Session: no single join for a turn's post-turn writers #2470

Nothing in a session joins all the writes that happen after a turn ends. Tests that measure "nothing changed after this turn" can see a late write, and quit or teardown have the same gap: there's no single point after which a finished turn will write nothing more. Low priority. Today this shows up as a rare test flake (TestAppIdleReleaseI1I2OpenViewKeepsItsFence, about 1 in 200 under load), when the setup turn's pace record is captured and published after the test starts counting.

Writers that can run after a turn has ended (app/v3 at 2c78de1):

1. Deferred turn-journal writes. internal/session/loop.go:971 defers journalTurnPace (appendPace, sessionfile.go:3277-3281), and loop.go:936/954 defers journalMarkRead. They run after sealTurn (loop.go:1835) has settled and announced TurnEnded (loop.go:1855-1856).
2. Naming. intentname.go:257-260 registers titleJobs, :288-290 writes intent_name, and the second-chance drain is at :305-307, :408 and :420, started from agent.go:2222. title.go:186-198 and :380 cover ordinary naming and publishing the title.
3. Memory and knowledge. loop.go:1365 starts learnFromTurn after EventTurnDone. memory.go:867-893 runs under memoryJobs. The standing scanner is at knowledge.go:129/251, and job compilation is at loop.go:348-352 and knowledge_usage.go:60/96/133.
4. Turn cleanup writes.
   • agent.go:2156-2283 hands back tasks, drains the queue and retires questions, and queues team traffic after closing done (:2235).
   • Team traffic is written at teamevent.go:92/126, joined at :184.
   • placemeta.go:126-138 SettleWrites joins six metadata writers.
   • Provider receipt accounting is tracked separately at usage_pending.go:31.
5. Capture scheduling. sessionfile.go:3449-3450 onWrite and loop.go:1863-1865 Tail.Wrote feed the trailing timer at internal/cellstore/sealwatch.go:176-189, which runs OnTail after a quiet period (:205-213). That can start a new capture after cellstore.WaitCapture has returned.

Why each existing signal falls short:

• EventTurnDone (loop.go:1355) and TurnEnded fire before naming, memory, pace and mark-read.
• Agent.WaitIdle waits on done, which closes at agent.go:2235 without joining titleJobs or memoryJobs, and before team cleanup.
• Turn stream closure (agent.go:2155, remote/server.go:1635-1651) covers synchronous defers but not titleJobs, memoryJobs or the trailing seal timer.
• Agent.SettleWrites covers its six owners only, not turn completion, naming or memory.
• waitForTitle (title.go:455) is a shutdown join with a timeout. WaitReceipts covers only accounting already admitted.
• cellstore.WaitCapture (pipeline.go:77-94) joins one existing capture and doesn't prevent a later one.
• SealWatch quiet is time-based, not writer completion. OnTurnEnd runs before pace.
• Close and StopWork cancel the work instead of letting it finish.

Proposed seam (not implemented): WaitPostTurnWrites(ctx, turn) in internal/session, next to startTurnLocked, reachable from tests through the test transport, since the agent runs in another process. It would, in order:

1. join the turn goroutine through its last cleanup;
2. join that turn's titleJobs and memoryJobs, plus any admitted receipt or knowledge work that writes to the journal;
3. drain SettleWrites;
4. flush SealWatch's pending tail through OnTail, so every capture has been admitted.

Only after that are WaitCapture and a peer-head comparison a sound "turn fully settled" check. The same join would give quit and teardown a definite end to a turn's writes. A wait for a single record kind, for idle, or for a quiet timer is not enough, because the next post-turn writer reopens the gap.

Found while fixing the idle-release tests (#2467). Low priority.

🤖 Generated with Claude Code