v1.4: preserve consecutive livery parent/child transforms

[Datamining00]

Fix derivative of v1.4-parser-structural-audit.

Root cause confirmed by the ownership trace

The supplied Parent Ownership Diagnostic shows that the earlier bare-parent recovery at offset 271968 is now correct: it owns exactly two child groups (272009 and 272321). However, the remaining central XANADU / THE NEW WORLD artwork is a different parser failure.

The raw stream around offset 269788 contains:

• a parent transform (-360.5, -31.0, 0.9000001, 3.5);
• immediately after it, the exact extended child marker 00 02 00 01 00 00 00 03 with child transform (0.0125732, 2.4793174, 0.05000036, 0);
• then the first counted child group at 269828.

Unlike the previous bare-parent case, this parent is already recognized by the pinned decoder as a normal zero-marker livery transform. The bug is that walk_step() stores only one pending_transform; when the extended child transform follows, it replaces the already-recognized parent instead of preserving parent/child ownership.

The result is structurally valid layer counts but a wrong GroupNode tree, which explains why Structural Audit v2 correctly reported unframed_transform_candidate_count = 0 while the artwork was still misplaced.

Fix

Adds a narrow consecutive-transform grammar rule:

• requires an existing clean pending parent whose marker is exactly 00;
• requires the current record to use the exact extended child marker 00 02 00 01 00 00 00 03;
• requires a valid livery group boundary immediately after the child transform;
• materializes the pending parent as an implicit two-child GroupNode;
• keeps the second transform pending for the first child group;
• relies on the existing stack completion logic to keep the parent open for the second child.

No car ID, section name, phrase, creator, source offset, or coordinate is hard-coded. No layer reversal, mask rule, source-offset normalization, or renderer/projection behavior is changed.

Validation

• Windows CI: success.
• Full regression suite: 178 tests, OK.
• New consecutive-transform regression tests: all passed.
• PyInstaller one-file build: success.
• EXE SHA-256: 5a386de6a7f61b27e5cfdbbbb5606685e8fda5fe4356856f73ae60f63af595b1.
• Artifact ZIP SHA-256: fbf532baebfaeb72b0316070010031e1aa1054f2a3e1d7d72644f7534607bea6.

Validation target

On the supplied Livery_0343 Right section, this should move the currently central XANADU/THE NEW WORLD group family under its real parent transform instead of leaving it around the character face.

This PR stays draft until the real livery is visually checked.