[wchwawa]

TLDR

Prepare Holt 0.8.8 to reject damaged file-backed blob reads on NoKV's current dependency line.

Problem

A changed byte in an active blob could reach a reader without a storage error. This is a private security follow-up; no public issue was opened under SECURITY.md.

Consequences

A caller could consume a wrong value, and a reused slot could expose bytes beyond a compact blob's persisted extent.

Example

A read of a damaged page now returns NodeCorrupt. A range read beyond the authenticated extent returns zeros.

Solution

Bind per-page checksums to each blob slot, sync the checksum record before the manifest mapping, and verify reads before returning them. Version 6 stores migrate on writable open. Compact reads fetch only the authenticated prefix; large Linux batches verify independent blobs in parallel and preserve result order.

Gain

Persistent blob damage fails closed. Sparse blobs avoid full-slot reads, and batch verification no longer runs serially.

Tradeoffs

CRC detects accidental damage, not an attacker who can rewrite storage metadata. Existing damage at migration time cannot be authenticated. Version 6 read-only opens need a prior writable migration. Dense-frame checksums still have material cost, so performance acceptance remains open and this PR stays Draft. NoKV remains locked to published Holt 0.8.7 until 0.8.8 is released and separately integrated.

Validation

• cargo test --workspace --locked -- --test-threads=2 passed on macOS before a final formatting-only change. Exact committed source passed macOS file-store tests 52/52 and Linux/arm64 Docker io-uring file-store tests 58/58, including a 40-blob damaged-page batch.
• cargo fmt --all -- --check, macOS workspace Clippy, Linux io-uring Clippy, and git diff --check passed. cargo publish --dry-run --locked packaged and verified 128 files.
• A published 0.8.7 writer's checkpointed store was migrated and read by 0.8.8; a 0.8.8 writer's store was read by the fixed 0.9.2 line. The reverse 0.9.2-to-0.8.8 WAL path is not qualified.
• Matched 18-pair synthetic frame benchmarks with crc32fast 1.5.0: sparse full reads changed by -47.1% on macOS and -13.2% in Linux/arm64 Docker; dense full reads changed by +33.6% and +56.6%, respectively. Dense batch reads changed by +8.5% and +35.8%. These are fresh-open microbenchmarks, not NoKV product latency evidence. Linux used relaxed Docker seccomp for io-uring.
• The existing CI workflow does not trigger on pull requests targeting release/0.8. The soak and benchmark lockfiles now match Holt 0.8.8, and both locked harness checks pass. A manual Nightly Validation run with 60-second soak/fuzz inputs passed on this exact HEAD (Verus was not enabled); it is not a full nightly or a PR check.

This is the 0.8-line backport of #80. No crate was published.