From the OrdinalDB builder post-mortem (F7) + 2026-07-05 remediation reconciliation: cold-open verification is hash-bound and single-threaded (~290 MB/s single-stream SHA-256 measured by brief-02 at 300k–1M vectors; open ≈ 200× search cost per invocation).
Current state: hashing is already bounded-streaming with EINTR retry (sha256_file_bounded). Artifact-level parallelism (primary ∥ aux) helps little because the primary artifact dominates bundle bytes. The honest win needs chunked-parallel hashing of the primary artifact or an algorithm change (e.g. blake3-class tree hashing) — a format/design decision, not a patch. Deferred from the 0.2.0 hardening wave by decision; the wave ships verify-once composition downstream (hash once per open pattern) which captures the architectural win. A cached verification token was considered and rejected (mtime-keyed tokens are the byte-pin/TOCTOU trade the crate's docs explicitly refuse).
From the OrdinalDB builder post-mortem (F7) + 2026-07-05 remediation reconciliation: cold-open verification is hash-bound and single-threaded (~290 MB/s single-stream SHA-256 measured by brief-02 at 300k–1M vectors; open ≈ 200× search cost per invocation).
Current state: hashing is already bounded-streaming with EINTR retry (
sha256_file_bounded). Artifact-level parallelism (primary ∥ aux) helps little because the primary artifact dominates bundle bytes. The honest win needs chunked-parallel hashing of the primary artifact or an algorithm change (e.g. blake3-class tree hashing) — a format/design decision, not a patch. Deferred from the 0.2.0 hardening wave by decision; the wave ships verify-once composition downstream (hash once per open pattern) which captures the architectural win. A cached verification token was considered and rejected (mtime-keyed tokens are the byte-pin/TOCTOU trade the crate's docs explicitly refuse).