[hudsonaikins]

Released as v0.4.6 at 100% production traffic through the Release workflow. Release source 3c257da34faf637a4c8c70cb2e3799cb54e11347; Worker 2226a90f-baba-4123-9022-d1ac8e137890. The exact uploaded image passed a fresh two-leg trip and all four BUILD-04 outcomes before promotion. Release acceptance used two sequential Browserbase sessions (151.080 seconds), two FlightAware requests, and left zero active browsers. Temporary resources were removed. Release evidence records usage, rollback and limitations. Authenticated production health awaits renewal of the Cloudflare Access Google sign-in.

Numbered revisions of the same leg could be treated as alternatives, or an approved revision could stop after preparation. The dispatcher now completes each requested revision in order, following both dispatch and revision approval with generation and finalization while preserving earlier results.

The original WF-025 prompt, fixtures, clock, harness and verifier remain unchanged. Exact per-revision approval and existing failure/uncertain-Generate controls are preserved.

Validation:

• Standard local product validation passed on head ddd03f6cff915ad2f099441ed90287195e7b606d: 458 tests, 3 Postgres-only skips, Ruff and dead-code checks.
• Matched configured-model MCP replay with controlled flight providers: original WF-025 failed on the released baseline and passed with the fix, retaining both revisions through checkpoint recovery.
• Live acceptance on the exact candidate image passed both the four-to-five-passenger revision sequence and KIAD–KEWR–KIAD two-leg itinerary. Each dispatch was separately approved and generated once; prior snapshots and briefings remained retained. All four BUILD-04 outcomes passed using the existing database-derived checks.
• Across the two live batches: four sequential Browserbase sessions totaling 202.743 seconds and four FlightAware requests ($0.06 list-price estimate). Sessions completed, no browsers remained active, and temporary resources were removed.

The acceptance report retains before/after results, LangSmith links, usage, diagnostic client stops and limitations. The canonical live record is bound to source commit c5deeee3aa8ba2582cf67926f91568cc4ca93510; the evidence-only head has identical integration bytes. Historical release records are pinned to their original evidence.

This change adds no aircraft support, thrust formula, provider/model selection or operating-limit change. The targeted observations do not establish statistical quality or latency improvements.