[jepsontaylor]

Summary

Ideas taken from the kernel benchmark battletest that were still true in the tree, as tests and fixes. No speed ratio is claimed. Durability stays Strict. wal_pipeline stays off.

• Clock eviction walks enough times to decay a frame whose usage has saturated at 5. A clean unpinned page is no longer reported as unevictable after two visits.
• A row lock that is free but already has a waiter is not granted to a newcomer. A spurious wake grants the lock only to the waiter at the front.
• A morsel arena offset that does not fit in u32 aborts instead of truncating the stored offset in release builds.
• An exact 128-byte WAL segment fill stays at 128 bytes. The rotation path already refused to write past the limit; the test keeps that.

Not in this pull request: one-pass scan loads, positional I/O, physical WAL batching, route-versus-index admission, and any suite median. Those are later measurements on the official Normal lane.

Claim: AGENT_CHAT.md 2026-09-25T18:40Z fix/kernel-p0-gates.

Acceptance: cargo test -p redlinedb-kernel --offline --lib newcomer_does_not_cut fifo_handoff; cargo test -p redlinedb-kernel --offline --test storage_tests buffer_pool_evicts_a_clean_frame; cargo test -p redlinedb-kernel --offline --test wal_tests exact_segment_fill; cargo test -p redlinedb-sql --offline --lib an_offset_past_u32. Local just pr-ci exited 0 (/tmp/redline-pr-ci-p0.log).

Parent: 76ba2a52c469c8917929418ff1f4fe8f5b6623cf.

[neverhumanbot]

Clock eviction, row-lock FIFO, and morsel arena bounds. Exact WAL segment fill stays inside the limit. Required, audit, official parity, evidence guard, and packaging succeeded on 85bbec3.