[vincentchoqueuse]

Five independent reviews swept the whole library — core signal path, compensators/metrics/utils, optical, MIMO+FEC, and the chain machinery with OFDM — hunting for wrong axes, dtype loss, RNG misuse, numerical blow-ups and validation gaps. Every suspicion had to be verified by running a minimal reproduction before being counted: 37 confirmed defects, 37 fixed, each locked by a test in tests/test_audit_regressions.py (36 tests).

The numerical cores came back clean against independent references: closed-form SER curves cross-checked by Monte-Carlo and numeric fading integration (~1e-10), Viterbi against exhaustive ML (600 trials, metric-exact), ML against the sphere decoder bit-for-bit, LDPC min-sum against brute force, CD/DBP round trips to 1e-14, EDFA/Raman budgets against their closed forms, the (133,171) distance spectrum against the published tables. No block mutates its input.

Real mis-computations

• Sequential.seed() never reached blocks inside a nested Sequential or a wrapper chain (the OFDM chains) — a seeded composite chain was not reproducible; it now recurses, with the flat-chain child seeds unchanged.
• The AMP detector evaluated its posterior-variance term G at variance tau^2 instead of sigma^2 (1 + tau^2) (Jeon et al., Algorithm 2): at high SNR the state went NaN and was silently hard-projected to symbol 0. soft_projector also gains a log-sum-exp guard, which fixes the same divergence in OAMP.
• ErrorCounter raveled before truncating, counting phantom errors on every batch row after the first under a filter delay.
• CarrierAllocator silently zeroed complex pilots on real data; OFDMReceiver read the period of a 2-D scattered mask as N_fft.
• compute_acpr counted a bin sitting exactly on a band edge in both bands.
• SRRCFilter(rho=0) — the documented sinc case — crashed; a logarithmic split-step schedule at alpha_dB=0 was all-NaN (now the uniform spacing it is the limit of); the FIR CD compensator applied a sqrt(1/(4*pi*K)) gain when its design collapsed to one tap (now the identity it tends to); the optical 90° hybrid and the MZM crashed on any (2, N) input.
• Sequential.summary ran the chain without feeding wiring or recording taps.

State that lied

FrameField froze the caller's array in place; Deframer recorded views of the input buffer; tap() served signals from a previous run or a previous taps= configuration; DataExtractor.forward overwrote its declared selector; BWFilter silently turned real signals complex.

Contract mismatches

A nested chain exported to JSON but failed to import; int-keyed callbacks were documented but never fired; ViterbiDecoder broke exactly at the empty message; the distribution matcher pair died on a zero-length stream; print_data died on degenerate tables.

Validation gaps closed with clear errors

Negative AWGN variance, cross-QAM orders in the square-QAM closed form, Serial2Parallel method, Complex2Real part, Upsampler phase, SampleRemover windows, CyclicPrefixer(N_cp > N), 1-D Parallel2Serial, short-signal SRRCFilter(method='fft'), numpy-integer sizes in the generators, empty compute_ser.

DataAidedMixin becomes a plain empty-slots class so the slotted compensators stay expressible under pyright's new disjoint-bases check.

Verification

• 1011 tests + 1293 subtests green (including the new regression file), examples smoke tests green
• ruff check . clean, pyright 0 errors, sphinx -W clean

🤖 Generated with Claude Code

https://claude.ai/code/session_01Pn2F8X4HZhS8fEJcL5enzw

---

Generated by Claude Code