Refactor Makefile and enhance disk/memory eviction tests

- Updated the 'bench' target in the Makefile to run all benchmarks for MemoryFS and DiskFS, improving clarity and coverage.
- Added explicit post-eviction consistency checks in DiskFS tests to ensure on-disk files are removed after eviction.
- Introduced new benchmarks for memory eviction strategies under pressure, enhancing test coverage for memory management.
- Improved error handling in benchmark tests for both disk and memory file systems, ensuring robustness during performance evaluations.
- Refactored key generation in tests for consistency and clarity.

vfs/disk/disk.go

@@ -66,7 +66,7 @@ func (d *DiskFS) shardPath(key string) string {
 }
 
 // pathForKey returns the full on-disk path for a key (sharded + normalized).
-// Extracted to reduce duplication in Evict*/Delete/Open paths (addresses review nit19; still safe to call under lock for evict).
+// Extracted to reduce duplication in Evict*/Delete/Open paths (still safe to call under lock for evict).
 func (d *DiskFS) pathForKey(key string) string {
 	shardedPath := d.shardPath(key)
 	path := filepath.Join(d.root, shardedPath)
@@ -584,7 +584,7 @@ func (d *DiskFS) EvictLRU(bytesNeeded uint) uint {
 			d.LRU.Remove(key)
 			delete(d.info, key)
 			path := d.pathForKey(key)
-			_ = os.Remove(path) // best effort
+			_ = os.Remove(path) // best effort; performed under WLock (reverted from post-unlock) to guarantee on-disk deletion is coordinated with metadata removal. This eliminates resurrection via lazy Stat/Open discovery and prevents late unlinks from deleting content of same-key recreates (critical for in-memory metadata safety model + user's explicit non-race requirement on hot eviction path).
 			d.size -= fi.Size
 			evicted += uint(fi.Size)
 			shardIndex := locks.GetShardIndex(key)
@@ -607,6 +607,9 @@ func (d *DiskFS) EvictBySize(bytesNeeded uint, ascending bool) uint {
 	var candidates []evictCandidate
 	for key, fi := range d.info {
 		candidates = append(candidates, evictCandidate{key: key, size: fi.Size})
+		if len(candidates) >= maxEvictBatch {
+			break
+		}
 	}
 	d.mu.RUnlock()
 
@@ -631,7 +634,7 @@ func (d *DiskFS) EvictBySize(bytesNeeded uint, ascending bool) uint {
 			d.LRU.Remove(key)
 			delete(d.info, key)
 			path := d.pathForKey(key)
-			_ = os.Remove(path)
+			_ = os.Remove(path) // best effort; performed under WLock (reverted from post-unlock) to guarantee on-disk deletion is coordinated with metadata removal. This eliminates resurrection via lazy Stat/Open discovery and prevents late unlinks from deleting content of same-key recreates (critical for in-memory metadata safety model + user's explicit non-race requirement on hot eviction path).
 			d.size -= liveFi.Size
 			evicted += uint(liveFi.Size)
 			shardIndex := locks.GetShardIndex(key)
@@ -655,6 +658,9 @@ func (d *DiskFS) EvictFIFO(bytesNeeded uint) uint {
 			key   string
 			cTime time.Time
 		}{key: key, cTime: fi.CTime})
+		if len(candidates) >= maxEvictBatch {
+			break
+		}
 	}
 	d.mu.RUnlock()
 
@@ -676,7 +682,7 @@ func (d *DiskFS) EvictFIFO(bytesNeeded uint) uint {
 			d.LRU.Remove(key)
 			delete(d.info, key)
 			path := d.pathForKey(key)
-			_ = os.Remove(path)
+			_ = os.Remove(path) // best effort; performed under WLock (reverted from post-unlock) to guarantee on-disk deletion is coordinated with metadata removal. This eliminates resurrection via lazy Stat/Open discovery and prevents late unlinks from deleting content of same-key recreates (critical for in-memory metadata safety model + user's explicit non-race requirement on hot eviction path).
 			d.size -= liveFi.Size
 			evicted += uint(liveFi.Size)
 			shardIndex := locks.GetShardIndex(key)
@@ -702,6 +708,9 @@ func (d *DiskFS) EvictLFU(bytesNeeded uint) uint {
 			accessCount int
 			aTime       time.Time
 		}{key: key, accessCount: fi.AccessCount, aTime: fi.ATime})
+		if len(candidates) >= maxEvictBatch {
+			break
+		}
 	}
 	d.mu.RUnlock()
 
@@ -726,7 +735,7 @@ func (d *DiskFS) EvictLFU(bytesNeeded uint) uint {
 			d.LRU.Remove(key)
 			delete(d.info, key)
 			path := d.pathForKey(key)
-			_ = os.Remove(path)
+			_ = os.Remove(path) // best effort; performed under WLock (reverted from post-unlock) to guarantee on-disk deletion is coordinated with metadata removal. This eliminates resurrection via lazy Stat/Open discovery and prevents late unlinks from deleting content of same-key recreates (critical for in-memory metadata safety model + user's explicit non-race requirement on hot eviction path).
 			d.size -= liveFi.Size
 			evicted += uint(liveFi.Size)
 			shardIndex := locks.GetShardIndex(key)
@@ -753,6 +762,9 @@ func (d *DiskFS) EvictHybrid(bytesNeeded uint) uint {
 			accessCount int
 			aTime       time.Time
 		}{key: key, accessCount: fi.AccessCount, aTime: fi.ATime})
+		if len(candidates) >= maxEvictBatch {
+			break
+		}
 	}
 	d.mu.RUnlock()
 
@@ -777,7 +789,7 @@ func (d *DiskFS) EvictHybrid(bytesNeeded uint) uint {
 			d.LRU.Remove(key)
 			delete(d.info, key)
 			path := d.pathForKey(key)
-			_ = os.Remove(path)
+			_ = os.Remove(path) // best effort; performed under WLock (reverted from post-unlock) to guarantee on-disk deletion is coordinated with metadata removal. This eliminates resurrection via lazy Stat/Open discovery and prevents late unlinks from deleting content of same-key recreates (critical for in-memory metadata safety model + user's explicit non-race requirement on hot eviction path).
 			d.size -= liveFi.Size
 			evicted += uint(liveFi.Size)
 			shardIndex := locks.GetShardIndex(key)