0e4005d7

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

deps: pin oneloopmaster main 149ddff

Move every oneloop pin (workspace, the ill-conditioned and masters_native
manifests, all four lockfiles via `cargo update -p oneloop`, manifest.json
and the README table) from 3023fa6 to oneloopmaster main
149ddffb496b8837f8e176dbe24d133c2490d88d (massless D0 collinear loss and
near-Landau discriminants of massless/one-/two-mass boxes).

149ddff changes no call NinjaRS evaluates in the benchmark. With perf
CLIs built on 3023fa6 (b215cba) and 149ddff, `reduce` gives identical JSON
(timing fields removed) on examples/graphs/*.json, the physical pools and
the CLI test graphs in f64, double-float, float 256, --engine fast and
--r2 fdh (135 runs, the 20 expected failures with identical errors), and
case-bench (27 cases) and pipeline-bench (7 diagrams) give identical
reference checks and lane values. The masters_native harness returns the
same bits for every native master, relabellings included, on all 34
recorded call streams (no box there has fewer than three massive lines),
and the 700/1000-bit, double-float and fast pentagon outputs match the
committed ones. The benchmark data measured with 3023fa6 therefore stay;
the report's environment paragraph says so.

Report fixes: the arbitrary-precision timing compares like for like
(interleaved on CPU 245: 700 bits 7.8-8.6 s on 264a6ce against
11.1-11.6 s on 149ddff, 1000 bits 9.1-9.9 s against 14.4-14.8 s, about
1.4x and 1.5x, from oneloopmaster 3cbe08d, reported upstream; recorded in
timing.txt with the run.sh wall time on 149ddff); the 9c5490b Gram-limit
D0 regression is noted as reported upstream; tensor bubbles are
3.8-12x faster than olo_b11; the 149ddff fix joins the oneloopmaster
defect list. The ill-conditioned README now says that the n6r6 rerun's six
master_check records make no quadruple-precision comparison, and no longer
names the branch.

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