[lsimon125]

Demonstrated Problem

The legacy one-loop source path drops finite widths from tree propagators and loop bridges. In e+ e- > u u~, this shifts the Born result and every virtual coefficient by the same relative amount (about 1.28e-7 at one of the tested points). Setting only the reference Z width to zero removes the discrepancy.

Minimal Fix

• Add the source fixed-width term +i*m*width to the shared tree/bridge denominator helper. Born, local counterterms and loop bridges use the same helper.
• Retain the corresponding declared width parameters in the source scalar binding, including live parameter updates.
• Add exact zero-width, finite-width and renamed-width regression cases.

This is a fixed-width tree/bridge prescription, not a complex-mass renormalization scheme. Integral-family masses and the v2 engine are unchanged. No process-specific colour or Born-factor shortcut is used.

Validation

• Three exact scalar regression cases pass. They need no native graph-worker fixture and were run with the parent collection-wide graph licence gate excluded:

PYTHONPATH=src python -m pytest --confcutdir=tests/unit \
  tests/unit/test_loop_graph_basis.py -q \
  -k tree_bridge_denominator_retains_symbolic_fixed_width

• Independent MadNkLO/MadLoop comparison: 56 Born/finite/single-pole/double-pole checks pass across two kinematic points, two alpha_s values, three scales, and additional zero/enlarged Z-width updates on the loaded runtimes. Worst relative difference: 4.49e-13. Acceptance tolerance remains 1e-7.
• Generated artifacts are evaluated through the dynamically linked C SDK; no numerical normalization adjustment was introduced.
• git diff --check passes.

The numerical fixture was run in the isolated MadNkLO integration based on loop at 36f43250, with the package-scope fix from #30. The three files touched here are unchanged between that base and this PR base. Generated artifacts/reports remain local; this PR contains only the source fix and small regression.