Provider-neutral, bounded, resumable execution for durable LM batches.

Requests have stable caller-supplied IDs and JSON-safe payloads. A dispatcher receives each normalized request plus attempt metadata and must return one of:

• {:ok, output}
• {:transient, reason} - the request did not run and may be sent again
• {:terminal, reason}
• {:malformed, reason}
• {:ambiguous, reason} - the request may have run; it is not sent again

A request that may already have run is never re-dispatched: sending it again could run and bill it twice. Only a :transient outcome is retried, up to :max_attempts. A dispatcher that raises, throws or exits, or that is still running at :timeout, may have sent its request first, so its outcome is :ambiguous.

The batch owns the wait before a retry as well as the retry. When the failure carries the provider's retry-after header (seconds or an HTTP date), the request is not sent again before then; otherwise the wait grows exponentially with the attempt, with jitter. A request that is waiting holds no dispatch slot: other requests are sent meanwhile, and the batch sleeps only when every remaining request is waiting. A process that traps exits and is stopped by its parent during that sleep exits at once with the parent's reason; a dispatch wave that is running is still waited for, for up to :timeout. The parent is the first of the process's $ancestors, which a process started by Task, GenServer or another OTP behaviour has; a process started with bare spawn/1 or spawn_link/1 has none, and its sleep cannot be interrupted. Backoff is capped at :max_retry_wait. When the provider asks for longer than :max_retry_wait, or the wait would end after the Imp.Deadline bound to the calling process, the request is not retried in this run: it stays :transient_failure with the attempts it has used, the summary is not complete?, and resume/3 retries it. The time a request may be sent again is kept in the checkpoint as a UTC time, and resume/3 waits for it under the same rules.

The checkpoint is an auditable JSON document. Every dispatch intent and outcome is appended to its event history before execution continues. Writes use a synced temporary file followed by an atomic rename. On resume, a request that was dispatched but has no committed outcome becomes :ambiguous for the same reason.

A checkpoint written by Imp 0.5.0 (schema version 1) retried timeouts and dispatcher crashes as transient failures, and its stored reason does not always say which a failure was. On resume, its :transient_failure requests become :ambiguous and are not sent again; the checkpoint is rewritten at schema version 2.