feat(llm): allow a separate attempt timeout for fallback LLMs

[roshaans]

Motivation

llm.FallbackAdapter applies one attempt_timeout to every provider, so a fallback inherits the primary's latency target. For voice agents the two needs conflict: the primary's timeout should stay tight so a slow turn fails over quickly, while a fallback typically runs cold — no warmed prompt cache, no session affinity — and needs a longer first-token window than the primary is allowed.

With a single shared value there is no good setting: raising it to accommodate the fallback slows every ordinary failover, while keeping it tight regularly cuts off a healthy fallback mid-attempt and turns a one-provider blip into a failed request. We hit exactly this in production, which is what motivated the change.

Change

New optional fallback_attempt_timeout on FallbackAdapter.__init__, applying to every LLM after the first (serving attempts and recovery checks alike):

FallbackAdapter([primary, fallback], attempt_timeout=5.0, fallback_attempt_timeout=8.0)

• Default is None, meaning attempt_timeout applies to all providers — existing behavior is unchanged.
• Implementation is a small _attempt_timeout_for(llm) lookup used by FallbackLLMStream._try_generate where it previously read the single shared value.

Tests

• test_fallback_attempt_timeout_gives_fallbacks_a_longer_window: primary misses its tight timeout, fallback's first token lands after the primary's window but within its own longer one → request served, and only the primary is marked unavailable (proving it was cut at its own timeout, not the fallback's).
• test_attempt_timeout_applies_to_all_llms_by_default: without the new parameter, the shared timeout still applies to every provider (backward-compat pin).

Both use a local fake that enforces conn_options.timeout on the first token, the way provider plugins do.

Proposal issue per CONTRIBUTING.md: #6916 — happy to adjust scope/naming based on discussion there.

🤖 Generated with Claude Code

[CLAassistant]

All committers have signed the CLA.

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.