[YoungTreeTree]

What and why

Pak v9 archives were skipped because their footer adds a FrozenIndex byte and moves the magic to 0xcd bytes from EOF. When that flag is set, accepting the version alone is insufficient: the index is a frozen UE4 memory image.

This reads bounded relative pointers, UTF-16 strings, FPakEntry arrays and nested sparse directory maps, retaining entry-relative compression blocks and the existing serialized-index path when FrozenIndex is clear. Encrypted-index key discovery checks the pointed-to mount. Shader discovery reads the directory before distinguishing inline/global-cache paks, shared shader archives and audio-only paks.

This addresses the Pak parsing/extraction failure reported in #2134. Related: #1618 handles ordinary v9 indices but explicitly rejects frozen indices; the Demon Slayer archives tested here use FrozenIndex=1.

Finding: Unreal Engine frozen pak indices (version 9).

Checked

• On the updated branch (ce370e1 base): all 63 test_uepak*.py tests (including eight new v9 tests), all 6 documentation tests, and the pinned Python lint ratchet pass (zero new findings).
• Synthetic coverage includes both index layouts, sparse-map holes, signed backward pointers, UTF-16 names, multi-block Zlib data, invalid ranges/references, deleted entries, encrypted-index key search and shader classification. No game-derived fixtures.
• Real archives: 1,284 + 88 entries agree with their serialized data headers; sampled configuration/audio and the 8,191,424-byte global shader cache decode to the indexed sizes.
• The shader split finds 57,982 shaders, with 1,519 first found in the pak's global cache. All 57,982 build for the Mac; the host and pipeline prebuild complete.
• A 45-second bounded random parser-fuzz run, adding the synthetic v9 pak to the existing harness seeds, completed 414,168 executions with no findings.
• GPU tests: make -C dm/gfx/tests -j2 GAME=0100309016E7A000 run passed.
• Game run: Demon Slayer -Kimetsu no Yaiba- The Hinokami Chronicles, 0100309016E7A000, 1.53 / v589824. Confirmed the Simplified Chinese autosave/new-save prompts and reached the title menu with the character-pack notification. Combat, full levels, long-term stability and iOS were not tested. The game remains on a provisional build profile; this does not add it to the supported-games table.
• Device: Apple M2 Pro, 32 GB, macOS 27.2.1, Xcode 27.2.

The real-game build/run, GPU tests and regression report below were collected on 547c5db plus this parser change before rebasing onto 034e91a. Upstream did not change the pak reader in between; parser/documentation/lint checks were rerun after rebasing. Runtime was not rerun against the newer host.

Regression report

scripts/regress.py check --games 0100309016E7A000: PASS, full mode, 324.7 seconds.

Baseline 547c5db647a2fc3880661d1357041d4dd1aae371; candidate the same revision plus this patch.

game	status	shaders (base → candidate)	newly failing	newly building	changed MSL	interface (new / known)	automated boot
Demon Slayer 0100309016E7A000	PASS	57982/57982 → 57982/57982	0	0	0	0 / 0	—

This was the only local game build available and directly exercises frozen v9 paks. Other games are not runtime-validated; existing v4/v8/v11 synthetic tests cover their parser paths. There is no registered smoke route for this provisional profile, so the report's boot/performance fields are empty; the manual menu observation above is separate. The harness compares the prepared shader package, while the real-archive and unit checks validate extraction.

CI checks

The branch is based on ce370e1. The save-sync runner-model check now comes from upstream; this PR no longer changes that test.

A separate test-only commit addresses the clang-tidy failures in the graphics tests: check nullable Metal resources before use; initialize and check the Vulkan swap chain; check the wide-line fixture paths and run its translator through NSTask; and free the sparse-copy test's linear command records after submission. The full LLVM run also regenerates the shrinking baseline to remove resolved findings. Runtime and shader-chain code are unchanged by this CI commit.

• All seven affected GPU tests pass on the M2 Pro with Xcode 27.2: t_res_commit (310 checks), t_upload_src (41), t_zero_depth (56), t_sparse_copy (43), t_vk_swapchain_life (141), t_wline (22), and t_pfv (zero failed). The latter four also ran with Metal API validation enabled.
• All 63 pak tests, 6 documentation tests and 19 lint-tool tests pass. The pinned Python lint ratchet reports 321 existing findings, zero new findings.
• LLVM 23.1.4 full clang-tidy validation passes across 312 sources: 204 findings, exactly matching the baseline, zero new findings and zero stale counts. The baseline shrank from 211 to 204 via lint_c.py tidy --update; no checks were disabled or findings added. Final check: python3 dm/tools/lint_c.py tidy -j 6 --against ce370e1.

The real-game and shader regression evidence above predates this rebase. No new game-runtime or iOS validation is claimed for the updated base.

Docs

• Reference finding added and python3 dm/tools/docs_symptoms.py run.
• No game files, keys or extracted game data in the change or attachments.
• Shader translation-chain inputs are unchanged; no chain-version bump is needed.