feat(photocd): read Kodak Photo CD, the first format the bit-field decode unlocked

[swackhamer]

Why

#640 made 993 masked fractional entries decodable and measured the corpus delta at exactly zero — because ~980 of them sit in tables no caller looks up. Only about twenty of the 612 generated tables are reachable through find_table at all. That PR said so rather than claiming a win. This one spends some of it.

PhotoCD::Main is the highest-value unreached table in the corpus: it declares 24 fields, ExifTool reports all 24 on combined-samples/PhotoCD.pcd, and OxiDex reported none. PCD was detected but unparsed — a correct File:FileType: PCD beside no PhotoCD group at all, the failure mode that looks healthy in oxidex -j.

Three of those tags are why this format waited for #640: ImageWidth, ImageHeight and CompressionClass are 1538.1/.2/.3, fractional entries sharing byte 1538 with Orientation, each masking a different slice. decode_binary_table refused every fractional entry before #640, so the table could not have been read from.

Measured coverage: +24 tags

Full 4246-file corpus against the pinned 13.59, before and after on this exact base:

	before	after
distinct keys	6825 / 8535 (79.96%)	6849 / 8535 (80.25%)
tag instances	352925 / 364486	352949 / 364486
missing_in_oxidex	1668	1644
value_differences	42	42
regressions	0	0

Every per-format metric outside PCD is byte-identical. PCD moves matched_tags 0 → 24, missing_in_oxidex 24 → 0.

The harness also reports extra_in_oxidex 0 → 2 for PCD. Both are Composite:ImageSize/Megapixels, which ExifTool does emit with identical values — a direct exiftool -G -j -a comparison is 29 keys each side, 0 missing, 0 extra, 0 value differences. 273 composites are already classified that way corpus-wide; it is how the harness invokes ExifTool, not an over-claim.

The table alone was not enough

Nine fields carry a ValueConv, three a RawConv, five a Condition — none of which the transcription runs. Each is implemented against the Perl it reproduces and cited by line, rather than skipped:

• int8u[2] version pairs are tr/ /./, a join: 7 61 is release "7.61". Read as a number it prints 7.61 by luck and turns "1.0" into "1". A pair of 255s is RawConv'd to "n/a".
• ScannerPixelSize is join(".",unpack("H2H2",$val)) — hex digits, so 0x11 0x48 is 11.48 micrometers, not 17.72.
• CreateDate/ModifyDate are ConvertUnixTime($val,1): local civil time with the offset in force at that instant, reusing the same helper the File group's timestamps use. 0xffffffff means absent, not 2106.
• ImageWidth/ImageHeight are ($$self{Orient} & 0x01 ? 512 : 768) * ($val * 2 || 1). The DataMember comes from Orientation's RawConv, and Perl's || makes resolution code 0 a multiplier of one rather than collapsing the image to nothing.
• Five scene-balance tags are Condition => '$$self{HasSBA}', gated on the literal bytes SBA at 225; CopyrightFileName additionally needs CopyrightStatus == 1. A file without the marker must lose those tags, not render them from whatever occupies the offsets — there is a test for exactly that.

Detection

A PCD_IPI signature at offset 2048, where PhotoCD.pm:454 seeks. ExifTool publishes no %magicNumber for PCD, but this is not an over-claim: the pinned 13.59 reports FileType: PCD for a copy renamed .dat and for one with no extension, because ProcessPCD is what confirms the type. The offset is past the 1 KiB probe, so it resolves through the same short-read fallback ISO 9660's CD001 at 32769 already needed.

The parser contributes a bare FileType, which normalize_identity_tags uses only to fill an Unknown — the one identity answer a parser is allowed to supply.

Verification

• Direct group-qualified per-file diff vs pinned 13.59: 29/29 keys, 0 missing, 0 extra, 0 value differences
• cargo test --workspace — 47 suites, 4876 passed, 0 failed (11 new PhotoCD tests, all synthetic — they build the record in-memory rather than reading the sample corpus, which CI does not have)
• just verify-tables — PASS vs 13.59, 0 mismatches
• cargo clippy --all-features -- -D warnings, cargo fmt --all --check, tools/ci/check-corpus-guards.py, scripts/sync_tag_stats.py --check — all clean

Rebased onto 303b33ad (#657, which also touched detection); the one conflict was two adjacent imports, and both formats' wiring verified present afterward.

🤖 Generated with Claude Code