Audit and improve safe library FIX rate

[rsasaki0109]

What changed

• advance gnssplusplus-library to the independent satellite-PAR surplus validator
• add canonical FIX/correct-FIX/false-FIX auditing on one epoch population
• add surplus monitor/active A/B analysis with input hashes
• add leave-one-city/time-block-out nested CV
• record full Tokyo/Nagoya, raw-fault, latency, oracle, and Pareto evidence
• add six focused analyzer tests

Result

• Tokyo: 50.37% -> 54.36%, +476 FIX; 3/6484 false (0.046%), zero >1 m; p95 62.25 ms
• Nagoya: 66.05% -> 69.38%, +253 FIX; zero false; p95 46.72 ms
• nested CV: 672/672 correct held-out candidates
• 8 raw RINEX fault replays: zero false FIX globally and in fault windows; reacquisition p95 <=3.02 s

The 70%/80% stretch targets are not reached. The included audit shows that unguarded oracle sources have 6.17%-19.17% wrong shares, far above the 0.1% budget; the adopted configuration is the tested safe Pareto point.

Validation

• Ruff passed
• pytest: 6 passed
• py_compile passed
• gnssplusplus MSVC Release build passed
• submodule PR: Add independent satellite PAR surplus validation gnssplusplus-library#401