[wingleeio]

Fixes the four v0.2.12 field reports on healthy networks:

1. Flashing warnings / phantom "Queued" / "Reconnecting" before sends

compute_connectivity reported raw socket state, so a chat room mid-join (every navigation), an idle link waking for a send, or any sub-second dial read as degraded — flashing the amber notice for ~300ms and stamping sends "Queued" on enterprise-grade connections.

Degradation is now hysteretic (DegradeGrace, engine-side): a source (OS path / registry room / per-chat room) must be raw-degraded for 4s continuously before it's reported; recovery reports instantly. All consumers (pill, composer notice, row Queued/Failed, send_queued) inherit the calm from the one stream. Unit-tested over injected instants.

2. "Couldn't stage the attachment locally"

A real race shipped in the parallel-chunk upload path: Uploads::append runs sweep() before writing, and sweep reclaimed empty staging dirs unconditionally — deleting a sibling chunk's just-created dir between its create_dir_all and first write. 3-wide chunk concurrency made the window live, and the no-delay lockstep retries re-collided. Empty dirs are now judged by their own mtime against the 10-min TTL (regression test included), and chunk retries stagger by seq.

3. Reconnect UI restyle

The amber-bordered strip with the raw transport-error line is gone. Degraded states render as one quiet frost pill — neutral dot + muted 11px label, fade-in entrance — with amber reserved for hard OS-offline. The composer notice drops its warning box for a single caption line ("Messages will send once the connection recovers."). Retry-countdown and last-failure text removed from the sidebar (they read as debug output; the detail still lands in logs).

4. Atomic new-chat sends (the "empty transcript" report)

The staging failure had a worse half on new-chat sends: the pipeline created state before it was entitled to (CreateChat mutate ran before staging; the composer had already selected the minted chat and cleared the input), so a staging failure stranded the user inside a just-created empty chat with their message gone.

• Attachments now stage first — before the chat row or worktree spec exist anywhere. Staging is chat-independent (uploadId-keyed), so a failure aborts with nothing created. Both the queued local-stage and legacy host-stage flows move together.
• A post-mutate failure (QueueCommand) best-effort deletes the minted row — no orphan either way.
• Failure restore targets the canvas for new chats: back to the new-session canvas with prompt text, attachment stash, and the error banner — handling both re-key orderings (the deleted row's watch event can re-key the composer before the failure handler runs; caught live in the rig).

Verified E2E on the headless rig — read-only uploads dir + new-chat image send: canvas retained, no stranded sidebar row, banner + prompt text + attachment chip all restored:

Tests

• New: degrade_grace_tests (blip-never-reports, persist-then-report, hide-fast, per-source independence), sweep_spares_a_fresh_empty_staging_dir.
• Full sweep: 709 engine/ui/rpc + 38 sync, 0 failures (re-run after every commit on the branch).

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

Screenshots

Degraded posture (dead edge, >4s — so the grace has expired and the state is real). Reconnecting is a transparent grayscale spinner + faint caption — no surface, no border:

Detail — sidebar spinner line and composer caption:

On healthy networks neither element ever renders — sub-4s blips (room joins, idle-link wakes, navigation dials) report Connected.