[lsimon125]

Summary

Two fixes to NativeRuntime::evaluate_with_precision (feature symbolica-runtime) for compiled artifacts.

1. Contracted-color topology replay in double-double and arbitrary precision. Full-colour (and NLC) pure-gluon
   processes always use the contracted-color topology replay, and run_double/run_float rejected it ("compiled/eager
   contracted-color topology replay currently exposes f64"). run_generic_with_color_topology_replay now evaluates the
   materialized sectors at every replay relabelling in the generic precision, routes their summed groups to the physical
   groups (gather_color_topology_replay_generic, mirroring gather_color_topology_replay_samples) and contracts them
   (reduce_color_topology_replay_generic). The group contraction of evaluate_raw_sums_generic is factored into
   contract_groups_generic and shared; run_generic_materialized and the replay share the extracted stage evaluation
   evaluate_generic_stages.

2. Artifact constants at the working precision. Above 32 digits the results had binary64 accuracy, padded with
   digits, for every compiled artifact (with or without replay). The generic path converted binary64 constants
   (normalization, colour weights, model parameters, masses, source-state factors, replay weights, process-group
   factor) with T::from(f64), i.e. a 53-bit numerica Float, and a Float product is rounded to the smaller
   precision of its factors. RusticolHighPrecisionNumber::from_f64_constant now converts them at the working
   precision (Float::with_val(binary_precision, value); unchanged for DoubleFloat); the generic source fill uses
   masses converted the same way via CarrierMasses. Unit crossing signs and phases are unchanged (numerica
   multiplies by ±1 without rounding).

The constants themselves remain the artifact's binary64 values (as before); this change only stops them from capping
the precision of everything they multiply.

Tests

rust/crates/rusticol-core/tests/compiled_color_topology_replay_precision.rs (runs when
RUSTICOL_COMPILED_COLOR_REPLAY_ARTIFACT points to
pyamplicol generate 'g g > g g' DIR --color-accuracy full --execution-mode compiled): exact massless integer points;
ratios of totals between points cancel the couplings and must equal the exact tree-level ratios of
9/2 (3 − tu/s² − su/t² − st/u²) to ≤ 1e-27 at 32 digits and ≤ 1e-58 at 64 digits; totals agree with binary64 to
≤ 1e-12.

Ratios to the exact analytic values:

	before	after
g g > g g, 32 digits	error: replay exposes f64	4e-33, 7e-33
g g > g g, 64 digits	error: replay exposes f64	within 1e-58 (test passes)
u d > u d (no replay), 40 / 64 / 100 digits	1.5e-16 each	5.1e-40 / 2.1e-64 / 5.7e-101
u d > u d, u u~ > g g, 32 digits	4.8e-32, 8.7e-32	unchanged

Run on loop at bae3b9e plus these two commits (candidate build, artifact from that generator): the new test
passes at 32 and 64 digits; compiled_binary64_with_symbolica passes; the 152 rusticol-core unit tests matching
generic/precision/double/float/exact/replay pass (with PYAMPLICOL_TEST_PYTHON set for the Python-producer
round trip). The ratio values in the table were measured with the same two commits on 79df367 and the
u d > u d / u u~ > g g artifacts of that generator.

cargo check -p rusticol-core with default features, symbolica-runtime and loop-v2: no new warnings.

🤖 Generated with Claude Code