decoder-core: replace MWPM greedy fallback with scalable matching

[DennisWayo]

Context

The current in-repo BlossomMWPM backend uses an exact bitmask dynamic-programming solver for small graphs and falls back to a greedy matcher beyond the configured node cap. That is useful as a lightweight implementation, but it makes larger-distance threshold comparisons and neural-guided MWPM results hard to interpret.

This should be addressed before making stronger claims about neural or tensor-network decoder performance.

Proposal

Replace the default greedy fallback path with a scalable minimum-weight perfect matching backend, keeping the current exact solver as a small-instance regression oracle.

Possible implementation routes:

• integrate a proven Blossom V / Kolmogorov-style backend behind a small adapter,
• add an optional external matching dependency with a deterministic fallback,
• keep the current DP path for small test fixtures and cross-checks.

Acceptance Criteria

• Exact small-instance outputs match the current DP solver for representative d=3/d=5 cases.
• Larger surface-code cases do not use greedy fallback by default.
• Decoder behavior is deterministic under fixed seeds.
• Benchmarks cover at least d=3, d=5, d=7, and one larger distance.
• CI includes focused regression tests for syndrome reproduction and matching parity.

Notes

This is foundational for mwpm, neural_mwpm, and any future tensor-network or neural comparison work.