[arhancanli]

When storage pauses self-play, the available match cores change. The controller previously compared an existing gate with this new concurrency, failed to recognise its progress, and could start a ready training round before the gate finished. Storage recovery could similarly replace a full-budget gate or oversubscribe it with a new generator.

Keep the checkpoint's saved concurrency when every test input, setting and execution profile still matches and the allocation fits the current CPU budget. Plan generation around that allocation, and refuse to replace a matching gate when another generator leaves too few cores. New or changed tests retain the normal allocation and identity checks.

Validation: reproduced the bug read-only on the actual 5,056-game, nine-worker cont4 checkpoint: storage-ready was recognised, storage-paused was not. The fixed predicate recognises both states and preserves nine workers on the later 5,120-game checkpoint. All 46 controller tests pass, including four new regressions for storage pause, storage recovery, unchanged-input requirements and an integration test proving ready training stays deferred while the same gate advances. Two of the initial regressions failed before the fix. No playing binary, network, gate limits, completed games or search options change.