Solver limitation register (40-case audit)

An adversarial audit built 40 commercial-plant test cases against every steady solver mode, found 35 silently returning a plausible-but-wrong number (6 of them critical), and used that register as the drumbeat for a multi-session coverage-extension epic that closed every one — the register now stands at 40/40 correct-or-loud.

• You want the historical evidence for why the honesty register exists — this audit is what proved the platform needed one.
• You're deciding whether a new physics area needs its own honesty detector — check whether a similar commercial case is already in this register before assuming it's covered.
• You're explaining to a stakeholder why correct-or-loud matters — this document is the concrete before/after.

Method: build real commercial-plant scenarios (not synthetic edge cases) across every solver mode, run them, and classify each result as correct, silently-wrong, or already-loud — a fundamentally different check than unit-test coverage because it hunts for confidently-wrong answers, not code paths that error.