[vincentchoqueuse]

Twenty commits. The thread running through them: every claim was
measured before it was acted on, and three of the measurements
contradicted me.

Three defects

Nothing ran the nine SLOW examples. The smoke test skips them —
that is what SLOW is for — but no other job ran them either. A
regression lived in CD_compensation_part1.py for two commits:
TypeError: 'Constellation' object is not subscriptable on its first
pass, every check green. They now run weekly and on demand, and the
runner imports SLOW from the test module rather than repeating it, so
the list has one home.

A serial monte_carlo without seed= was silently unreproducible.
The parallel path refuses outright; the serial one said nothing. It
warns now — a first look at a curve does not need to be reproducible, but
a curve nobody can reproduce is not a result.

N_s did not mean what it said. channel(N_s * oversampling_sim)
handed a symbol count to the generator and the Upsampler applied the
oversampling again, so both optical examples simulated six times the
data their variable names promised — and the page quoted "8192 symbols
per point" for what is 49152.

The fibre is seven times cheaper, and says the same thing

apply_chromatic_dispersion rebuilt its transfer function on every
half-step: an fftfreq, a w**2 and a length-N complex exponential,
a thousand times per propagation, for something that depends only on
(n, dz, beta2, fs, alpha_dB, direction). Split into
linear_step_transfer (what the step is) and apply_frequency_response
(what it does), with step_transfers caching one row per distinct step
length — one for a linear schedule whatever StPS is, StPS for a
logarithmic one, so the loop indexes a table and never branches on the
schedule.

With StPS measured down from 200 to 50 — 0.005 dB at the sweep's most
nonlinear power — the launch-power sweep went from ~1400 s to 144 s
and the one-shot page from 73 s to 13 s. Every measured decibel on
both pages is unchanged; nine configurations agree with the old code to
3e-15.

The AWGN tutorial derives its reference curve

The Monte-Carlo page asked the constellation for a closed form and drew
it, which teaches that an object returns the right numbers rather than
why they are right. It now states the expression — a square QAM as two
independent PAM constellations, the argument of Q, Proakis and Salehi
section 4.3 — evaluates it in plain NumPy, and only then shows
constellation.metrics returning the same array to 3.3e-16.

It also no longer stops at one constellation. A closing section sweeps
4-, 16-, 64- and 256-QAM at equal energy per bit and reads off what
the density costs: 6.8, 10.5, 14.8 and 19.4 dB of Eb/N0 for a BER of
1e-3. Both floors of that figure are named rather than glossed — the
Gray-mapping approximation runs a factor 2.2 under the measurement for
256-QAM at 0 dB before settling onto it below 1e-2, and 100 000 symbols
per point cannot resolve a 4-QAM BER under 5e-6.

Two more defects surfaced there: the page still named sweep in two
cross-references after the rename, and the figure it displayed was
written by no script — monte_carlo_awgn.py ended on plt.show() and
saved nothing, so the committed PNG was whatever a human last produced
by hand.

Where I was wrong, three times

A batched FFT is not faster than B separate ones — 40.2 ms against
39.1. Nor was the gain the per-call overhead I then blamed: building the
per-span amplifier is 20 µs and a block dispatch 0.25 µs, together 0.03 %
of a pass. It was the transfer function all along. The batch axis is
still worth having and is now documented, tested and discoverable, but
for what it is: 1.76x by amortizing that exponential, and the
precondition for a GPU backend.

Smaller things

FiberLink.budget() puts the amplifier-noise formula next to the block
that generates the noise, where three tutorials had written it by hand
with three spellings of the polarization factor. ErrorCounter makes an
error rate a ratio of totals rather than a mean of per-trial rates.
sweep becomes monte_carlo, which says that its points are random
draws. Two "helpers" in validation/ turned out to be the library's own
functions under other names and were deleted. wdm.py was calling
np.fft instead of the backend, so that block could never have run on
cupy.

validation/optical_dbp_hager.py restores the reproduction of Häger and
Pfister's curves on their own setup — Gaussian stimulus, logarithmic
steps for the low-step back-propagators — which successive rewrites had
turned into a teaching script.

CONTRIBUTING.md now says which NumPy the strict-typing ratchet is read
against: the 1.26 stubs require ndarray to carry type arguments and
the 2.x stubs do not, so linting against the pinned oldest NumPy buries
a real error under several hundred that are not.

Also removed: pyright_all_tmp.json, a scratch copy of the ratchet list
committed by accident.

Verification

ruff check ., 860 tests + 1145 subtests, doctests, and
sphinx-build -W pass. 44 new tests.