[Alok-Ranjan23]

Summary

• Preserve the existing fast path for complete quantized REPACK transfers.
• For partial transfers, untile into canonical bytes, update or copy only the requested range, and retile after a SET.
• Extend the Hexagon buffer regression across Q4_0, Q4_1, Q8_0, IQ4_NL, and MXFP4, including nonzero-offset writes and a sentinel-protected short read.

Error corrected

On SM8750/v79, an opt-in non-HOSTBUF Q8_0 128×33 buffer accepted a 136-byte transfer but its repacker processed the entire 4,488-byte tensor. A short GET overwrote the sentinel tail of the caller's full-size diagnostic buffer; a short SET changed rows outside the requested range. The safe baseline diagnostic exited 1. The corrected package preserves both tails and exits 0. The default HOSTBUF=1 path remains unchanged.

Validation

• Exact speech@de1d0f5c Android host/v79 DSP package: 15/15 device hashes match before and after.
• Expanded five-type buffer regression exits 0 with HOSTBUF=0 and HOSTBUF=1; baseline short-transfer diagnostic exits 1, corrected diagnostic exits 0.
• Fixed-package HTP matmul CPU comparison passes 5/5 and records five completed HTP operator profiles. Baseline selected v79 suite recorded 112 completed operator profiles.
• Focused host build passes; shared-buffer, opcache, runtime-info tests pass 3/3, and the HTP-only host case skips with code 77. git diff --check passes; independent source review found no blocking issue.

Partial SET now uses a temporary canonical tensor and full untile/retile, so frequent partial updates can cost memory and time; complete transfers retain the original path. These checks do not establish whole-model parity or a performance improvement.

Audit subtask: https://app.asana.com/1/45238840754660/project/1214153063536860/task/1220036025374365