MOUILLAY GPU MINER 1.0.3-rc2 — SM120 EXPERIMENTAL PERFORMANCE UPDATE

RTX 5060/5070/5080/5090 (CUDA compute 12.0) only for the new compute path.
RTX 3000 and RTX 4000 production compute/preparation binaries are unchanged.

START_MINER.bat: normal mining; edit your pool and wallet.worker.
The default SM120 path now uses split2 with a 16x16 grid, packed exact U8
operands and XOR-linear partial transcripts. B packing is cached per job.
SM120 compile report: 128 registers, no stack/spill, 16 KiB dynamic shared.
The old kernel used 255 registers, 16-byte stack and register spill loads/stores.
Power reporting, dev fee and PEARL+NOCK handling are unchanged.

COMPARE_SM120.bat: three 30-second real mining runs, SAME 16x16 grid,
legacy -> split2 -> split4. Edit its pool and wallet too. Each run performs
the selected-device CPU/GPU startup self-test. Send logs/compare-*.jsonl.
To force a variant manually before START_MINER.bat:
  set PEARL_BLACKWELL_VARIANT=split2
  set PEARL_BLACKWELL_VARIANT=split4
  set PEARL_BLACKWELL_VARIANT=legacy
The split4 kernel uses 96 registers with zero stack/spill.

What was validated here: actual RTX 3070, SAME SOURCE compiled as SM86,
21,504 complete transcript comparisons, 10,760 matching candidate digests,
32 Rust-verified proofs across both new variants, plus 16 proofs against the
integrated new default. Nonce boundaries, target boundaries and guards pass.
Controlled 16x16 medians on that regression build (21 scans per variant):
  old Blackwell prototype 14.50 TH/s; split2 41.67; split4 38.17.
  Desktop miner cooperatively paused for ~3 seconds and resumed afterwards.
  Earlier contended timings were superseded and are not used for this claim.
These are NOT RTX 5070 results and NOT a comparison with the fast RTX 3070
production kernel. No physical SM120 GPU was available on the development PC.
There is no verified RTX 5070 hashrate for this new build yet.

TEST_PEAK_SM120.bat: separate original PeakMiner 2.20.2 native SM120 diagnostic.
Its PEARL GPU module, preparation and native records are original/unmodified,
not authored by Mouillay. It is embedded only for this explicitly requested
private test. No competitor miner program or pool connection is started.
It requires physical SM120 and a sufficiently recent NVIDIA driver.
The complete native work/transcript/consensus contract is NOT yet qualified:
this diagnostic records native execution times and raw records, NOT TH/s,
and DOES NOT submit shares or replace our active mining backend.
Send logs/peak-sm120-console.txt and logs/peak-sm120-diagnostic.json.
Do not confuse this diagnostic with a validated Peak-based mining integration.

This is a separate local preview. Existing public releases, Desktop miners,
Linux and Docker packages have not been overwritten or republished.
