fix(ssd-cache): inline LRU unlinks so eviction frees queue capacity

[cfbraun]

Summary

_enforce_size_limit_for_new_block enqueues evicted file unlinks as ("unlink", path) items onto _write_queue — the same bounded queue that carries pending writes. Combined with the pre-eviction _write_queue.full() short-circuit at the top of save_block, this creates a deadlock under sustained save pressure:

1. Writer is saturated → _write_queue is full.
2. save_block's pre-eviction check sees full() → returns False immediately, before calling _enforce_size_limit_for_new_block.
3. Eviction never runs → cache stays at the size cap.
4. Every subsequent save drops; ssd_write_drops climbs forever while _total_size sits pinned at the cap.

Fix: inline the unlinks on the eviction-calling thread instead. Eviction typically removes a single block per save (evict_until_size stops as soon as total_size <= target), so this is one syscall per save in steady state. The deferred-unlink justification ("avoid blocking the inference thread with N file delete syscalls") doesn't materialise under normal load, and inlining removes the bounded-queue contention entirely.

Bounded inline burst. The ENOSPC-recovery path invalidates the 30 s disk-usage cache, which can shrink _get_effective_max_size sharply — evict_until_size would then return hundreds of entries at once and the inline-unlink loop would stall the inference thread on a syscall storm. Cap the burst at _MAX_INLINE_UNLINKS_PER_SAVE = 32 and reinsert the deferred metadata into the index so subsequent saves drain the remainder.

Also adds an evict_unlink_failures stats counter (eviction now decrements the index before the on-disk unlink; if Path.unlink raises OSError, surfacing the counter lets operators see when on-disk size has drifted above what the index reports).

The dead writer-thread ("unlink", file_path) dispatch branch is removed since no path enqueues such tuples anymore.

Test plan

• pytest tests/test_paged_ssd_cache.py::TestInlineLRUUnlinks — 4 passed
• pytest tests/test_paged_ssd_cache.py tests/test_hot_cache.py — 125 passed (4 new + 121 existing)

[jundot]

Thanks for tracking this down. The root cause makes sense: eviction was sharing the bounded write queue with pending writes, so once the queue saturated, cache pressure could not drain cleanly.

This looks good to me, and I am going to merge it. I verified tests/test_paged_ssd_cache.py and tests/test_hot_cache.py locally, and I will fold two small follow-ups into main after merge: preserving LRU order by sorting _lru by last_access after capped inline eviction reinserts deferred entries, and threading evict_unlink_failures through PagedSSDCacheStats so it shows up in the runtime/admin stats path.