By: ValentinHirschi <valentin.hirschi@gmail.com>
Add binary64 bubble and tadpole tensor kernels NativeTensorEvaluator<f64> (B11, B111, An) now runs a binary64 specialization of the OneLOop bub0-bub111/bnlog/solabc/tadpn port (src/native/tensor/binary64.rs). DoubleFloat and Float keep the generic code, which f64 can still select with the hidden set_binary64(false) for validation. Same sectors, exact zero tests, mass ordering, lips, root pair and series radius; what changes: - real kinematics stay in real arithmetic: one ln_1p per real root, the lip as +-pi(1-x^(n+1)) on a real direct formula, and every real root and -1/x one quotient of the inputs (complex masses keep the generic roots); - complex-conjugate roots of real masses below threshold take one bnlog: bnlog(x1) + bnlog(x2) = 2 Re bnlog(x1), exactly real; - the bnlog series is evaluated by Estrin's scheme on 8 (|y|_1 <= 0.01), 16 (<= 0.2) or all (<= 24, zero padded) coefficients, each meeting the generic 2^-57 truncation bound (checked at construction), instead of a Horner chain of up to 21 complex multiply-adds per rank. Each rank depends only on the row, so B111's first nine outputs are bitwise B11's. Timing (in-process, interleaved with OneLOop 3.6 gfortran -O3 through Ninja's _cm wrappers, CPU 52, medians of 9 blocks >= 20/25 ms, two rounds, before = d10eaa0) on Ninja's 27 recorded master streams: - every distinct bubble call (285 rows): worst 1.31x -> 0.57x OneLOop (syn-n2-r1/r2, syn-n4-r4, syn-n6-r6: one massive, one massless line below threshold, 1.05-1.31x -> 0.39-0.57x); median 0.23x; - bubble group per case 0.15-1.22x -> 0.08-0.52x; bubbles + tadpoles 0.19-1.05x -> 0.10-0.55x; syn-n2 totals 0.97-1.08x -> 0.32-0.55x; - An vs olo_a0 0.46-0.62x -> 0.42-0.58x; 1,200 random rows over every mass sector: worst 1.28x -> 0.90x (a massless An, ~12 ns). D0/C0/A0 code is unchanged. Accuracy: tests/native_tensor.rs passes unchanged (worst binary64 error against 320 bits 2.1e-14 -> 1.2e-14, Feynman-parameter rows 2.1e-14 -> 1.8e-14). New tests/native_tensor_binary64.rs: 8,000 random rows over every sector against the generic binary64 code and 320 bits (error quantiles 50/90/99/99.9/100% 0/1.6e-16/1.4e-15/9.8e-15/1.2e-12 vs generic 0/1.6e-16/1.5e-15/9.4e-15/1.1e-12; never 16x worse above 1e-14), An bitwise equal to the generic code, every bubble/tadpole call of the 27 streams (tests/data/ninja_tensor_rows.txt, generated by tests/generate_ninja_tensor_rows.py) against OneLOop (<= 9.4e-14) and 320 bits (<= 3.4e-14, as the generic code), and 2,090 regression rows against the symbolic b0/b1/b11/b111/an expressions at 256 bits (worst 4.4e-14; 3 random outputs at the generic code's own 5e-14-2.5e-13). The expressions of this commit put opposite-sign real masses at p^2 = 0 on the wrong lip and lose finiteness for m1^2 = 0 where a root rounds to 1-ulp; the test reports both, and both are empty with claude/expression-fixes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>