[leiysky]

Summary

Foreground staging no longer treats lock collisions as buffer pressure. Retain Region capacity requests through the final write, including when staging is full and its next flush requires alignment padding, so one capacity rejection is enough to trigger flush and rotation. A staging record-count limit only rotates when the next record cannot fit; duplicate wakes leave the new Region idle.

Reuse detached, exclusively owned L1 allocations of the exact size and skip eviction-counter updates when no eviction occurred. Caller-held values remain immutable and retain their memory charge. Count CacheSnapshot::served_bytes from the final returned value after L2 promotion, including when a concurrent publication supplies a newer L1 value of a different size.

Regression coverage includes CLOCK/S3-FIFO ownership and accounting, concurrent publication and final-owner drops, full buffers, automatic rotation, fixed routing after reclaim, and warm reopen. Permanent boundary tests cover 1,152 alignment/rotation combinations, 12 staging exact-fit/overflow cases, and 16 promotion/retained-charge cases, including empty values and statistics on/off.

Validation

• cargo x check
• cargo x test: 369 Rust tests, including extended tests and doctests
• cargo test --workspace --release --all-features: 367 tests
• cargo x lint: all checks pass, including dependency advisories with main's Tokio 1.54.1 update (chore: bump tokio to 1.54.1 for security fixes #65)
• git diff --check
• The two reported defects reproduce in the new regressions before the fixes and pass afterward
• Combined perf: reuse L1 allocations and avoid contention-driven flushes #66/test: qualify benchmark timing and paired comparisons #67 mixed-size soak checks pass in Disabled/Adaptive modes with statistics on/off; application-visible bytes match library counters when enabled
• GitHub CI: all checks pass on commit 579c1dd, including Linux/macOS on Rust 1.98.0 and stable, AddressSanitizer, ThreadSanitizer, and the Required gate

Targets current main directly. Benchmark harness and comparison tooling are in #67.