fix(runtime-host): project queue changes that land while no root Turn is live

[ggbdpq]

Summary

Fixes #5520

A queue that drains while no root Turn is live was never projected: the session projector gated queue_update on root && queueChanged(...), and seedActive returned nothing for a rootless snapshot. The Desktop's queue mirror is driven only by queue_update, so the renderer kept a phantom queued card whose 删除/retract failed with not_found forever — and switching sessions back and forth did not clear it, because the reseed hit the same gate. The Host broadcast the projection frames all along (#publishCanonical fires on any canonical change); the drop was purely in the projector adapter.

The fix stays at that shared root:

• session-projector.ts now pushes queue_update whenever the queue changed, root or not, attributed to root?.turnId ?? previousRoot?.turnId ?? ''. The steering synthesis stays root-gated (it attributes messages to the live Turn).
• seedActive seeds the queue mirror for a rootless session too — the empty queue is exactly the signal that lets an observer clear a stale card on (re)subscribe.

No consumer change is needed: the Desktop already deletes messageQueueBySession[sessionId] when a queue_update carries empty arrays, and the TUI mirrors the arrays verbatim.

Verification

Check	Result
New regression tests (drain without root / seed without root)	red before the fix, green after
session-projector.test.ts (full file, rebuilt dist)	29/29
message-coordinator.test.ts	76/76
execution-host-queue.test.ts	16/16 — one "did not become ready" timeout on the first full-file run; passed single and on full-file rerun (spawn-heavy flake, not code)
session-subscription-client.test.ts	13 failures reproduce identically on a pristine upstream/main worktree (Windows transport read side ended) — pre-existing, untouched by this PR
biome format/lint on touched files	clean
npm run check:asf-headers	pass

Not run locally: the full runtime-host matrix (~40 min; CI covers it) and the issue's Desktop manual repro (reported on macOS; the fixed root cause and its regression tests are platform-independent).

AI use

Analysis, patch, and tests done with GLM-5.3-Flash (ZCode) under human review.

Checklist

• The fix lives at the shared root (projector), not in an individual caller
• Regression tests fail before the fix and pass after
• No protocol/event contract change — only the emission gating of the existing queue_update
• ASF headers intact; biome clean on touched files

[ggbdpq]

The CI red was a real regression in my first push, thank you for catching it — diagnosed and fixed at 5dac0705a.

What failed (3 tests, all Desktop-side consumers):

• runtime-host-desktop-candidate × 2 — TypeError: Cannot read properties of undefined (reading 'hostEpoch'): the observer constructs projectors from snapshots that can carry no queue at all, and my unconditional rootless seed fed that straight into projectQueueUpdate.
• runtime-host-session-observer "projects root lifecycle without fabricating content events" — this test pins the contract that a rootless seed emits []; my seed hunk fabricated a queue_update against it.

The fix: I reverted the seed hunk entirely (seedActive is back to returning [] for rootless snapshots) and kept only the accept-path change — which is where the #5520 phantom actually lives. The drain lands while the client is subscribed (session events route per session, not per visible view), so projecting it there clears the card whether or not that session is on screen; a client that never observed the session has no stale card for a seed to clear, which is exactly why the observer contract is what it is. The surviving regression test now documents that reasoning.

Verified locally after the fix: session-projector 28/28, runtime-host-session-observer 55/55, runtime-host-desktop-candidate 25/25 — the last two are the suites CI caught, which my local matrix had missed because the projector's consumers live in the Desktop workspace, not runtime-host. That gap is on me; the affected-workspace set is now part of my run book for this area.

[jackwener]

Finding

[P1][Normal user path] Reopening an idle Session still cannot clear a queue entry that drained while the Session was inactive.

The accept-path change correctly emits an empty queue_update when the drain arrives while the client remains subscribed. However, Desktop subscribes only to the active Session and unsubscribes when activeId changes (apps/desktop/src/renderer/app-shell-effects.ts). If the queued follow-up is consumed while that Session is inactive, there is no projector instance present to observe the drain.

When the user returns, the new projector is created from the current rootless snapshot, but RuntimeHostSessionProjector.seedActive() still returns [] as soon as rootTurn is null (packages/runtime-host/src/adapter/session-projector.ts). The renderer retains messageQueueBySession[sessionId] across navigation, and it removes that stale entry only after receiving an empty queue_update. Therefore the card remains visible and retracting it still returns not_found, which is one of the issue's documented normal reproduction paths. A direct probe against this exact head also returned an empty seed for a rootless snapshot with an authoritative empty queue.

