Original-code oracle matrix

The isolated oracle uses FS-UAE 3.2.35, Workbench 3.1, xpkmaster 4.16 (21-Aug-1997), and recovered compressor libraries. ROMs, Amiga binaries, private files, logs, and the complete workspace are not distributed.

xQuery reported these mode ranges:

Codec Default Ranges
NONE 50 0–100
NUKE 50 0–100
FAST 50 0–79, 80–100
RAKE 100 0–25, 26–50, 51–75, 76–100
HUFF 50 0–100
SHRI 100 0–14, 15–28, 29–42, 43–56, 57–70, 71–84, 85–100
DLTA 100 0–100
SMPL 50 0–100
HFMN 0 0–100
MASH 100 0–100
SQSH 100 0–100
LZBS 100 0, 1–10, 11–20, 21–30, 31–40, 41–50, 51–60, 61–70, 71–80, 81–90, 91–100
SLZ3 100 0–100
SDHC 50 0–7, 8–15, 16–23, 24–31, 32–39, 40–47, 48–55, 56–63, 64–71, 72–79, 80–87, 88–100
LHLB 100 0–100
BZP2 40 0–19, 20–29, 30–39, 40–49, 50–59, 60–69, 70–79, 80–89, 90–100
GZIP 65 0–9, 10–19, 20–29, 30–39, 40–49, 50–59, 60–69, 70–79, 80–89, 90–100
IMPL 100 0–10, 11–30, 31–50, 51–75, 76–98, 99–100

The matrix exercised every numeric mode on a canonical text vector and every range boundary/default on deterministic pattern and size vectors:

All 1,026 emitted containers begin with XPKF. Small representative fixtures for all six codecs are committed as reviewable hexadecimal files and validate the current pure-Python implementations.

A second isolated run added CBR0, RLEN, FRLE, RDCN, BLZW, and DUKE. Across the completed non-empty cases, all 372 original packed/plain artifacts decode byte-exactly in Python. Compact acceptance containers for each codec are committed as hexadecimal fixtures. The historical packers are run in separate or short-lived emulator sessions because repeated use can crash the original Amiga process; this is an oracle limitation rather than an accepted decoder failure.

A third set covers DLTA, SMPL, HFMN, MASH, and SQSH. Short isolated runs produced 35 byte-exact original pack/unpack artifacts across modes 0, default, and 100. Python directly decoded all 20 generated compressed DLTA, HFMN, MASH, and SQSH containers. The SMPL packer selected raw XPK chunks for all 15 vectors, so its decoder additionally has a source-derived synthetic prefix-code test rather than claiming compressed oracle coverage. DLTA hangs on a one-byte input and HFMN hangs on one broad byte-pattern input; those original-packer defects are excluded from the safe fixture stage.

Every compact committed acceptance container is recorded in the public manifest with its input, output, size, and SHA-256 hashes. Known hanging packer/vector pairs are machine-readable exclusions and are never emitted into a guest stage. oracle.py verify-python performs the direct comparison and records its results as JSON.

The CyberYAFA LZW2–LZW5 variants are now backed by preserved original libraries and headless UAE fixtures across every reported mode boundary. SHR3 remains source-derived because no corresponding original library has been located; it is additionally tested by converting both packed chunks from the 64 KiB SHRI continuation fixture to its headerless chunk representation.

Preserved original ACCA, FBR2, ILZR, and ZENO libraries were also located and exercised under UAE. Together with the LZW family, 157 generated artifacts round-trip in the original master library and decode byte-exactly in Python. Compact fixtures, private library hashes, and the preservation archive hash are linked in the public manifest; the binaries themselves are not redistributed. ARTM remains source-derived because its original library has not been located. Compact source-derived XPK containers for these codecs were also independently compared byte-for-byte with Ancient 2.3.0 where Ancient supports the emitted variant.

The next batch adds LZBS, SLZ3, TDCS, LHLB, SDHC, and CYB2. Preserved LZBS, SLZ3, and SDHC packers produced 83 completed artifacts across every reported mode boundary; Python decoded all of them byte-exactly. Low-mode LZBS output also confirmed that its original depacker stops a final literal run at the declared output size even when the packer rounds that run up. Representative original-Amiga fixtures for all three codecs are committed.

LHLB initially appeared unpack-only because its required lh.library was missing. Installing the original Aminet dependency exposes its 0–100 packing range; a mode-100 artifact now round-trips in the Amiga master library and decodes byte-exactly in Python. The dependency hash is recorded in the public manifest, but the binary is not redistributed.

CYB2 still reports no packing modes. Its preserved library also requires xpkcybhandle.library; with that dependency and its CYB1 companion installed, the original library continues to reject source-constructed CYB2 containers before unpacking. Consequently CYB2 remains source-derived rather than being mislabelled as original-Amiga verified. TDCS has no library in the preservation archive. Both retain synthetic success and malformed-stream coverage.

Preserved BZP2, GZIP, and IMPL packers generated representative compressed fixtures that round-trip under headless UAE and decode byte-exactly in Python. The complete boundary/default matrix now contains 36 BZP2, 42 GZIP, 24 IMPL, and six LHLB byte-exact successes with no Python mismatches.

PWPK, CRM2, and CRMS expose no packing modes in the preserved XPK libraries. An original Amiga PPMC utility generated the positive PowerPacker stream used by the PWPK fixture, which was also checked with Ancient. CRM2 and CRMS use positive streams generated and decoded by an independent Crunch-Mania implementation. Their XPK envelopes remain source-derived and are labelled accordingly.

FS-UAE runs behind Xvfb with Mesa software rendering, so oracle generation does not map a window onto the host desktop.

The recovered private wrapper beginning 01 80 63 68 05 61 01 0a was not recognized as packed data by the original master library. That result rules out using generated XPKF files to infer its outer framing.