fix(ai-gateway): keep long response streams alive - #214

Summary

• add bounded AI_GATEWAY_IDLE_TIMEOUT_SECONDS configuration
• default to Bun's maximum supported idle timeout of 255 seconds
• accept the complete supported 0..255 range (0 disables the timeout)
• propagate the setting to BunHttpServer.layer({ idleTimeout })

Root cause

A production-like read-only scout wave using gpt-5.6-sol exposed Bun's 10-second default request idle timeout. Long reasoning gaps caused Bun.serve() timeout warnings followed by transport ECONNRESET, even though the upstream model request was otherwise healthy.

Verification

• targeted configuration/CLI tests: 11 passed
• full AI Gateway suite: 21 files, 88 tests passed
• TypeScript check passed
• Effect migration policy: 509 paths, no violations
• @akua-dev/agentos package build passed
• git diff --check passed
• independent pre-commit review: no security concerns, logic errors, or suggestions

Operational note

This PR changes source defaults only. It does not deploy, resume the historical Fleet, alter credentials, or modify production.

Refs #85