v0.2.0 — Performance Campaign

FrankenFS v0.2.0 — the project's first tagged release, consolidating a solo, negative-evidence-ledger-first performance-optimization campaign against the mounted kernel filesystem (ext4 / btrfs).

Every kept lever is measured (honest same-worker A/B in one binary, gated on median self-time vs a paired null control, cv < 5%), isomorphism-preserving (behaviour proven byte-identical; btrfs images pass btrfs-check, ext4 pass e2fsck), and profile-first. One lever per commit. Rejected levers are recorded with their null-control result and a retry condition rather than deleted.

Headline measured wins

• Metadata: name-index closes the create existence-check O(N²) → 26× slower-than-kernel to kernel parity (1fcd0b62); delete htree fast-path O(N²)→O(log N), gap 5.3× → 2.1× (85d46923).
• Write allocation: coalesce contiguous extents O(N²)→O(N) — 120× at N=40k (2aa92946).
• Allocator: binary range-overlap reserved-check up to 3110× (af91cc18); 15.8× fewer delete preads (b296dbdb).
• Journal replay: memoize indirect-block resolution — fixes a 2024× re-read at every mount (fe00c75e).
• Checksum: incremental crc32c up to 24.7× vs full recompute (ff222d17 → 17380f69); incremental dir-block csum 10.3× (56cb5f94).
• Extent resolution: binary-search leaf+index O(E)→O(log E), up to 19× (2785e425); SWAR namespace hash 7.14× (96c27663).
• MVCC: skip per-read snapshot register/release — 5× parallel random read (9376f4d6); drop 4096-realign — 1.82× commit (eb229915).
• btrfs: read directly into caller buffer — 1.37× warm, RSS halved, beats kernel (54b0ae94); fan-out-gated prefetch — 4.3× metadata walk (18fb0e88); COW insert_many/remove_many batching across create/mkdir/delete/rename.
• CLI: jemalloc global allocator — create 1.26–1.6×, now faster than kernel single-thread (14f443cb).

Honesty methodology

Claims are median-gated against a paired null control (not best-of-N); wins are byte-identical to prior behaviour; and every rejected lever lives in the negative-evidence ledger (docs/NEGATIVE_EVIDENCE.md, docs/PERF_CAMPAIGN_FINAL.md). Where honest scrutiny contradicted an earlier claim it was corrected in the open — a previously-reported "1.7× faster than kernel" cold-read row flipped sign under a corrected harness (frankenfs is actually ~1.42× slower there), and that finding is documented rather than buried.

Built #![forbid(unsafe_code)]; release build is green (remote-only rch gate). Full per-subsystem win list and reject ledger in CHANGELOG.md.