refactor(tasks): declare-then-execute — ctx.tasks.start() replaces register()

start({ kind, label, owner, run }) preflights everything that can fail
(the attachSurface fence, validation, the owner-cleanup attach) BEFORE
invoking the producer's run() starter, then commits atomically —
'work started but never got a collectable id' is now structurally
impossible instead of a producer try/catch rollback obligation (the
P1 review fix, rebuilt on #185's declare/execute split). Producers
lose their catch-wraps; the leak tests now pin the stronger property
that a failed preflight never spawns anything. TaskRegistration splits
into TaskStart (identity + run) and TaskHooks (cancel/done/readOutput);
docs, type-equiv manifest, catalogs, and both RFCs move with it.

docs/cookbook/adding-a-tool.md

@@ -41,7 +41,7 @@ Registration is effect-based: disposing the plugin fiber unregisters the tool (w
 
 ## Long-running work
 
-Register the running work with the shared task runtime instead of inventing a task protocol: gate a `run_in_background` parameter behind your plugin's own defaulted `enableRunInBackground`-style config, start the work, and hand it to `ctx.tasks.register({ kind, label, owner: exec.agent, cancel, done, readOutput? })` (`@deepseek-ai/dsh-tasks`). The runtime issues the `<kind>-N` id, fences access to the owning session, cancels-and-awaits your task when the owner disposes, and the generic `task_output`/`task_list`/`task_kill` tools plus the completion notice come from `@deepseek-ai/dsh-tool-tasks` — your tool returns `started background task <id>` and is done. Your producer keeps its execution concerns: `done` must settle at quiescence (resources released), and a stream-kind `readOutput` owns its own truncation/spill formatting (bound buffers, spill full output to disk so nothing is silently lost — see tool-bash's `renderProcessRead`). Do NOT wire `exec.signal` to the background work after the id is returned; check `exec.signal?.aborted` once before starting, then leave cancellation to `task_kill` and owner cleanup. **A failed `register()` must not orphan the work**: `register()` is atomic (a throw — the no-control-surface fence, a bad owner — mutates no registry state), so wrap it in try/catch, cancel the just-started work, AWAIT its quiescence, and rethrow — the model never learns an id, so nothing else could ever collect or kill what you started (tool-bash's `proc.kill(); await proc.done` and tool-subagent's `run.cancel(); await done` are the templates).
+Hand long-running work to the shared task runtime instead of inventing a task protocol: gate a `run_in_background` parameter behind your plugin's own defaulted `enableRunInBackground`-style config, then call `ctx.tasks.start({ kind, label, owner: exec.agent, run: () => ({ cancel, done, readOutput? }) })` (`@deepseek-ai/dsh-tasks`) — the runtime preflights everything that can fail (the control-surface fence, validation, owner-cleanup attach) BEFORE invoking your `run()` starter, so work that started without a collectable id is structurally impossible (no try/catch rollback in your tool). The runtime issues the `<kind>-N` id, fences access to the owning session, cancels-and-awaits your task when the owner disposes, and the generic `task_output`/`task_list`/`task_kill` tools plus the completion notice come from `@deepseek-ai/dsh-tool-tasks` — your tool returns `started background task <id>` and is done. Your producer keeps its execution concerns: `done` must settle at quiescence (resources released), and a stream-kind `readOutput` owns its own truncation/spill formatting (bound buffers, spill full output to disk so nothing is silently lost — see tool-bash's `renderProcessRead`). Do NOT wire `exec.signal` to the background work after the id is returned; check `exec.signal?.aborted` once before calling `start`, then leave cancellation to `task_kill` and owner cleanup.
 
 ## Permissions / sandboxing