fix(stt): reconnect on send-side socket drops in remaining plugins

[flasun]

fix(stt): reconnect on send-side socket drops in remaining plugins

What

Follow-up to #6429 (deepgram) and #6467 (assemblyai/baseten/speechmatics/elevenlabs/gladia/openai/inference), which fixed a class of bug where a send-side WebSocket drop crashed the STT session instead of reconnecting. This applies the same fix to the streaming STT plugins those PRs didn't cover.

Closes #6473.

Why

SpeechStream._main_task only reconnects on APIError; any other exception kills the stream. When the peer resets the socket mid-stream, ws.send_bytes()/send_str() raises a raw aiohttp.ClientConnectionResetError (a ConnectionResetError), which isn't an APIError. In these plugins the send_task doesn't wrap the send loop, so that raw exception escapes and permanently terminates the stream — dead air, no failover. The recv path in each of these plugins already turns an unexpected close into the retryable APIStatusError (or an internal reconnect), so the fix just makes the send path symmetric.

Changes

Wrap each plugin's send_task loop in except (aiohttp.ClientError, ConnectionError), mirroring the plugin's existing recv idiom:

• Raise-style (recv raises APIStatusError → send raises the retryable APIConnectionError): xai, fireworksai, smallestai, telnyx, slng, gradium. The guard (closing_ws / self._session.closed) matches each plugin's own recv guard, so a graceful close or a session teardown returns quietly rather than raising.
• Internal-reconnect style (recv sets _reconnect_event and returns): simplismart — the send path now triggers the same internal reconnect instead of crashing.
• fireworksai and gradium additionally import APIConnectionError.

Test

tests/test_plugin_gradium_stt.py::test_send_side_socket_drop_raises_api_connection_error drives SpeechStream._run with a fake WebSocket whose first audio send_str drops the socket, and asserts _run raises APIConnectionError (retryable) rather than the raw ClientConnectionResetError. Verified it fails on the unpatched code and passes with the fix. gradium is used as the representative case since it already has a plugin test module.

Notes

• No CHANGELOG / manifest edits (per CONTRIBUTING — a maintainer handles versioning).
• Whether a drop surfaces on the send vs recv side is timing-dependent; this closes the send-side gap that (deepgram stt): fix reconnect on send-side socket drops instead of crashing the session #6429 was originally filed for.

[CLAassistant]

All committers have signed the CLA.

[flasun]

Good catch from the automated review on `slng` — thanks. Unlike the other six plugins (which propagate to the base `_main_task`, where any `APIError` is retried on the same connection), `slng` has custom per-endpoint failover that only grants same-endpoint immediate retries for `APIStatusError`; any other error permanently bumps `_active_endpoint_index`. Since `recv_task` already raises `APIStatusError` on an unexpected close, a send-side drop raising `APIConnectionError` would have permanently downgraded to a lower-priority endpoint instead of retrying the primary.

Fixed in c32c19d: the `slng` send guard now raises `APIStatusError` (matching `recv_task`), so both paths get identical same-endpoint-retry-then-failover semantics. Verified the other six do not have this type-keyed failover logic, so their `APIConnectionError` remains correct.