Original Peak SM86 GPU modules

This local TEST embeds unmodified Peak main/preparation GPU modules.
The host adapter, pool connection and CPU verification are Mouillay code.
The GPU modules are Peak components, not Mouillay-written kernels.

prefix_sm86_peak_original.cubin SHA256 fb170193159ada6121cf6b490096d806a6d8efcd5e0b65cb779a909189881d3d
tables_sm86_peak_original.cubin SHA256 f1e4fdb76b209765b2f99ae7b81e43ca48d1c1e891919c54b8afec20297b0407

TEST27 / SM89: unmodified original Peak SM89 modules, dedicated 0x240000-byte expanded-low/high preparation. Host adapter and full CPU512 candidate verification remain Mouillay.
prefix_sm89_peak_original.cubin SHA256 3536fef9b2940a705dcdf95f539c6e41f878f9e26097af7af8f016e643108e03
tables_sm89_peak_original.cubin SHA256 45ee5d386a7092a5a1c854d85ffa1cca85bf351998d760e96ccaac3ae34ff604

TEST30 / SM89: unmodified original PeakMiner2.20.5 QPTAB16 module.
quantus_sm89_peak205_original.cubin SHA256 354df94bd2606a4486c01c164598834e36d157856460c3537b8daf040b9b4b19
Main _qv3_19ca8d515 (40 registers), preparation _qv3_766f772b94 (16 registers).
This binary includes both functions and mutable job constants. It is a Peak
component, not a Mouillay-written kernel. Independent CPU preparation,
nonce reconstruction, exact thread-count launch, target conversion, result
bounds and full512-bit candidate verification are Mouillay code.
The separate low/high16 basis and 4,194,384-byte preparation differ from TEST28.
Source recovery and attribution evidence: research/session30-gap/peak-agent.md,
decode_peak205.py and peak205/decode-manifest.json. The original CUDA ELF bytes
are unchanged. This attribution does not change any third-party license terms.
