Try budgeted cache recovery before splitting a minimum wave

[bradhilton]

When a minimum microbatch wave fits only after splitting, give its rejected flat demand one cache-recovery opportunity under the existing budget before accepting the split. The existing WORLD fallback vote carries this opportunity uniformly across split, flat and empty local shares. Recovery runs ordinary fresh search once; denied or ineffective recovery keeps the split only if a fresh admission check still fits.

This intentionally changes the prior “a split fits, return immediately” behavior. It preserves the shared lifetime first attempt and subsequent 5% ledger, physical-free admission and error handling. It adds no ordinary-path collective, new budget, public API or art.megatron change.

Validation: 62 focused CPU tests and 9 subtests passed, with CUDA hidden and actual cleanup. The changed real-estimator denied-upgrade test also passed in the earlier run; its runtime/test bytes are unchanged. Ruff and diff checks passed. Rebased onto current main c1b99e36 with identical candidate/runtime/dependency files; the intervening change concerns trajectory-capture cleanup. Actual two-rank collective and candidate-native qualification remain pending.

Motivation comes from a separate baseline GPU diagnostic, not qualification of this candidate: two approximately 135 ms releases restored the exact natural 64,902-row flat plan. Its same-plan allocated peak stayed 93.849 GiB. Quiet flat forward/backward took 8.332 s versus 10.506 s for a historical split, with higher peak memory; this separate-run comparison is not a throughput guarantee.

Related to #848 and #870. This does not close their broader qualification scope.