fix(client-cohttp-eio): interruptible export waits at shutdown

[dijkstracula]

Problem

When an OTLP collector cannot be reached during a traced run, exiting the process takes several seconds. Two things combine:

1. The core function Collector.remove_backend calls B.tick() before calling B.cleanup(). As a result, during shutdown an export with force:false runs while the flag stop is still false; stop is only set in cleanup.
2. send_http waits for 3 seconds (Eio_unix.sleep 3.) when an export fails.

Because this failed POST blocks for the whole backoff period before cleanup runs, the loop cannot finish. Changing the sleep to be interruptible alone doesn't solve it, because no broadcast occurs while the tick is blocked; the signal that would stop the loop lives in cleanup, which happens after the blocked tick.

Fix

Move the backoff out of the sending logic so that an export, and thus a tick, never blocks on it:

• send_http now fails fast and has a time limit (send_timeout_s). If it fails, it does not sleep. The timeout also limits a black-hole endpoint (accepts the connection but never replies), which before had no timeout.
• The ticker controls the backoff as its wait between ticks. Both the normal cadence and the post-failure backoff compete with a one-shot shutdown promise (wait_or_shutdown). This means shutdown can interrupt them immediately.
• The ticker remains a non-daemon fiber, so an in-flight final export is awaited instead of being cancelled mid-send (which would drop a batch it had already popped). The forced flush in cleanup is wrapped with Cancel.protect.

No change to core Collector.remove_backend, and no daemon fiber required.

Validation

Measured with a standalone harness and mock collector for three cases: refused, reachable, and black‑hole:

scenario	before	after
unreachable (connection-refused), shutdown mid-backoff	~3000 ms (deployed ade04b2)	0.9 ms
reachable, final batch flushed only at cleanup (force:true)	delivered	delivered (no loss)
black-hole (accepts, never replies)	unbounded	bounded at send_timeout_s

Relationship to #4

This approach is an alternative to #4. While #4 removes the backoff and daemonizes the ticker, our fix keeps the backoff but makes every wait interruptible. Therefore we don't need a daemon or changes to final batch flushing. The shutdown latency for connection‑refused remains the same, and we also bound the black‑hole case.

Note: Because remove_backend runs a tick before cleanup in core, this stall also affects main, upstream, and the ocurl/lwt backends. We plan an upstream report and PR as a follow-up.

🤖 Generated with Claude Code