【BUG】router: cache-aware switching can block judge-requested tier upgrades

[lcp2021211]

Description

When automatic model selection and smart routing are enabled, a conversation can become stuck on the simple tier.

After an earlier request selects a simple model and gets a cache hit, later requests that require a stronger tier may still use the same simple model. The cache-aware switching logic compares costs without distinguishing between upgrades and downgrades. Since upgrading is normally more expensive, the previous cached model may keep winning the comparison.

The screenshot below shows three turns using the same simple model. The final request asks for in-depth research and a long report, but the router still keeps it on simple.

Expected behavior

Cache cost should not prevent a judge-requested tier upgrade. It should only affect downgrade decisions.

[lcp2021211]

I have submitted PR #596 to address this issue.

The PR contains a minimal change to the cache-aware switching logic:

• Read the previous tier from the sticky session state or request metadata.
• Allow judge-requested tier upgrades immediately without comparing cache costs.
• Keep the existing cache-cost comparison only for strict tier downgrades.
• Prefer the latest judge result when the transition direction cannot be determined.

I also added regression tests covering an allowed upgrade, a downgrade that retains the cached model, and a cost-effective downgrade that switches models. The TypeScript build and all router tests pass.

I would greatly appreciate the maintainers reviewing PR #596, and I hope this fix can be merged upstream. Thank you!