CL-6379: Turn event pipeline — serialize collector events, classify terminal flips

[TheGreatAxios]

Problem (CL-6379)

On a fresh stack, every chat turn logged:

• WRN hub·event-collector: Dropping 'text'/'step-start'/'step-finish' part: no active turn for session 'ses_…'
• ERR hub·lookups: Ignoring terminal event for run 'turn__1': it does not belong to source deployment 'run_…'

and an agent's turn could be left permanently "running" — stuck busy status, no clean termination, live output lost.

Root causes

1. The event collector processed events concurrently. The registry's dispatch fires collector.onEvent(event) without awaiting (deliberately, so the websocket loop never blocks), but the collector had no internal ordering. An onTrigger section's end-of-turn burst (inference.done, connector.reply, next inference.start back-to-back) interleaves across the collector's DB awaits:

• connector.reply's finalize nulls currentTurnId while inference.done is still inserting parts → parts dropped as "no active turn", and persisted ordinals come out scrambled (a real run showed text at ordinal 0, step-start at 1, step-finish dropped).
• A finalize processed while the NEXT turn's beginTurn insert is in flight marks the new turn finalized at birth → its own connector.reply short-circuits and the inference_turn row stays running forever (permanently busy status, idle-sleep blocked, "typing dots" that never resolve).

Fix: the collector chains onEvent/abandon through an internal tail promise, so events settle in wire order while dispatch stays fire-and-forget.

2. The terminal-event flip misclassified occurrence child runs. receiveWorkflowRunPack treated any newly-terminal run id without a workflow_run row anchored on the source deployment as an ownership violation. A section occurrence's child run (turn__<n>, CL-6329) is repo-local by design — it never mints a row — so every completed chat turn logged a spurious ERR, indistinguishable from a real violation. Fix: the flip classifies through the new pure decideTerminalRunFlip: repo-local ids skip quietly, a minted run_<hex> id with no row still reports loudly, a row anchored elsewhere is still refused.

Verification

• Red/green: vendor/intx/hub-sessions/src/event-collector-registry.test.ts dispatches two full turns fire-and-forget and asserts every part persists in order and both turns finalize (failed before the fix exactly as in production); decideTerminalRunFlip classification tests in hub-session-lookups.test.ts.
• Live proof on a scratch stack (real hub + sidecar + Ollama, two messages into a fresh chat): before — both WRN drops and both turn__N ERRs, second inference_turn stuck running, parts out of order; after — zero WRN/ERR, both turns completed, parts step-start/text/step-finish at ordinals 0/1/2.
• vendor/intx/hub-sessions, packages/chat, packages/folded-runs, packages/insights test suites green; repo typecheck green; VENDORED.md records the hub-sessions local modifications.