Bounded delivery of supervised chunk results to a lazy caller-side stream.

The consumer's demand drives admission: at most :buffer tasks are outstanding, so the caller's mailbox never holds more than :buffer undelivered results (plus their :DOWN notices). A slow consumer throttles admission instead of growing a mailbox — the normative bounded-delivery constraint from the production-pipeline design.

Task crashes become per-chunk events via the monitor :DOWN reason; halting the stream early kills all outstanding tasks.

Why hand-rolled

Without drain, this module should collapse into Task.Supervisor.async_stream_nolink(..., max_concurrency: buffer, ordered: false, zip_input_on_exit: true) — the bounded/unordered/ crash-to-event core duplicates it. Drain is what it can't express: reporting never-started chunks requires detecting the dead supervisor at admission time. Layering that detection on top was also considered and rejected: a Stream.transform can't catch the supervisor-down exit (it erupts between elements, inside the inner enumerable's reduction), so the wrapper needs continuation-driving via Enumerable.reduce plus OTP exit-shape matching plus reconstructing the never-started set by diffing observed byte ranges — more fragile than this explicit loop, and dependent on uncontracted stdlib event ordering. If drain semantics are ever dropped, delete this module in favor of async_stream_nolink.

Drain

When the runner shuts down mid-stream (deploy, supervisor restart), chunk tasks trap exits, so the supervisor's shutdown signal gives each in-flight request up to :shutdown milliseconds to finish naturally — completed results still reach the consumer. Chunks never started are emitted as {:error, %ChunkError{reason: :drained}}, detected when admission hits the dead task supervisor. Consumers already handle per-chunk errors, so drain adds no new consumer code paths.