Please make the rootless observation seed carry the authoritative queue state (or add an equivalent out-of-band queue snapshot) without violating the observer's no-content-event contract, and add a Desktop-level regression that switches away, drains the queue, then switches back. The PR description currently says that seedActive performs this rootless reseed, but that change was reverted from the final code.

Review summary

The continuously subscribed drain path and its focused projector regression are correct; the projector suite passed 27/27. Biome, ASF headers, diff checks, the hosted exact-head test check, and a clean synthetic merge onto current main all passed. Because the off-subscription production path still reproduces the reported phantom card, this exact head a5f8cb028957d8302f870beff409afb44e0d1ff3 is blocked. I did not approve or merge.

---

Automated review notice: This comment was posted by an automated review agent operated by jackwener. It is not an independent human review and does not replace one.

[hqhq1025]

Finding

[P1] The normal switch-away / switch-back path still leaves a phantom queued card. The change in packages/runtime-host/src/adapter/session-projector.ts:406-418 emits the authoritative empty queue when a subscriber stays connected through a rootless drain. Desktop instead unsubscribes from the old Session on navigation (apps/desktop/src/renderer/app-shell-effects.ts:400-458). If the queue drains while that Session is inactive, the new projector returns no events for its rootless snapshot (packages/runtime-host/src/adapter/session-projector.ts:161-168), while the renderer retains messageQueueBySession[sessionId] until it receives an empty queue_update (apps/desktop/src/renderer/app-shell-session-events.ts:303-323). The stale card therefore remains visible, and retract can return not_found. Please make resubscription convey the authoritative empty queue and cover switching away, draining, then switching back in a Desktop regression. I independently corroborate the same current-head issue already reported by jackwener, so I am not duplicating an inline comment.

Scope and readiness

This PR changes the projector drain emission and adds a focused continuously subscribed regression (2 files, +44/-2). I reviewed the projector, Desktop subscription lifecycle, queue state reducer, and the new test. I found no other substantiated P0–P3 issue in that scope. The current-head test check passed, and the diff and synthetic merge onto fetched main are clean; the remaining normal navigation path still blocks merging. I did not run the suite locally or exercise a real Desktop session switch. No schema or migration files changed.

Automated review notice: This comment was posted by an automated review agent operated by hqhq1025. It is not an independent human review and does not replace one.

[ggbdpq]

Fixed at 65cb72209: seedActive no longer returns an empty seed for a rootless snapshot — it emits the authoritative queue once (empty or not, through the same unplacedQueue filter the update path uses), which the Desktop queue projection keys by Session and uses to retire the stale card.

Covered at both layers: a projector seed test (rootless seed → one empty queue_update) and a Desktop queue-projection regression for switch away → drain → switch back. The full navigation test through app-shell-effects would need the e2e harness; the two-layer coverage pins both halves of the contract.

[hqhq1025]

Reviewed the current head eba044200976a9e72ca15b76033ea7a369a3cc98. I found no substantiated remaining P0–P3 issue in the changed queue-projection path.

The projector now emits queue_update when the queue changes even without a live root Turn, and seedActive() also sends an authoritative queue for a rootless snapshot (packages/runtime-host/src/adapter/session-projector.ts:161-175,416-429). This addresses the previous resubscription gap: when Desktop navigates away while the queue drains and then returns, the rootless seed reaches its existing empty-queue reducer, which deletes the stale card (apps/desktop/src/renderer/app-shell-session-events.ts:303-321). Added projector and Desktop tests cover the rootless drain, rootless seed, and stale-card clearing. No schema/migration files change.

At review time the current-head test check succeeds and the PR merges cleanly with freshly fetched main at the Git merge-tree level; GitHub still reports a blocked merge state, so this is not a merge-readiness claim. Diff-check is clean. I did not run tests locally or a real Desktop navigate-away/re-subscribe scenario. Human validation and merge decision remain outstanding.

Automated review notice: This comment was posted by an automated review agent operated by hqhq1025. It is not an independent human review and does not replace one.

[Astro-Han]

Approved at @Astro-Han's explicit request: a small, focused fix with no blocking findings in the automated review of this head and green CI.