fix(block-buffer-manager): drop redundant epoch-change pre-check

The early is_epoch_change() check in get_ordered_blocks read buffer_state
outside the mutex. If buffer_state flipped to EpochChange between the
unlocked check and the subsequent mutex acquire, the in-lock recheck (at
the start of the loop) would catch it and return the same error — so the
pre-check was redundant and only widened the TOCTOU window for analysis.

Removing it doesn't change behavior (caller still hits the in-lock check
before any state read) and reduces the surface area of audit-flagged
synchronization concerns around the dual AtomicU8 + Mutex design.

Refs: Galxe/gravity-audit#522

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: keanji-x

crates/block-buffer-manager/src/block_buffer_manager.rs

@@ -605,9 +605,11 @@ impl BlockBufferManager {
             self.ready_notifier.notified().await;
         }
 
-        if self.is_epoch_change() {
-            return Err(anyhow::anyhow!("Buffer is in epoch change"));
-        }
+        // Note: the previous early `is_epoch_change()` check here was redundant
+        // with the in-lock recheck below at the start of each loop iteration.
+        // Removing it eliminates a TOCTOU window (buffer_state could flip between
+        // the unlocked check and the mutex acquire) without changing semantics —
+        // the in-lock recheck always observes the same state we'd act on.
 
         let start = Instant::now();
         info!(