PROPOSED ONLY — actual cleanup, real QA and new generation have not run. Purpose: correct the tan color cast in the generated grating material of 203c6/c7 while retaining generated local luminance and all nonselected artwork. The existing quiet V2 raw remains untouched and HOLD2 remains in force. Selection: the complete authentic gray grating cohort, Source indices 11..30 inclusive in fixed Source geometry regions [21,3,31,31] and [2,0,31,12], respectively. All 179/192 cells are selected, including every gray shade in those structures; selection never reads candidate color or pixel error. The Source-only audit publishes sorted coordinates, exact indices, RGB colors, hashes and aggregate means. No original full-frame upscale is called HD. Single fixed transform: crop each 887-square panel from the verified RGB raw; disclose Pillow Lanczos3 resize to 128 square as the baseline. For selected 4x logical cells only, retain the actual generated per-pixel luminance Y, subtract one cohort-wide generated mean and add the authentic Source cohort aggregate gray RGB. This is equivalent to a single constant luminance shift plus neutral chroma anchor. A fixed absolute shift cap of 32, no clipping and <=0.5 luminance rounding error are mandatory. No per-cell Source RGB fill, gain/sigma profile, random grain, flat fallback, shape cleanup or texture synthesis. Nonselected resized-parent RGBA and every alpha byte remain exact. Equality to the old Rust-resampled raw bank is not claimed. Detail provenance: selected spatial luminance differences come only from this untouched generated raw. Source supplies semantic geometry and one aggregate color per frame. Require genuine same-material raw 3x3 local luminance witnesses before any correction. Witnesses/variance are technical evidence, not visual approval. Original-owned geometry and aggregate color receive no generated-detail credit; direct Original pixel-copy operation count is zero (coincidental numerical matches are not asserted absent). Input/publication: strict complete Source32 headers, ordinary opaque class, six-bit palette, exact RGB/RGBA PNG bounds, verified-byte decoders, fd-walk no symlinks and NONBLOCK regular final descriptor, bounded exact-length stable fstat read and SHA. Fresh atomic no-replace output, reauthenticate inputs/tool/recipe before publication. Existing output is never overwritten. This private helper is not installed publicly. Synthetic validation: 21 tests pass including independent baseline/alpha equality, two selected generated luminance differences retained, repeatability, semantic index/region boundaries, clipping and nonneutral rejection, malformed inputs, palette/header limits, symlink/FIFO/oversized/short/changed reads, fresh receipt and no-replace. Tests render synthetic inputs only. First 14-test draft passed before FIFO/short-read fixes; the final 21-test log is authoritative for this helper. Next gate: different code peer reviews this proposal and fixed pins; Root physically reviews authentic Originals and untouched raw. Only after explicit Root release may a single real derive run create private two-frame outputs, then one current G2 QA and complete native-scale material review. No tuning iterations or automatic qualification. Any remaining material mismatch stays HOLD. No canonical/source/registration/Git mutation is authorized here.