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.