Original-resolution car preparation above the graphics handoff repair

Only this additive Patch Pipeline folder is authored. Final vehicle_package.cpp
and vehicle_rt64.cpp inputs are copied outside the checkout. No Package ABI,
guest function, allocator policy, wire format, peer ID or launcher design is
changed. Submodules, RecompiledFuncs and RecompiledPatches remain read-only.

Only the explicit original-resolution-v1 profile gets tile preparation. Unknown
profiles remain rejected. The native-4kib-rgba32-safe-v1 promise and all existing
PNG, archive, decoded-memory, LOD, ownership and index checks remain. Existing
profile/no-profile packages bypass this preparation and retain identical parsed
data. The original ZIP bytes remain the identity and online distribution source.

Material interiors are at most 30 texels in each split axis, with one neighbour
pixel on both edges. Whole axes already fitting 32 texels remain whole. This
preserves point/filter edge samples, wrap repeats including negative periods,
clamp overflow, exact PNG RGBA8 channels and the image's original resolution.
Each stored tile fits the existing two 2 KiB RGBA32 TMEM banks. No new GPU API,
replacement-texture layer, render target or larger TMEM limit is introduced.

Preparation clips source triangles at material boundaries using canonical edge
intersection order. Prepared pieces retain mesh colour, culling and wheel ID;
wheel pivots are computed from the original geometry before splitting. The
existing per-piece interpolation index remains bounded by 64 draw groups. Limits
remain 32 prepared tiles, 8192 vertices and 4096 triangles per LOD, with a bounded
clipping work budget. All six LODs are prepared before publishing the result.
The renderer continues using its existing material loads and 30-vertex batches.
More geometry is the cost of full original artwork on this existing path; no
per-frame clipping or resampling occurs. Existing preparation memory accounting
counts the resulting arrays.

The explicit dkrkart.per-base-driver revision 1 contract is validated separately:
coordinate axes, raw native anchor units, Studio seat units, transform exclusions
and the supplied anchor-table hash/values must agree. Without that contract,
older per-driver/legacy Studio-unit rules remain. The existing authored_anchors
flag is used only with PerBaseDriverV1 to mark this explicit raw-native contract;
its older shared-seat meaning is retained. No struct layout change is required.
The render adapter subtracts raw anchors in native model coordinates, then uses
the existing seat/world conversion. It never rescales the driver's animation.

Disabled revision-1 boost metadata preserves the native effect position. Enabled
custom boost anchors are not implemented by this adapter and fail admission
clearly. Unsupported explicit revisions or inconsistent anchor tables fail before
any installation, RDRAM or rendering work. The supplied file has this feature off.

Tests compare the unmodified importer with the new importer on the exact supplied
archive, cover repeated/clamped RGBA sampling and area coverage, bound geometry,
retain pivots and anchor-unit modes, and compare semantic fingerprints for nine
installed older vehicles. Negative archives test unknown profiles, unit/revision
and hash mismatches, RGBA32-safe budget violations and unsupported effects. All
archive tests are data-only; no mod/game code, ROM, RSP or GPU is executed.
Linux sanitizer compilation includes the final prepared package translation unit.
The retained graphics handoff/rendering, memory and online regression chain is
also required. Static archive read-back and ELF dependency closure stay required.
