5f35b071

By: ValentinHirschi <valentin.hirschi@gmail.com>

docs: correct the benchmark report after a claim-by-claim review

A review of every factual claim of the report against the recorded data
and the code found 24 confirmed issues; all are fixed:

- references re-timed on the NinjaRS core (CPU 250) with the same
  executables; they agree with the CPU 29 runs within ~1 % except OneLOop
  on the n = 5 cases (validation/benchmark/cpu-crosscheck.md);
- MadLoop's Ninja-interface glue is counted with its numerator, where
  NinjaRS counts its own propagator and S-matrix preparation, instead of
  inflating the reduction reference;
- Figure 1 labels use measured totals; Table 2's difference column uses
  the identical-input comparison;
- stale or imprecise numbers and attributions corrected (master
  agreement, tensor-numerator speed-up, rank penalty, mu^2 of gg -> H,
  component sums, run-to-run spread, dilogarithm timing, Symbolica vs
  expression fixes);
- evidence for the ill-conditioned synthetic lanes and the
  arbitrary-precision timing added under validation/benchmark.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>