[BUG] ISA-L backend (auto-selected) produces PFSC blocks that the PS5 hardware inflater rejects #132

🐞 What happened?

On macOS arm64, mkpfs pack folder auto-selects the isal compression backend. The image passes mkpfs verify with 0 errors. On the PS5, a small number of PFSC blocks cannot be decompressed by the console's hardware decompressor (ZCN). The console returns wrong data for those 64 KiB blocks, and the game crashes a few seconds after launch.

The affected blocks are byte-identical on the console and the PC, they decompress correctly with software zlib, and they equal isal_zlib.compress(block, 2) exactly. Rebuilding the same folder with --compression-backend zlib fixes it completely: the game runs and no ZCN errors occur.

📦 Versions (required)

• MkPFS version (mkpfs -V): MkPFS 1.0.0 (pip), isal 1.8.0
• Python version (python --version): Python 3.14.7 (macOS 26.5.1, Apple M4 Pro, arm64)
• PS5 FW Version: 13.42 (standard PS5), ShadowMountPlus 1.7beta2
• Target image format/strategy you attempted (check all that apply):
  • .exfat->.ffpfsc
  • .ffpkg->.ffpfsc
  • raw-folder->.ffpfs (direct / single-pass)
  • raw-folder->.dat->.ffpfsc (wrapped / double-pass)
  • Other (please specify): raw-folder->.ffpfsc via pack folder (default: exFAT-wrapped, single pass, "no temp image")
• Did you use a UI to convert or the command line?
  • UI (please specify which)
  • Command line

💻 Command input (required)

# Failing build (backend auto -> isal)
python -m mkpfs pack folder --cpu-count 10 --temp-folder ~/Downloads/ps5-images "staging/PPSA11386-app0" "PPSA11386.ffpfsc"

# Working build (only change: zlib backend)
python -m mkpfs pack folder --compression-backend zlib --cpu-count 12 --temp-folder ~/Downloads/ps5-images "staging/PPSA11386-app0" "PPSA11386-zlib.ffpfsc"

📤 Command output

Output of the failing build (isal). It completes without errors:

MkPFS 1.0.0 - https://github.com/PSBrew/MkPFS
PFS Image Builder - Parameters
  Source path:       ~/Downloads/ps5-images/staging/PPSA11386-app0
  Output path:       ~/Downloads/ps5-images/PPSA11386.ffpfsc
  Version:           2 (PS5)
  Compression Setup: PFSC (0x43534650)
  Block size:        64 KiB (65,536 bytes)
  PFS mode:          0x0008  (case insensitive)
  Compression:       enabled
    Skip executables: yes
  CPU cores:         10
  Zlib level:        7
  Max PFSC ratio:    100%
Wrapping ~/Downloads/ps5-images/staging/PPSA11386-app0 into an exFAT and compressing to ~/Downloads/ps5-images/PPSA11386.ffpfsc (no temp image)...
Inner image: PPSA11386.exfat (80.90 GB)
Compressing inner exFAT (80.90 GB)...
Successfully wrote 51.52 GB image
Detected fakelib/libSceAmpr.sprx; generating ampr_emu.index...
Generated ampr_emu.index with 34 entries
Build Summary
  Uncompressed size:       80.90 GB (86,868,361,216 bytes)
  Final image size:        51.52 GB (55,319,986,176 bytes)
  Actual gain achieved:    36.32%
  Elapsed time:            99.79s

mkpfs verify on this image: Warnings: 0, Errors: 0.

📎 Extra context

Console kernel log right before the crash:

[IOD] ZCN ErrorNotification: ZcnError:0x22 opc:0x4e  sysEc:0x79 isIB:0 eRing:0x8  ringLoc:0xbc58
[IOD] process_completion: error opc: 0x4e, ec: 0x79, ering: 0x8, loc: 0xbc58, eaddr: 0x22, rid: 0x8bc58
[IOD] process_completion: conv error: 0x79 -> 0x5
ERROR: inflate(345) a53io_wait_completion() failed. 0 5

This happens identically from a USB stick and from internal storage (/data/homebrew).

Analysis:

1. I mounted the image read-only through ShadowMountPlus ("Mount R/O"). I then read 130 sampled 1 MiB ranges over FTP and compared them with the source files. Two samples each had exactly one wrong 64 KiB block, deterministically across re-reads.
2. I mapped the offsets through the inner exFAT to PFSC blocks #397847 and #844240 (of 1,325,506).
3. For both blocks:
   • the raw compressed bytes on the console are identical to the PC image;
   • zlib.decompress() gives the correct original data;
   • the stored block equals isal_zlib.compress(original, 2) (header 78 5e, 18017 / 19278 bytes).
4. With --compression-backend zlib, the same blocks become standard zlib level 7 (78 da), and the game runs without errors.

Suggestion: don't auto-select isal for PS5 PFSC images. Prefer zlib, or zlib-ng if it is verified, as the default, or at least warn about it in the help text. I haven't tested whether zlib-ng output is affected.