fix(provider): recycle sidecar after idle/terminal errors

[justin-carper]

Summary

• After opencode sits idle (15–60min+), every subsequent Cursor turn fails with Cursor run ended with status "error". /new does not help; only restarting opencode does.
• Root cause: the Node sidecar hosts @cursor/sdk, which memoizes its streaming transport with no reconnect. Cursor's backend drops idle sessions; unary RPCs still succeed (so runs get created), but the dead stream ends every run status:"error". The dead state is process-wide → fails fresh sessions too.
• Fix: recycle the sidecar child after 10 minutes idle and after any terminal run error. Next request spawns a fresh child (fresh transport). Pooled agents resume from Cursor's checkpoint store — same path as an opencode restart. Complements fix(session): retry resumed turns that fail against an expired Cursor agent #52 (per-agent resume retry), which cannot heal a shared dead transport.

Test plan

• npm run typecheck
• npm test (310 tests)
• npm run build
• Sidecar unit tests: recycle on status error, recycle on idle TTL, keep child across healthy turns, never recycle while a sibling request is in flight
• Manual: leave opencode idle 20+ minutes under Bun + Node sidecar; send a message — should succeed (or fail once then succeed on next message before this fix would require a restart)

[justin-carper]

Closing as superseded by #76 (v0.5.0 native experience overhaul).

Rationale:

• Native experience overhaul: reliability, streaming fidelity, transport (v0.5.0-next.0) #76 flipped the default Bun transport to http1 in-process and demoted the Node sidecar to opt-in rollback (OPENCODE_CURSOR_TRANSPORT=sidecar). This PR recycles the sidecar child, so it does nothing for the new default path that most users are now on.
• The Cursor run ended with status "error" after-idle symptom on resumed turns is now self-healed transport-agnostically by Native experience overhaul: reliability, streaming fidelity, transport (v0.5.0-next.0) #76's classified replayable retry (re-create fresh agent + full replay on the unknown/network/etc. error set).
• The likely reason Native experience overhaul: reliability, streaming fidelity, transport (v0.5.0-next.0) #76 chose http1 is that HTTP/1.1 opens fresh connections per request instead of holding one long-lived HTTP/2 session, which plausibly moots the idle transport-death on the default path.

Remaining gaps worth tracking separately (not addressed here):

1. Fresh-turn (/new) inactivity error is not retried on any transport (Native experience overhaul: reliability, streaming fidelity, transport (v0.5.0-next.0) #76's retry is gated on acquired.resumed).
2. No confirmation of whether the idle failure reproduces on the http1 in-process default.

If either proves to be a real problem, the right fix is a transport-agnostic reset/retry for fresh turns, not sidecar-child recycling.