feat(desktop): use the shared composer in WorkHub progress cards

[M4n5ter]

Summary

WorkHub progress cards keep the shared composer available in compact, expanded, docked and floating layouts. Focusing opens the input area, and accepted sends expand the conversation. Enter and the send button queue a follow-up while a Turn is running; Shift+Enter steers the current Turn, matching ordinary chat. Both paths preserve uploaded attachments and submission identity. An uncertain submission can only be retried with its original text, attachments and sending mode until Host evidence resolves it.

The pending panel uses the Host queue as its single authority, with editing, removal and ordering for steering and follow-ups. Local admission placeholders stay visible after an accepted response until matching queue, steering or transcript evidence arrives, including across a disconnected observer. Steering enters the transcript when Runtime consumes it; the provider boundary waits for pending steering admissions and durable queue edits and includes the entire batch in order. Follow-ups advance one per Turn and retain WorkHub coordination authority across recovery, including crashes before successor admission or before its Run starts. Cold recovery retains pending work until the original Desktop provider returns with both required WorkHub tools. Live binding verifies both tools and captures the authenticated principal, Client and capability owner fingerprint in one registry mutation; every new v2 root requires that evidence before durable admission, and successors inherit it. Unrelated providers cannot activate recovery, even with identical tool names. Older admissions without binding evidence remain pending. Stop receipts clear withdrawn rows even without retraction events. A shared question-mark tooltip explains the shortcuts. The unused display-anchor contract and two single-caller public facades are removed; compatibility epoch advances to 139.

Card and conversation transitions share a damped resize animation, preserve focus and IME input, and respect reduced motion. Compact composers resize with their content; expanded conversations allow native manual resizing. A fixed-height renderer canvas stays anchored to the native bottom during transitions, avoiding repeated conversation layout. Cancelling a transition restores normal input coordinates before docking. Warm summons retain painted content.

Verification

• Full build, repository-wide typecheck, lint, format, Knip (Desktop/UI), renderer architecture, AppShell hook scope, locale hygiene, E2E budget, ASF headers and Astryx inventory passed.
• Runtime Host: 1,833 passed (12 skipped); Storage: 1,232 passed (8 skipped); Core: 824 passed. The preceding Desktop and AI SDK runs passed 2,472 and 234 tests respectively. The earlier full-workspace run also passed Core, Storage, Runtime, UI and the other workspace suites.
• The production composer and AI SDK run against a local provider fixture: Desktop disconnects between live binding and root admission, reconnects before backend composition, then recovers across both cold-recovery crash cuts with the original authority evidence. An unrelated capability provider registers the same tool names first and cannot activate recovery or receive tool calls. The original Desktop later registers both tools and executes the successor exactly once. Separate coverage rejects changes to the provider principal, Client and capability owner while accepting legitimate connection and registration renewal.
• Queue regressions cover a delayed steering admission commit at the final boundary, edits and ordering, in-flight snapshots, withdrawal, response-before-observation visibility, and exact retries after unknown outcomes. The real preload and Desktop adapter also cover uploaded attachments in idle answers, running-turn queue modes, reads and foreign-Host rejection.
• Both native WorkHub E2E cases passed, covering repeated folds, renderer recovery, Enter/send-button follow-ups, Shift+Enter steering, Stop queue cleanup and tooltip bounds.
• Earlier GPU/compositor traces on a 60Hz display reduced Layout events from 79 to 30–31 per fold. Occasional 33ms presentation gaps remain; elimination of cold shader compilation is not claimed.

Before:

After:

AI use

• No generative tool made a substantive contribution
• Generative tooling made a substantive contribution

Tool(s) and scope: OpenAI Codex — implementation, regression tests, and validation.

Checklist

• Tests cover the change and fail without it
• Lint, format, typecheck and the affected suites pass locally

Does this PR entail a change in behavior?

• Yes — described under Summary above
• No

[jackwener]

Approved at exact head a88bf417f792b4bf4f34e8fa23f68a3e3d569c4b at the explicit direction of M4n5ter. No technical review was performed as part of this action; M4n5ter requested the approval and accepts full responsibility for subsequent handling.

At submission time, the hosted test check was terminal and successful. This approval is not a merge decision, and no merge was performed.

---

Review notice: This approval was submitted by an automated review agent operated by jackwener and is published at the direction of M4n5ter, who requested this action and is the human accountable for it.