2026-09-02 - (/loop) Packed typed-buffer formats begin: UINT single-word load, GPU-verified

Pivoted (at the user's push) from the corpus-wall crunch to the ahead-of-guest GPU shader work. First finding: the task the roadmap pointed at - G10 resource model, MTBUF translation - is already done (V# decoder D204, MUBUF/MTBUF-plain translation worklog 101, D203/D205), verified in model.rs. The roadmap doc was stale; corrected its G10 row.

The one capture-free piece that genuinely remained is narrow-component format conversion - the case typed_buffer_memory refuses as "needing component conversion". Started it (D466), lowest risk first:

• Added packed_buffer_memory in orbistoun-translate/src/model.rs. typed_buffer_memory now routes non-plain formats there instead of refusing; plain-word formats still go straight through buffer_access unchanged.
• Implemented UINT, single word: read the one packed word, and for each component shift it down and mask to its width - exact bit fields, no rounding, so it is correct by construction. Signed, normalised, float and sRGB kinds, multi-word widths, and stores are still refused, each with a detail that keeps the word "conversion" so the existing refusal test passes.
• Test: a_packed_uint_load_extracts_each_component_from_one_word stores a raw word (0x04030201) via the untyped path, loads it as BUF_FMT_8_8_8_8_UINT, and asserts each byte lands in its own register, low byte first. It executed on a real Vulkan device (no !! SKIPPED printed), so the extraction is validated on hardware, not just spirv-val-accepted.

clippy/fmt clean; all 104 execute tests plus the rest of the translate suite pass; the three existing typed-format refusal tests are unchanged and green.

Next increments, each device-validated before the next: SINT (needs a SHIFT_RIGHT_ARITHMETIC op for sign extension - check it exists in orbistoun-spirv::op, add if not), then UNORM/SNORM (int->float convert, divide by the max, clamp for SNORM), then FLOAT16 (half unpack), then SRGB; then packed stores (pack + write), then multi-word widths. The execute.rs mtbuf/mubuf helpers and this test are the template.