Harden GPU pipeline selection and recall validation

[JustAResearcher]

Closes #10
Closes #11
Closes #14

Summary

• remove the duplicated SOLVER_PRELAUNCH_NEXT mining loops and keep the smaller context-owned edge-buffer path in both the standalone solver and pool miner
• derive automatic pipeline depth from free VRAM after the first context, with an automatic cap of 4 only on Windows WDDM and a cap of 5 on Linux/Windows TCC
• add exact proof-level GPU recall regression tooling, a conservative reference build profile, build/runtime architecture binding, and hard gates for hidden edge loss
• compile both release and reference solver profiles in CI and run the platform selector plus dependency-free recall verifier fixtures on Windows and Linux

Validation

• Windows CPU/protocol/reliability/pipeline/consensus suites: pass
• Linux CPU/protocol/reliability/pipeline/consensus suites: pass
• recall verifier fixtures: 30/30 pass on Windows and Linux
• Windows CUDA 13.2: release solver, pool miner, and reference solver compile for sm_86, sm_89, and sm_120
• Linux CUDA 12.8 / Ubuntu 18 compatibility build: release solver, pool miner, and reference solver compile for all three architectures; every binary requires at most GLIBC_2.14
• RTX 5090 WDDM:
  • automatic selection chooses pipeline 4; explicit pipeline 5 is honored
  • tightened metadata-contract 128-graph release/reference repeatability and pipeline-4 exact proof parity: pass, 4 identical proofs, embedded/runtime sm_120, compiled trim defaults 48/50, zero verification failures
  • 4,200-graph exact recall sequence: pass, 110 identical proofs across both release runs, both reference runs, and pipeline 4; zero verification failures
• RTX 4070 Ti SUPER Linux (sm_89, idle Vast host with no active container):
  • automatic selection chooses pipeline 2
  • 420-graph deep sequence: 12 identical proofs across release repeats, reference repeats, and pipeline 2; zero verification failures
  • tightened metadata-contract 128-graph sequence: 4 identical proofs; embedded build target, runtime architecture, and compiled trim defaults all match sm_89
  • validation files copied off-host; temporary remote files removed; Vast service left active and GPU returned idle
• current Windows pool miner, pipeline 4, loopback job: 197 graphs, 4 submitted/accepted shares, zero verification failures, clean exit
• pre-change path A/B on the RTX 5090, three fixed 128-graph runs each:
  • shipped double-host-buffer path median: 25.701 graphs/s
  • retained context-owned-buffer path median: 26.187 graphs/s

Boundaries

• sm_86 is compile-tested only because no verified Ampere test device is available.
• Windows TCC selection is unit-tested and CUDA-compiled but not runtime-tested because the available Windows GPU uses WDDM.
• Reference binaries are written under bin/validation; package_release.ps1 copies only the six exact release binary names, so reference binaries are not packaged.