oMLX OMLX_DISABLE_PRESSURE_RECLAIM and request_pressure_reclaim scheduler path

Facts

• Whether 5 grace polls cover a slow prefill chunk, since the chunk-boundary drain waits for the end of the current chunk. [source]
• `Scheduler` keeps `_pending_reclaim_request` for idle reclaim and `_pending_pressure_clear` for hard-pressure reclaim; the second "drains at the next step boundary even under load". [source]
• `request_pressure_reclaim` only sets `_pending_pressure_clear = True`, and its docstring says the actual `_sync_and_clear_cache` runs on the inference thread at the next `step()` boundary. [source]
• `_consume_pressure_clear` returns False when the flag is unset, otherwise resets it and returns True, and it is deliberately not idle-gated. [source]
• `has_requests()` includes `_pending_pressure_clear`, so an idle engine loop keeps calling `step()` until the clear fires. [source]
• The external prefill loop calls `_consume_pressure_clear()` after each chunk and then `Scheduler._clear_cache(self)`, because a single `step()` can last minutes. [source]
• A comment at that site says the abort escalation was measured firing 0.1 GB over the watermark with 3.7 GB pooled. [source]
• `Scheduler._clear_cache` calls `_sync_and_clear_cache(self._stream)` and records the pooled bytes for the prefill transient tracker when flat-overhead tracking is on. [source]
• `_process_pending_reclaim` returns without consuming the idle request while any request is running, prefilling or waiting, and cites issue 2179. [source]
• The enforcer's grace gate requires level not ok, not emergency, `OMLX_DISABLE_PRESSURE_RECLAIM != "1"` and fewer than `_PRESSURE_RECLAIM_GRACE_POLLS_MAX` (5) polls, and returns early after a successful request. [source]
• At hard level the enforcer calls `_request_scheduler_cache_reclaim(freed_hot)` after the hot-cache shrink unless the env var is "1", and then recomputes the pressure level from current usage. [source]
• `_request_scheduler_cache_reclaim` calls `request_pressure_reclaim` on each pooled engine's scheduler, tolerates exceptions per scheduler, and returns the count of accepted requests. [source]