fix(evals): Stabilize model-backed scenarios

[dcramer]

The flaky model-backed evals now fail fast and retry at the correct boundaries instead of drifting into live-network, unrelated retrieval, or inconsistent routing paths.

• Tool model settings are parsed once into typed botConfig values and passed through tool composition instead of read from ambient environment variables.
• Web fetch replay stores one canonical max-size source per URL, then applies any model-selected max_chars value through shared production truncation logic. Strict harness coverage verifies two limits replay from the same fixture.
• Eval AI Gateway requests use an origin-scoped 25-second body-idle timeout. A real Pi stream integration verifies the resulting assistant error enters Junior's existing provider retry path; non-AI eval HTTP is unchanged.
• Memory provenance scenarios contain the facts required for their drafting/recap requests, and the resource-event fixture skill only handles synthetic provider discovery.
• High-complexity or uncertain routing conservatively selects the handoff profile, while Slack delivery policy explicitly prevents presenting unsupported top-level channel posts as delivered.

The 60-second eval reply budget and normal shard parallelism remain unchanged.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
junior-docs	Ready	Preview	Jul 24, 2026 12:25am