Discount cached prefill cost in WaitAndWiden routing

Problem

WaitAndWiden estimates the full input-prefill cost for a backend selected by cache-affinity hashing. For a continuing session, most input tokens may already be in that backend cache. Treating all of them as cold can place the affinity backend in a later TTFT bucket and route the request to a cold backend too early.

Proposed behavior

Allow WaitAndWiden to scale the current request prefill estimate for cache-affinity candidates. Keep queued work unchanged, keep cold candidates at the full input cost, and retain the current behavior by default.

Acceptance criteria

• The scale is bounded to the inclusive range 0 through 1.
• Only affinity candidates receive the discounted current-request prefill cost.
• Existing configurations behave unchanged when the setting is omitted.
• A full affinity candidate can anchor the first TTFT bucket, while cold candidates become eligible through normal widening.

Start at the WaitAndWiden routing entry point and trace cache-affinity candidate selection and prefill-cost estimation. Confirm the configuration path and focused routing tests, then verify the scale bounds, affinity-only discount, unchanged defaults, and TTFT bucket behavior described in the acceptance criteria.