fix(admission): seed clean peers under cache-owner pressure

[DavidBellamy]

Problem

Cache-aware routing can concentrate a shared-prefix workload on one cached owner while healthy peers retain issuable capacity. Aggregate engine-feedback admission can then reject new work before worker selection, preventing recovery.

Required behavior

• Compute fail-closed per-worker issuable headroom from complete fresh rank telemetry.
• Permit a bounded, default-off distribution seed only after a redeemed fair-share credit supplies a one-shot scheduler admission proof.
• Bind the seed to one exact clean worker revision before selection and verify it again before dispatch.
• Release the distribution lease and preserve existing scheduler cancellation and refund behavior on any mismatch.
• Keep non-allowlisted partitions unchanged.
• Do not publish provisional cache ownership before backend success.

Dependencies

Stacks on #60. Related cache pressure routing context: #11.