Verify the hunk CRC32 that v3/v4 map entries carry

[rtissera]

CHDv3/v4 map entries hold a CRC32 of the decoded hunk plus a flag to opt out of it. map_extract() has always parsed the field into entry->crc, and nothing ever read it, so on the pre-v5 path a hunk that decompressed cleanly to the wrong bytes was handed back as valid data. The v5 path has checked its CRC16 all along. MAME rejects the same files with a decompression error.

The gap also undercut the reasoning behind DR_FLAC_NO_CRC: dr_flac's own frame CRC is dropped because the hunk CRC is said to cover the same data, and CHDCOMPRESSION_AV on a v3/v4 file routes through the same libchdr_flac.c with no hunk CRC being checked at all.

The check mirrors MAME: COMPRESSED, UNCOMPRESSED and MINI entries, the 0x10 opt-out honoured, self- and parent-referenced entries covered when the hunk they point at is read. v1/v2 entries have no room for a CRC and map_extract_old() already sets the opt-out flag for them.

Testing

tests/corpus/mklegacy.py builds the v3/v4 layouts directly, since chdman only writes v5. It emits an intact pair, two files whose stored hunk no longer matches the CRC the map advertises, and one that sets the opt-out flag on that same hunk. All ten cases agree with MAME 0.264 — the intact and opt-out files verify, the other two are refused. tests/legacy_decode.c asserts that, and fails on master with four got "no error", want "decompression error".

Also checked against three real CHDv4 images (76, 75 and 1817 hunks), which decode identically to chdman extractraw, and against the v5 corpus plus a large real disc, whose output is unchanged. Built and run under the default configuration, CHDR_LOWRAM_TARGET=ON, WITH_SYSTEM_ZLIB=ON and CHDR_VERIFY_BLOCK_CRC=OFF.

The fixtures need only a Python interpreter, unlike the CHDv5 corpus, so the CMake job builds them; the test skips itself where they are absent, as parent-decode does.

Not covered: no v3/v4 AVHuff fixture. The check sits after decompress() regardless of codec, so AV is covered by construction, but nothing exercises it.

Cost

About 1.1 KB of text on RV32 at -Os, nearly all of it mz_crc32's table. Nothing extra against a shared system zlib, and nothing when VERIFY_BLOCK_CRC is off. mz_crc32 joins the ESP ROM rename list, since libchdr's own sources now name it.

🤖 Generated with Claude Code