 fail on every dev head after the #6337 merge.

• Owner: @Yeachan-Heo opened the revert revert: #6337 managed logical owner cleanup fencing (breaks dev session-directory tests) #6343 (2026-10-04T21:00Z). The bisect in its body shows 22e47b0f at 110/0, 69311476 (fix(session): fence managed logical owner cleanup and cold-scope authority #6337 merge) at 106/4, and the revert at 110/0.
• Current blocker on revert: #6337 managed logical owner cleanup fencing (breaks dev session-directory tests) #6343: Affected path validation / plan stops at the stale-base guard, so the shards have not run on its head. Fix: rebase revert: #6337 managed logical owner cleanup fencing (breaks dev session-directory tests) #6343 onto current dev. feat(sdk): persist fenced logical owner deletion dispositions #6339 imports managed-session-scope.ts symbols, but the revert keeps them, and git merge-tree reports no conflict.
• No separate fix PR will be opened from this issue, because it would duplicate revert: #6337 managed logical owner cleanup fencing (breaks dev session-directory tests) #6343.
• Re-landing fix(session): fence managed logical owner cleanup and cold-scope authority #6337 is a follow-up for @snowykr: re-open it with session-directory.test.ts and session-resident-lifecycle.test.ts passing, and include the setProjectDir cleanup from c8643d9 (test/move-session-tool.test.ts:48-50).

Close condition: #6343 merges and Dev CI on the next dev head is green on shards 4 and 7.

[probepark]

gajae-merge-routine 2026-10-04T21:51Z: dev is RED again at a1588149 (https://github.com/Yeachan-Heo/gajae-code/actions/runs/37234274435). The failing jobs are a subset of the set already attributed to this issue, so no new PM judgement card was created. Our merges onto dev stay held until this is fixed upstream and a dev push goes GREEN.

[probepark]

dev CI is still red at 58e868b3 (run 37234297546). Verdict: regression, not flake. Evidence level: CODE (blob comparison against the last green run, 22e47b0f). Nothing was re-run locally.

• shard-5: the 3 managed session write protocol tests fail with managed_gc_scope_authority_unavailable. These are the same failures as 42d7e061, moved to a different shard. managed-session-scope.ts changed 54ff196575e6 -> 1a793b4999ea (fix(session): fence managed logical owner cleanup and cold-scope authority #6337), and session-directory.test.ts is unchanged at 2d531d2f175a.
• shard-8: resident cache ... cleans resident cache on session deletion fails, same as before. There are also 3 new failures in sdk-broker.test.ts (binds session.delete to the requested session header..., keeps transcript completion pending through canonical artifact reappearance, keeps retained transcript side authority pending...). They expect cleanup_pending and get invalid_input ("Managed task-artifact-owner authority could not be established for deletion." / "managed scope security could not be verified"). The test blob is unchanged at e1aa90a1022d. lifecycle.ts changed 462f9fad67ca -> d8e5bb938ca1 and broker.ts changed 10bbf1965367 -> ab076fc6d5d6, both from feat(sdk): persist fenced logical owner deletion dispositions #6339 (merged 21:00:11Z, built on fix(session): fence managed logical owner cleanup and cold-scope authority #6337). The new deletion path is lifecycle.ts:6475.

Note for #6343: it reverts only #6337. #6339 calls the owner-scope authority added by #6337, so verify that the sdk-broker.test.ts session.delete tests pass with #6343 applied before merging. If they don't, #6339 needs to be reverted or fixed too.

[probepark]

Status update (2026-10-05T00:30Z). Triage stays B (in progress). No separate fix PR will be opened from this issue.

• The fix path moved from the revert revert: #6337 managed logical owner cleanup fencing (breaks dev session-directory tests) #6343 to snowykr's fix-forward fix(session): establish actual authority for cold managed cleanup replay #6346 (head 34546327). The owner keeps revert: #6337 managed logical owner cleanup fencing (breaks dev session-directory tests) #6343 open only as a fallback until fix(session): establish actual authority for cold managed cleanup replay #6346 merges. Right now fix(session): establish actual authority for cold managed cleanup replay #6346 has 15 checks passing, 0 failing, and it is waiting for review.
• Dev is still red at 5552a9a2 (merge of feat(gc): retire logical task owners through immutable native journals #6341, run 37243777038). The failures are the same 7 as before: 3 managed session write protocol tests failing with managed_gc_scope_authority_unavailable, 3 sdk-broker session.delete tests, and the resident-cache deletion test, now spread across shards 1 and 6. The green run on c6827542 skipped the affected shards, so it does not show a recovery.

Close condition (updated): #6346 or #6343 merges, and Dev CI on the next dev head is green on the coding-agent shards that run these tests.

[probepark]

#6346 merged to dev at 2026-10-05T00:48Z (merge commit f4b3d1ad). It was approved at exact head 3454632, and its code-event Dev CI passed (run 37245072855).

The close condition is not met yet. Dev CI on f4b3d1ad still has to pass on the coding-agent shards that run session-directory.test.ts, session-resident-lifecycle.test.ts, and sdk-broker.test.ts. The 3 sdk-broker session.delete tests came in with #6339. The PR run did not exercise them, because its affected-path plan did not select sdk-broker.test.ts, so the dev run is the first real check of them. #6343 can be closed once that run is green.

[probepark]

Dev RED ba4c776 is a regression, not a flake: shards 3 and 4 fail deterministic tests whose blobs changed from the last full-shard green, and the changed SDK lifecycle/runtime code is in the #6339/#6344 merge range. The existing upstream tracker is #6342; no coder card (not an ACP failure).