Skip to content

srmech codebase audit — abs() + numpy-math residue (rc32)

Done 2026-06-04 via three parallel sub-agent sweeps (amsc / qm+signal_processing+spectral / dsl+bus+cli+mcp+introspect+rbs_lm) at the user's direction, after three standing directives:

  1. "abs() is never fine" — magnitude/sign is an explicit Class-K sign-branch (x if x >= 0 else -x / np.where(x>=0, x, -x) / real-imag composition), never an ALU abs() / np.abs() / math.fabs() / sqrt(·²) stealth-abs.
  2. "never use numpy math when srmech can do it with a cascade" — when srmech ships the primitive (a C op or a Python cascade), use it; don't delegate the math to numpy. (Container/layout ops — np.zeros/array/asarray/reshape/ tolist/copy/astype/sum/mean — are NOT "math" and are fine.)
  3. "audit entire codebase for numpy math residue in this rcN."

Category A — abs() residue: FIXED codebase-wide in rc32 (value-preserving)

Every site below was replaced with a value-identical explicit Class-K form and is now guarded by a permanent AST ratchet test (tests/test_laplacian_numpy_free.py::test_no_abs_calls_anywhere_in_srmech — fails CI if any real abs()/np.abs()/math.fabs() CALL reappears anywhere in srmech/; docstring mentions don't trip it).

  • amsc/rational.py ×2 (abs() → sign-branch feeding math.gcd)
  • amsc/kepler.py (Newton-step convergence abs(delta) → sign-branch)
  • amsc/harmonics.py ×5 (sqrt(x*x) / sqrt(·²) stealth-abs → np.where / sign-branch; the docstring's false "no abs()" claim corrected)
  • amsc/laplacian.py (signed_laplacian np.abs(A_off) → np.where)
  • amsc/cascade/hypercomplex_dft.py (my rc31 ℍ-closure guard np.abs(v[4:]).sum() → np.any(v[4:] != 0.0) — a presence test needs no magnitude at all)
  • qm/bell.py ×3, qm/sm.py ×2, qm/pseudo_hermitian.py ×3 (np.abs(real) → np.where; abs(scalar) → sign-branch; abs(phi)**2 → phi.real**2 + phi.imag**2)
  • spectral/__init__.py (np.abs(complex) → np.hypot(real, imag))
  • signal_processing/: 21 sites across 13 files — complex np.abs(X)**2 → X.real**2 + X.imag**2; complex np.abs(z) → np.hypot(z.real, z.imag); scalar abs() → sign-branch (form_function_rotation.py, ica_jade.py, and the closed_form_ops/ DSP family: cross_spectral, mlse, multitaper, music, spectral_subtraction, spectrogram, wiener, fsk, ofdm, psk_qam + path_b_ops/wiener).

Verified clean (already abs-free): qm/octonion.py (octonion_norm routes through the Class-K cascade _magnitude, not abs — the documented claim holds), qm/triality.py, qm/so8.py, and all of dsl/, bus/, cli/, mcp/, introspect/, rbs_lm/, llm/, adapters/.

Category B — numpy math where srmech HAS a cascade: PARTIALLY done; rest STAGED to rc33

Done in rc32 (the §22-core Class-L spectral op): - amsc/laplacian.py jacobi_eigvals no-native fallback: np.linalg.eigvalsh → srmech's own pure-Python Jacobi cascade (_jacobi_eigvals_py). The Class-L eigenvalue math is now srmech's, never LAPACK. Native C stays primary. - The real Laplacian builds use srmech's pure-Python builder (not numpy elementwise) in the no-native path.

STAGED to rc33 (engine-routing — behavioural, needs per-site parity tests; NOT silently dropped): - ~11 np.linalg.eigh/eigvalsh on Hermitian/symmetric matrices (qm/gauge, qm/potentials, qm/single_particle ×4, qm/so8 Killing form, signal_processing esprit/heat_kernel/ica_jade/music) → route through laplacian.hermitian_eigendecompose (srmech's Class-L primitive). - ~15 np.cos/np.sin/np.arctan2/math.cos/math.sin (windows, DCT bases, CKM/Weinberg angles, rotation eigenvalues) → srmech.amsc.rational / asymptotic_calculus Class-N exact-rational trig.

Keep numpy (genuinely scientific-tier, no srmech equivalent — §22): np.linalg.{svd,qr,eig(non-Hermitian),lstsq,inv,solve,pinv,matrix_rank,norm}, np.einsum, np.kron, np.outer, np.trace, np.convolve/correlate, complex np.exp time-evolution phases. np.fft.* is architectural — both the Path-A and Path-B signal_processing FFT ops delegate to numpy (there is no hand-rolled srmech FFT); building a native srmech FFT cascade is its own future voxel, not a residue fix.

Why staged, not all-in-rc32

rc32 = §22 laplacian-numpy-free + the absolute abs() sweep (value-preserving, low-risk). The Category-B engine-routing swaps WHICH solver computes a result (numpy LAPACK → srmech primitive) — value-equal to round-off but a different code path that needs per-site parity verification, and §22 itself keeps the heavy qm linear algebra numpy-backed. Bundling 26 engine swaps into the same rc as the 16-file abs sweep would make one un-reviewable, regression-prone rc. rc33 does the routing as a focused, separately-verified voxel.

rc33 CLOSURE (2026-06-04) — Category B routed

Done in v0.7.0rc33. The premise that "full trig routing needs new infrastructure" was WRONG (user-corrected): the Class-N Taylor cascades (sin/cos/atan_series_truncate) were always globally convergent and pi_cascade_digits always shipped — the only gap was a float-returning, range-reduced wrapper composing them. That wrapper now exists: srmech.amsc.rational.{cos,sin,tan,atan,atan2} (π-cascade range reduction → Class-N anchor → Taylor → project to float; atan three-band √2∓1 reduction). Measured vs libm: cos/sin ≈6e-16, atan/atan2 ≈2e-16 (machine ε). +5 tool-schema entries (describe 210→215).

  • Trig sites routed — qm.sm CKM/Weinberg angles + signal_processing window/basis/rotation trig (cross_spectral, dct, multirate, multitaper, stft, ica_jade, form_function_rotation). The amsc-internal solver trig (kepler Newton, kuramoto coupling, hypercomplex_dft twiddle, laplacian magnetic-phase) was NOT in this audit's "~15" and stays as its own future consideration (those are inside iterative solvers / DFT twiddles where the substitution is a separate fidelity Q).
  • Hermitian eig routed — the 11 np.linalg.eigh/eigvalsh sites in qm.{potentials,gauge,single_particle,so8} + sp.{esprit,heat_kernel,ica_jade,music} → amsc.laplacian.hermitian_eigendecompose (native C complex-Hermitian Jacobi on the native path). Real-symmetric inputs .real the eigenvectors; complex-Hermitian keep complex128. Per-site parity tests (eigvals ≤1e-9 + reconstruction).
  • Still scientific-tier per §22 (no srmech cascade yet): svd / qr / non-Hermitian eig / lstsq / einsum / kron / complex-exp / FFT.