feat(ai): retry transient failures on queued generation reads - #51635

Why the change

A single 429, 5xx, or network blip while polling a queued generation (Veo, Runway, fal, BFL, AssemblyAI, …) failed await() even though the provider job kept running and billing, so this retries those read-only calls within the caller's poll.timeout.

Special things to note

• Only the status and result HTTP exchange retries, never the decoded document: a terminal failed generation also decodes to ProviderInternal (HTTP 200), and re-reading it would spin until the deadline. start (submit, which can bill twice) and cancel never retry. Inline and stream routes are unchanged.
• The retryability rule is not new: isRetryable moves unchanged from packages/core/src/session/runner/retry.ts to @opencode/ai (provider-error.ts), and SessionRunnerRetry re-exports it so every Core caller stays the same.
• resultEvents (behind Video.stream / queued Image.stream) used to fetch the result with no deadline. That fetch is now bounded by the same poll.timeout, so retries cannot run past it. The policy is internal: there is no new option.

Change outline

The retry policy (src/route/media.ts). A single blip recovers in about 1s, which is less than the default 5s poll interval. Each gap is capped at 30s, close to how often polls already run. The 8-retry cap (about 2 minutes) bounds a direct result() or resume, which has no deadline of its own.

READ_RETRY = max(min(exponential(1s), spaced(30s)), recurs(8))
  |> jittered
  |> delay = min(max(delay, reason.retryAfterMs ?? 0), 30s)   // honor retry-after, capped

Where it applies inside MediaRoute.queued:

 generationRoute(token, http, execute)
   poll(operation)
     transport.call("GET", path)
+      .retry({ schedule: READ_RETRY, while: isRetryable })   // RateLimit, ProviderInternal, Transport, …
     operation.decode(response)                               // not retried
   status = poll(protocol.status)
   result = poll(protocol.result)
   cancel = transport.call(cancel.method, …)                  // not retried
 start → transport.submit(…)                                  // not retried

Deadlines in src/generation.ts: await and events already interrupt retries at poll.timeout. The fix is the result fetch in the stream path:

 resultEvents(generation, expand, options)
+  start = now
   generation.events(options)            // each poll bounded by the deadline (unchanged)
-  ++ generation.result()                // no deadline
+  ++ within(generation.result(), deadline = start + poll.timeout)

Files:

 packages/ai/
 ├── src/provider-error.ts      + isRetryable (moved from core, exported from index)
 ├── src/route/media.ts         + READ_RETRY on status/result GETs
 ├── src/generation.ts          ~ shared `within` deadline helper; result fetch in resultEvents bounded
 ├── docs/media-design.md       + retry policy; interrupting a wait does not cancel the job
 └── README.md                  + same, user-facing
 packages/core/src/session/runner/retry.ts   - isRetryable body, re-exports the ai one
 packages/ai/test/video.test.ts              + 7 tests (Runway, TestClock)

Tests: 503 status and 503 result, then success · 429 retry-after: 10 (no poll at 9s, polls at 10s) · 400 not retried · persistent 503 ends with Timeout at poll.timeout · streamed result retries end with Timeout · submit 503 not retried · cancel 503 not retried. The existing Veo and Runway tests where a terminal failed generation surfaces as ProviderInternal still pass without advancing the clock, which confirms decoded failures are not retried.