A single chunk request through the runner's budget, with the runner's retry policy.

Every attempt acquires from the Limiter first and releases after. The policy per outcome:

• 429 — release_and_pause in one cast (so waiters cannot slip in between release and pause), then retry. Rate-limit waits never consume the chunk's retry budget: the server asked us to wait, not to give up. A server-provided retry-after is honored verbatim — an hour-long quota reset is waited out, not retried against; when the header is absent the pause escalates exponentially from one backoff period, capped at 30s. A chunk retries up to :rate_limit_retries times after a 429 (default 10); the next 429 past that cap fails the chunk with the rate-limit error — bounded, unlike a budget, only by persistence of the 429s — while still pausing the runner with that final 429's deadline, so giving up never hands siblings a hot slot.
• 5xx / transport error — jittered exponential backoff, consumes one unit of chunk_retries; budget exhausted returns the last error.
• any other error (4xx, parse-level) — returned immediately; a bad request never gets better by retrying.

Emits [:lang_extract, :chunk, :retry] before each retry with the attempt number and a compact reason (:rate_limited | :server_error | :transport_error).

Internal — no stability guarantees; see the README's "Stability" section. Documented because it explains how the library works, not because it is API.