feat(decopilot): background slow generate_image as a durable job

[pedrofrxncx]

Summary

Backgrounds slow generate_image calls so they no longer freeze the user's turn or the per-thread gate. On the cluster the tool enqueues a durable DBOS workflow (generate → append result → re-enter the thread gate to react) and returns a "started" handle immediately; desktop daemons POST the work back to a fence-authed cluster route.

How it works

• makeBackgroundable(name, tool, dispatcher) wraps a slow built-in. With no dispatcher (desktop without link, tests) it returns the inner tool unchanged — backgrounding is a cluster capability, not a behavior change everywhere.
• Cluster: createBackgroundToolDispatcher enqueues backgroundToolWorkflow on a per-org partitioned DBOS queue. Steps: runHeavyTool (memoized — replays don't regenerate a different image) → appendResult (assistant message rendered via the normal generate_image card) → react (re-enter the per-thread gate so the agent acknowledges/continues).
• Desktop daemon: createHttpBackgroundDispatcher POSTs to /api/:org/threads/:threadId/background-tool, authed with the run's temp bearer + run fence token, which runs the same cluster workflow.
• No bespoke job table — DBOS owns durability, idempotency, status, and cancellation; the jobId is the workflow id.

Review fixes included

• Background card no longer spins forever. The originating call returns a handle; the card is now terminal ("Image queued — generating in the background") and the finished image arrives as its own message below.
• Reaction nudge hidden from the transcript. The reaction turn's role:"user" nudge ("…do NOT call generate_image again") would otherwise render as a user bubble. Added an internal flag to message Metadata; useMessagePairs filters it, so the model still sees it but the user doesn't. Round-trips on both v1 (saveMessages) and v2 (PartEmitter finish anchor).
• Validated the route's tool input against the generate_image schema instead of forwarding an unvalidated z.unknown() cast.
• Deterministic reaction-message id so a workflow replay doesn't mint a new one.

Testing

• bun run check — clean
• oxlint on changed files — clean
• bun test on backgroundable, decopilot-events, run-registry, thread-gate-workflow — pass

Notes

• The "X agents working on N tasks" home badge (running-threads summary) was bundled onto this branch and has been removed here — it's an independent feature for its own PR.
• apps/mesh/src/tools/connection/create.ts (an unrelated ON_MCP_CONFIGURATION firstRun change) is left out of this PR.
• This branch's merge-base is behind main; the running-threads removal restores files to their pre-feature state, so a main merge/rebase is still pending separately.