bug(routing): knn backend drops generation settings (temperature / max_tokens / reasoning_effort) on per-model endpoints

[0xKT]

Summary

The knn routing backend routes calls to per-model endpoints through PerModelProvider, but the per-model sub-providers it builds never receive .generation. As a result, the configured agents.defaults.temperature / max_tokens / reasoning_effort are silently ignored for any call that routes to a knn endpoint; those calls fall back to the hardcoded GenerationSettings() defaults (temperature=0.7, max_tokens=4096).

Introduced by #163 (opt-in knn backend). Off by default, so impact is limited to users who set routing.enabled=true and routing.backend=knn.

Mechanism

• Generation defaults reach a call via chat_with_retry, which reads self.generation.{temperature,max_tokens,reasoning_effort} and injects them into chat() (raven/providers/base.py:516-521). The low-level chat() itself does not read self.generation.
• On the normal path, make_provider sets provider.generation from config (raven/cli/_helpers.py:134), so the values reach calls.
• On the knn path, PerModelProvider.__init__ builds one CustomProvider per model with only api_key / api_base / default_model and never assigns .generation (raven/providers/per_model_provider.py:23-30). PerModelProvider.chat_with_retry delegates to the sub-provider (per_model_provider.py:49), whose .generation is the default, so the configured values are lost.
• gateway_commands.py:50 also never sets .generation on the PerModelProvider itself.

Impact

• knn-routed calls ignore user-configured temperature / max_tokens / reasoning_effort.
• Low severity: knn is opt-in and off by default (routing.enabled=false, routing.backend=ecoclaw).

Fix

PerModelProvider should inherit the fallback provider's .generation and propagate it to the per-model sub-providers (the fallback is already configured by make_provider before knn wrapping at gateway_commands.py:196):

# in PerModelProvider.__init__, after building self._by_model
self.generation = fallback.generation
for sub in self._by_model.values():
    sub.generation = self.generation

Note

This same wiring gap also blocks the configurable llm_call_timeout proposed in #150 from reaching knn endpoints. The #150 fix will apply the 3-line propagation above as a side effect, which incidentally corrects the temperature / max_tokens loss described here. Filing this separately to record the root cause; the cross-reference on #163 keeps attribution clear.

Related: #150, #163.

[0xKT]

Resolved by #204 (merged): PerModelProvider now propagates its .generation to the per-model sub-providers, so knn-routed calls honor the configured temperature / max_tokens / reasoning_effort (and the new llm_call_timeout). Closing as fixed.