perf: seed symmetry colour refinement by survivors on wide-sparse domains

[rasros]

Advances #938 (the liner-sf half).

What changed

• Seed the symmetry pass's Weisfeiler–Leman colour refinement (domainSeed) from a variable's survivors rather than its holes when the domain is sparser than dense.

Why

domainSeed fingerprinted each integer variable by enumerating its domain holes, building an O(span) LongArray key. On liner-sf-repositioning — domains spanning up to 20M for a handful of live values — that key is tens of millions of words, and every RefineKey hash/compare across the refinement rounds then costs O(span). BREAK_SYMMETRIES alone took 741ms of liner-sf's 800ms presolve (and it finds no symmetry to break). Fingerprinting by whichever of holes or survivors is smaller drops liner-sf presolve to ~175ms — the liner-sf goal in #938, though the bottleneck was the symmetry colour seed rather than Element GAC as the issue hypothesized. 2DBinPacking is untouched, so #938 stays open.

The two seed forms carry a distinct marker word, and within a form equal domains yield equal keys and distinct domains distinct keys, so the colour partition — and every generator found — is identical.

Testing

• Byte-identical presolve oracle (dry-run-presolve, budget=0) across all 130 mzn-bench instances.
• Bench-parity solve on the budget-truncated instances: ma-path-finding and oocsp prove the identical optimum with identical assignments; no wrong-optimum.