feat: fall back for streamed requests that fail before their first token

[Yash-Chindam]

What

Streamed requests now get the same declared fallback as buffered ones (§7.1, §16 "no silent routing or fallback changes"), and an engine that fails before producing anything returns a real error status instead of a 200 with a broken stream.

How

The gateway draws the first token before it sends any byte. Until then the status code and route headers are still open, so:

• if the local engine fails at that point and the request is entitled to the external provider, it falls back; the headers, every chunk, the metrics and the trace name the model that actually answered;
• if there is no fallback, the caller gets the ordinary 502/503 JSON error, and the admission slot is released;
• once a token has been sent there is no fallback: one model's output is never continued by another.

Tests

Four new integration tests: fallback before the first token; no fallback gives a real status and frees the only slot; a failure after the first token does not switch models; an empty stream still ends cleanly.

Local: ruff and mypy clean, 483 tests pass, Playwright 9/9.

Limits

Time to first byte for a stream now includes the engine's time to first token, since nothing is sent before it. The role chunk used to go out immediately.

🤖 Generated with Claude Code