Revision history for Cavil::Matcher

1.04   2026-08-27
        - FpIndex now caches opened segments and reuses them across searches instead of reopening every
          segment on every query. Opening a segment is a fixed per-query cost that otherwise dominated search
          time on a long-lived query server, regardless of the query. Segments are immutable append-only
          files, so a cached handle always maps the same bytes; a handle whose segment has been compacted out
          of the manifest is dropped on the next search. Results are unchanged.
        - score() and FpIndex::search() take an optional minimum containment; matches below it are dropped
          inside the scorer instead of being returned for the caller to filter. A query dominated by common
          fingerprints can match hundreds of thousands of contents, almost all coincidental, so applying the
          floor at this level avoids building and returning them. Defaults to 0 (return everything).
        - score() and FpIndex::search() also take an optional max_df: query fingerprints appearing in more
          than that many records of a segment are ignored (boilerplate that matches nearly everything, adding
          only coincidental hits and skewing containment). Off by default (0). Intended for a compacted
          single-segment index, where a segment's record count for a fingerprint is its document frequency.

1.03   2026-08-27
        - Fingerprint scoring now reports, for each matched fingerprint, the query fingerprint value it was,
          appended to each region tuple after the existing start line and span. This lets a caller map
          matches back to their query positions - enough to tell an aligned copy from scattered coincidental
          hits, and to derive each query fingerprint's document frequency from a full result set. Backward
          compatible: the (start_line, span) pair at the front of every region tuple is unchanged.

1.02   2026-08-27
        - Snippet code search: a new winnowed-fingerprint index for "where does this code already exist"
          provenance queries, alongside the license-pattern matcher. fingerprint_file winnows a file's
          tokens (k tokens per gram, w grams per window) into content fingerprints; a fingerprint segment
          stores them keyed by a 128-bit content hash rather than filenames (so it stays lean and joins to
          the Cavil database), and scoring ranks candidates by both-direction containment with exact
          matched-line spans for highlighting.
        - Cavil::Matcher::FpIndex is the pure-Perl lifecycle over a directory of fingerprint segments
          (incremental add_segment, manifest/generation, memory-mapped shared reads), reusing the same
          on-disk discipline as the pattern index: versioned, CRC-checksummed, structure-validated on
          every open, with corrupt or foreign files rejected rather than mis-read.
        - Byte-identical files are de-duplicated by content hash, so the same source shipped across many
          package versions is stored once, with no package-name parsing.
        - Additive only: the license-pattern format is unchanged and no migration is needed.

1.01   2026-08-04
        - Big indexing speedup: the scan path (load/attach) no longer re-checksums a segment's whole
          payload on every open. A compiled segment is an immutable, atomically-published cache that is
          already CRC-checked when written, so re-CRCing hundreds of megabytes in every forked index job
          was almost the entire cost of a load (~97% in a 160 MB corpus: ~293ms -> ~8ms). Structural
          validation still always runs, so bad segments are rejected and scanning stays memory-safe;
          only per-scan detection of on-disk bit-rot is dropped.
        - CRC is now verified where it is meaningful: once at publish (dump() fully verifies before the
          atomic rename) and on demand via the new Engine verify($file) method (an fsck for a cache dir).
        - Same treatment for the closest-match "bag of patterns" cache (Bag::load, used by analyze): the
          load path trusts the payload CRC (the record parse still bounds-checks every read), with a new
          Bag verify($file) method and a publish-time check.
        - No on-disk format change: existing segments load as-is, no reindex, and old/new engines
          interoperate on the same files (in-place upgrade is seamless).

1.00   2026-07-22
        - Initial release: the next-generation license pattern matcher for Cavil.
        - Segmented, incrementally-updatable index: adding or removing a pattern no longer
          rebuilds the whole compiled cache (base + append-only delta segments + tombstones).
        - Versioned, CRC-checksummed, fully-validated on-disk segment format (no unchecked
          reinterpret_cast); corrupt or hostile segments are rejected, never mis-read.
        - Segments are memory-mapped read-only and queried in place, so index workers on a host
          share one physical copy instead of each materialising its own.
        - Pure-Perl manifest / segment lifecycle (Cavil::Matcher::Index, ::Manifest); a small,
          frozen C++ core (tokenizer, scan, SpookyV2 hashing) kept bit-for-bit compatible with
          Spooky::Patterns::XS, so no database migration is needed to switch engines.
        - Hardened against arbitrary input (binaries, malformed samples): never crashes.
        - 100% statement/branch/condition coverage of the Perl surface; developer-only
          differential tests against the previous engine in xt/.
