fix(core): generate titles at the lowest model variant - #51939

Issue for this PR

No linked issue.

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

When a provider has no small model, title generation ran on the session model at the session's selected variant, so reasoning models spent their full reasoning budget on a title (Kimi K2.6 used ~1,300 reasoning tokens, Meta Muse ~1,450).

Title model selection is now:

1. The provider's small model at its lowest variant (unchanged).
2. Otherwise, the session model at a lower variant:
   • none or minimal when the model has one, since those always reduce reasoning.
   • low only when it ranks below the variant the session selected. On some models (for example Claude) the default variant does not reason and low turns reasoning on, so low is never chosen over the default.
3. If that attempt fails, the session model at the selected variant (the existing fallback in title.ts).

If no lower variant applies, the title uses the session model at its selected variant in a single attempt.

How did you verify your code works?

• Added tests in packages/core/test/session-title.test.ts for the lower-variant path, the fallback to the selected variant on failure, no lower variant, a default selection that skips low, and a default selection that uses none. All 20 title tests pass; bun typecheck passes in packages/core.
• Live run on moonshotai/kimi-k2.6: the title request used about 524 input, 6 output, and 1 reasoning token (previously ~1,336 reasoning tokens).

Screenshots / recordings

Not applicable.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR