"opencode run": zero-chunk provider stream stall never surfaces — no timeout, no retry, no exit (1.18.30)

Symptom

Three parallel headless opencode run workers (one-shot implement agents, distinct
sessions, same provider/model opencode-go / muse-spark-1.3-contributor) all stalled
within 17 seconds of each other. Each session's LAST log line, overall, is the
stream open:

timestamp=2026-09-12T15:58:30.124Z message=stream providerID=opencode-go modelID=muse-spark-1.3-contributor session.id=ses_aaa step=6
timestamp=2026-09-12T15:58:42.856Z message=stream ... session.id=ses_bbb step=9
timestamp=2026-09-12T15:58:47.110Z message=stream ... session.id=ses_ccc step=9

Zero follow-up events on all three: no chunks, no finish, no retry, no error, no
timeout. Grep for retry|error|timeout|chunk across the three sessions returns
nothing. The processes later vanished with no DONE, no commits, no exit diagnostics
— from the driver's side this is silent work loss (one worker had green tests,
uncommitted).

Why existing fixes don't cover it (1.18.30)

• v1.18.21 "continue on unknown finish" needs a finish reason — here there is none,
  the stream never starts delivering.
• v1.18.27 5-minute header/chunk timeouts never fired (nothing logged at +5min or
  after, across all three sessions). Either the chunk watchdog arms only after the
  first chunk/headers, or the opencode-go provider path bypasses it.
• #26167 (empty-truncation retry) and #40061 (fail loud on missing finish) likewise
  key off stream events; a zero-event stream triggers none of them.

So the never-starting stream — message=stream with zero subsequent events — is
still an unbounded silent stall on latest.

Expected

A watchdog on event-less streams in run mode: after N minutes with zero stream
events, fail loudly (non-zero exit + stderr naming the session and the last event
age), so external drivers can kill + relaunch instead of discovering silent death
via file mtimes.

Repro shape (from our logs)

• 3x opencode run "<long implement prompt>" in separate dirs, same model,
  provider hiccup window; all stall at message=stream, steps 6-9, mid-task.
• Discriminator vs slow run: absolute zero log lines after the stream line
  (not slow progress — no progress).

Happy to provide full session log excerpts on request.