v1.4: recursive occlusion isolation and native-mask counterpart diagnostics

[Datamining00]

Follow-up test branch after the consecutive-transform fix was visually verified on the previous livery.

Current real-world findings

1. The Left-side occlusion probe first localized the large black blocking artwork to layers 2416-2761; the supplied refined sheet narrows it further to layers 2718-2761. The other seven refined ranges contain only small/sparse artwork.
2. The separate Right-side failure is confirmed to be a mask at layer 655 with shape word 0x0100, type 0x100100, and no resolved native resource. It must not be silently skipped.
3. The v2 raw window revealed a diagnostic addressing mistake: decoder source_offset is relative to the C_livery layer-data body, while v2 sliced the full decompressed payload directly. In the supplied sample the real direct 02 00 01 ... record is exactly one layer-data base (0x48 / 72 bytes) after the v2 nominal position. v3 resolves the body base before reading raw bytes.

v3 simple diagnostic policy

• Rendering behavior remains unchanged.
• The dark-occluder probe now recursively follows only the range with the largest near-black coverage until at most 8 layers remain, then renders those layers individually.
• Outputs include *-occluder-pass-N.png, *-occluder-candidates.png, and *-occluder-candidates.json with source offset/type/shape/mask metadata for the final candidates.
• Fatal missing native masks still fail closed.
• The v3 fatal-mask JSON resolves the body-relative source offset to the actual raw record and ranks mask layers from the opposite side by mirrored transform similarity. This is intended to identify a valid Left-side counterpart for a Right-side unknown mask (or vice versa) without guessing geometry.
• Missing ordinary visible native shapes retain the existing skip-and-log test policy.

No car ID, creator, livery name, section coordinate, layer number, shape word, or resource alias is hard-coded into runtime behavior. Existing parser transform fixes remain unchanged.

Validation

• Windows CI: success.
• Full regression suite: 189 tests, OK.
• v3 source-offset and opposite-side mirror-ranking tests: passed.
• PyInstaller one-file build: success.
• EXE SHA-256: b2ad040ba360f37b19764f3102b85f2aae92afa2d221b065978499bade91af51.
• Artifact ZIP SHA-256: 64057d545b8b18ade19f966d41dbe79d577ff95e2ca7f7f2aaa260595b6d5669.

This PR remains draft until the final Left candidate sheet/JSON and the new Right v3 mask report are checked on the supplied real-world cases.