[thehobbit85]

Fixes #1296. Short NDSR shader packs can end in a partial 16-byte record. The scanner now scores only complete records when choosing between header offsets 0x10 and 0x14. It also retains main’s 8-byte walk past non-shader entries.

Current validation

Tested cd03af43 against main b2d46b86 on Apple silicon. The saved 127-byte input triggered Python’s record-length error on that main revision. The rebased scanner passed 128 tests covering NDSR layout and pairing, NDVE, related splitters, fuzz seeds, docs and the shader-chain header. Pinned Ruff and Pyright reported 371 baseline findings, zero new and zero stale. The symptom index was regenerated from the combined reference entries.

Donkey Kong Bananza and Splatoon Raiders use the two header layouts. Star Fox embeds these packs inside VIPR containers. The regression harness found no Mac builds of those three games. No game resplit, gameplay or device result is available for this revision.

Earlier implementation and validation

Short NDSR packs can leave a partial 16-byte record in either candidate header slice. struct.iter_unpack then raises before the reader can choose the layout, including for the saved 127-byte synthetic fuzz input. Score at most 16 complete records at each offset, preserving layout selection, shader contents and vertex/pixel pairing.

Fixes #1296. Related to #313's parser robustness work.

Adds the saved input to the automatically replayed fuzz corpus, plus tests for short lengths, both record starts, trailing-byte remainders, valid shader extraction and pairing. Updates the existing NDSR reference and symptom index; no shader translation chain or runtime changes.

Validation on Apple silicon macOS, using existing tools and synthetic fixtures only:

• The exact saved input failed unchanged main f8a5618d with struct.error@shader_split.py:ndsr_shaders>records_from; it passes with this change. The additional boundary tests failed 268 subcases before the fix and pass afterward.
• All 35 corpus tests and all 3 NDSR boundary tests pass.
• Ruff 0.17.2 / Pyright 1.1.414 ratchet: 389 existing findings, no new findings and no stale baseline entries.
• Shader-parser random fuzz engine: 266,803 cases over 60 seconds, no findings.
• Python CI file loop: all 151 files pass (1,564 tests, 13 optional skips). An initial Rosetta-launched run failed 14 architecture-sensitive files; all 14 pass when explicitly launched with arch -arm64 using the existing ARM64 .NET 10 runtime, without code changes.
• make -C dm/gfx/tests GAME=010015100B514000 -j3 run: passed on the Mac (existing compiler warnings and optional external reflection samples reported by the suite).

Impacted-game rationale: Donkey Kong Bananza uses the 0x10 record layout, Splatoon Raiders uses 0x14, and Star Fox embeds NDSR packs in VIPR containers. These are the narrow affected paths. No local game inputs are available for this task, so no shader resplit, gameplay, or device validation is claimed. The original archived fix and seed remain preserved.

regress.py impact f8a5618d..HEAD reports 3 broad matches, 1 unaffected profile and 141 unknown profiles; its file-level shader-split rule is broader than this NDSR-only change. regress.py check --baseline f8a5618d --candidate f72ee4e7 --games 040074A01BF12000,0400380020056000,0400E88025B1A000 was attempted against an explicitly empty input directory: exit 2, no Mac builds found, so no gameplay report or validation stamps were produced.

Coverage check before implementation: all 426 issues and 863 PR descriptions, relevant issue timelines, and 104 relevant open/draft/closed/merged PR diffs. #1255/#1258 retain the old scorer; #1259's embedded-pack support and #1278's NARC bounds fix do not resolve this failure. Coverage was rechecked after validation: main remains f8a5618d; the 865-PR inventory and the issue timeline contain no equivalent solution. PRs that changed during this pass were inspected too, including #1297's separate NDVE reader, which leaves the NDSR scorer unchanged.

[seedds]

Landed on main as ff33639, unchanged, with you as co-author. Thanks! I confirmed the new test's 268 subcases and the saved corpus input fail on the old scorer and pass with yours. Wherever the old score did not raise, the new one gives the same count, so no game's shader split changes. #1296 is closed by the commit.