Route output facts should outrank uncatalogued compatibility guesses #5460

Problem

effective_max_output_tokens_for_route always intersects an uncatalogued
model's conservative compatibility guess (8192) before a concrete
RouteLimits.output_tokens fact. As a result, an unknown OpenAI-compatible
wire alias with a resolved route limit of 24576 is still sent as 8192.

The fail-closed rule is sound when the route publishes nothing: absence is not
permission. But once the resolved route carries a positive output maximum,
that exact route fact is stronger evidence than a catalogue-absence guess.
Today it can narrow the guess but can never replace it.

This affects hosts and route resolvers that already provide an explicit output
limit for a custom or gateway model that is not in Codewhale's static model
catalogue.

Proposed behavior

• Unknown model, no route output fact: keep the conservative 8192 ceiling.
• Unknown model, route output limit 24576: use min(requested_cap, 24576).
• Unknown model, route output limit 64000: use min(requested_cap, 64000).
• Known model: its documented compatibility ceiling remains authoritative;
  a route limit may only narrow it.
• A route limit must never raise the configured/requested output cap.

This keeps "absence is not permission" intact while treating a concrete route
fact as permission for that exact route.

Acceptance

• Regression tests cover unknown Openai and Custom routes with and
  without explicit output limits.
• Documented model ceilings remain authoritative.
• Formatting, focused route_budget tests, and clippy pass.