[ruslan-splynx]

Closes #515

Matching priors above 0.5 previously reached Rust's graph builder as negative-weight edges even though Sparse Blossom cannot handle them and the CUDA-Q adapter rejected them. Probability 1 and non-finite mechanisms could also disappear silently.

All matching constructors now reject non-finite priors and values outside [0, 0.5] with InvalidMatchingPrior { column, prob }. Validation precedes edge dropping and merging, so an invalid source prior cannot hide in an empty column or a parallel-edge combination. Native Python and CUDA-Q share the Rust range validation; inputs are never clamped. Zero and 0.5 remain supported; DEM construction and BP-family priors retain [0, 1]. Split constructors report split-model column indices.

Boundary tests also exposed an infinite log-odds weight for the smallest positive f64. A log-difference fallback handles overflowing ratios while preserving the original expression and tested bit-identical weights for ordinary priors. Decoder algorithms, quantization and tie-breaking are unchanged.

Validation:

• 65 targeted Rust tests passed on each of macOS and Linux, including all seven new prior regressions in both debug and release, DEM/matching/splitting suites and doctests.
• 15 native Python tests passed locally; 33 passed on Linux with the actual CUDA-Q QEC SDK, including 18 adapter/registration tests, with no skips. The isolated native overlay's ten source hashes match the committed code.
• 12,000 real Stim syndrome/decoder cases at distances 3 and 5: zero differences against the existing matrix-roundtrip reference, including syndrome and observable projections. This is a representation/reference comparison, not an independent optimal-matching oracle.
• Workspace/all-target Clippy, Python-feature Clippy and rustfmt passed locally; QEC/Python all-target Clippy passed on Linux. Full workspace/platform tests are left to PR CI.
• Manual self-review found no blocking issues. This correctness change makes no throughput claim.

Policy, reproduction commands and recorded evidence. Genuine MWPM retrace ties and infeasible syndromes remain outside this change.