[P0][Plan 2] SF-006 — Replace destination burst with bounded generation concurrency

[Ankit6149]

Confirmed problem

After strategy generation, all selected destinations launch together with Promise.all. With twelve destinations and one revision attempt per channel, a campaign can create up to 25 provider requests in a concentrated burst, risking rate limits, high cost, inconsistent latency, and serverless timeout failure.

Scope

• Split destination generation into independent requests.
• Add a client/server generation queue with a default concurrency of two.
• Support provider-specific and local-model concurrency limits.
• Give long-form destinations a separate lane or scheduling policy.
• Add cancellation through campaign-level and request-level abort controllers.
• Record request count, duration, retries, provider/model, and destination status.

Acceptance criteria

• No campaign launches all destination requests simultaneously.
• Queue progress is visible and updates as each destination starts/completes.
• One destination failure does not cancel unrelated destinations.
• Cancel stops queued work and aborts active work safely.
• Rate-limit/payment errors are not automatically retried.
• A full twelve-destination run respects the configured concurrency ceiling.

Required verification

• Concurrency tests measuring maximum simultaneous calls.
• Partial failure, cancellation, timeout, and long-form tests.
• Provider capability tests for hosted and local routes.

[Ankit6149]

Closing as completed against the issue's stated acceptance criteria.

Merged evidence:

• Finish generation quality contracts and recovery states #322: bounded destination generation foundation and non-retrying safe provider error classes for rate-limit/payment/quota failures
• Stop active and queued generation on cancellation #326: campaign/request abort propagation; queued work stops starting; active provider requests abort safely; partial drafts remain intact
• Bound and measure provider requests during generation #328: hard provider-request budget + request/duration/provider/model/destination telemetry
• Stream truthful per-destination generation progress #329: truthful queued/generating/revising/complete/needs_review/failed/cancelled progress
• Schedule generation by provider capacity and content length #330: hosted ceiling 2, local ceiling 1, configured values can only lower the ceiling, and long-form destinations run in a separate sequential phase
• Verify full generation concurrency ceiling #332: explicit full twelve-destination hosted acceptance test proving the concurrency ceiling and long-form phase at the exact scale named in this issue

The broader durable cross-restart job infrastructure remains separate work and is not part of #27's original acceptance contract.