Skip to content

Release readiness for 0.3: audit fixes - #883

Merged
jtravs merged 393 commits into
mainfrom
chore/release-0.3-readiness
Nov 18, 2026
Merged

jtravs merged 393 commits into
mainfrom
chore/release-0.3-readiness

Conversation

@jtravs

@jtravs jtravs commented Nov 17, 2026 •

Copy link
Copy Markdown
Contributor

Fixes from the pre-release review of the tree for 0.3.0. Draft. #878 has merged, so this diff is this work alone.

What changed

  • Output schema 3 (was 4): the schema bumps at most once between releases, and 0.2.0 wrote 2. A 0.2.0 or 0.1.0 file is refused naming the file and the release to read it with, tested on every reading path.
  • Scans: a scan that ends early gives its claimed points back (An interrupted Local or Pool scan skips its running points on restart #168); a recipe scan on a queue or Slurm keeps its recipe; an interrupt in a pool worker stops the point instead of failing it; run_scan(stale_age=) is removed.
  • Runs and resume: a failure to close a run stopped by its own flag is raised; a resume runs on a copy and no longer reseeds the caller's Simulation; read_resume_state replaces output.resume.
  • Grids and statistics: field grids refuse ffac < 2.2, so every grid zeroes its Nyquist bin, and the on-axis spectra use the exact end-bin weights on both backends (on_axis_spectrum: native and the Python backends differ at a field grid's Nyquist bin when a hand-built grid keeps it active #155).
  • Modes: the zero-dispersion search counts a sign change of beta2 only where beta1 has a resolved extremum, and takes near= (modes.base.zdw reads an even number of zero crossings as no zero, and says none exists #157, _zdw_near reports sign changes of the differenced beta2 beside a gas resonance as zeros #527); the antiresonant effective-radius coefficient is Rosa 2021's printed 20 (The docstring and the guide disagree about what Rosa 2021 prints for the antiresonant effective radius's coefficient #160); a named gas needs its pressure.
  • Coupling, calibration, material data: regrid crops the band it reports; a lower-level calibration reports the row the store keeps; optmat's internal tracker ids no longer reach warnings or records (Warnings from optmat's Raman ranges reach users carrying optmat's internal tracker ids #166); the ideal-gas path refuses non-finite states.
  • Service and front end: a direct run inside a Slurm allocation is cancelled by its own signal, never by scancel of the allocation; a Stop pressed while a job is queued is obeyed; a resumed run is not read as lost; the retired step window is deleted (about 5,400 lines).
  • Native: a required device is refused before any calibration runs; a CPU build without OpenMP is refused at configure; C++ refusal markers are checked against the sources.
  • Tests: a relative tolerance has no absolute floor (pytest.approx(x, rel=...) kept pytest's 1e-12, which accepted any result for small SI values: 122 assertions could not fail); every np.allclose in tests states its atol; resume normalisation, check_build's refusals, the schema refusals and about forty other behaviours gain tests shown to fail on a planted defect.
  • Documentation: stale statements corrected across the guide and the developer pages; the SEP archive stays in the repository and leaves the built manual; benchmarks.md becomes a short reference; the roadmap and open-questions pages go, their entries being SEPs and issues.
  • Changelog fragments: rewritten so the release notes describe the net change from 0.2.0 and do not contradict each other (200 fragments to 151).
  • Repository: poe docs-clean no longer deletes committed gallery pages; the release checklist gains the GPU lane, the benchmark history and the schema check; two holes in CI's lane filter are closed.

Known state

  • 13 strict expected failures in tests/physdata/test_sellmeier.py record a precision loss in refractive_index_minus_one; the fix is on fix/r03-sellmeier-cancellation, awaiting the maintainer.
  • Committed gallery output for two tutorials predates the coefficient change and needs regenerating.
  • Deferred work is on the 0.4 milestone.

Closes #155, #157, #160, #166, #168, #527.

🤖 Generated with Claude Code

jtravs and others added 29 commits November 17, 2026 22:58
structure.md no longer describes a shared step driver; dependencies.md
and cpu-performance.md give SLEEF's real scope and drop single-precision
FFTW and the rocket-fft row; benchmarking.md no longer says snapshots
need step planes.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The run folder holds the service record and no index or thumbnail,
Stop writes the stop file, every run outlives its window (D21), the
window keeps no pipe to a run, the catalog holds nine classes, launch,
displacement and tilt are delivered, and the front end is published.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The stored transverse DFT, the negative stored k_x of a beam towards
+x, and the signs of displacement and tilt were stated only piecemeal
in SEP 23 and docstrings.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
decisions.md says that where a later record changes an earlier one,
both say so, but many earlier records did not. The earlier records now
name what amended or superseded them: ADR-013, 014, 029 and 056 are
marked superseded (in whole or in part), and the later records' 'stay
as written' clauses now except this log. ADR-074 and ADR-076, merged
with their PRs, read accepted. Records naming hooks, classes, files or
options that no longer exist gain a short 'Now' note of what the
mechanism is called today or what removed it, without rewriting the
decision; ADR-063's note names what ADR-066 removed.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
SEP 26 listed Slurm job ids and cluster node names, and decisions.md
and cpu-performance.md called a machine 'the Xeon'; the public tree
names hardware by its model.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…text

SEP 0 reserves Superseded for a SEP replaced by a later SEP and lets a
Final SEP gain a Superseded-by note. SEP 42, replaced by ADR-066, goes
back to Final (as ADR-066 says), and SEP 33, whose feature ADR-077
removed, names ADR-077 with a note; SEP 0 now says this once. SEP 0's
seam rule names the run request instead of the removed driver surface,
the template's phase reports follow the one-branch rule and no longer
name solprop-kokkos, and the index loses a stray paragraph.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The FFTW-without-NEON Raman rule was described as 'the previous' one;
it now states the rule raman_fft.hpp applies.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
_AdiabaticAmplitude took the paths' kinks and the span, documented them,
needed a positional-parameter suppression, and then deleted both. The
reason it needs no kink check now sits in its docstring.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The protocol blocks keep the corrected member order without the removed
device-memory query; the kernel contract keeps the build-check wording
of the backend-seam fixes.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Docstrings and comments in spatial, responses and ionisation described an
earlier design ("exactly as before", "today's to the last bit", "its
original arithmetic", "used to discard", "contract version 2"); they now
state the invariant. rate_table_key spoke of a cache file where the cache
is in memory, and _finalise copied all twelve fields of a rate table by
hand where dataclasses.replace does.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
- low_storage_cases.assert_parity (and assert_compiled, its only callee)
  and host_cases_envelope_propagation.propagation_studies (and
  refinement_record, used by it alone) had no caller in the tree.
- The radial real-BLAS check compared numpy_ref with itself on its
  numpy_ref instance and skipped its metadata there: it now takes the
  compiled spatial backends only, which uses the spatial equivalence
  selection the audit had found unused, so that selection stays. The collar
  diagnostic, a host computation, runs once rather than once per backend.
- test_a_compiled_cpu_backend_accepts_the_shift constructed and closed a
  backend that the carrier-frame runs construct anyway; the seam freeze's
  table of retired member names restated what the frozen member set already
  holds.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
RamanKernels said a propagation transforms its kernel once per density,
but a run carries only the undamped rows and each backend damps and
transforms them itself. The damped spectrum is what the reference
backend's tests compare against, so the class is kept and its docstrings
say exactly that.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
- The planar statistics test said both backends form the field statistics
  through one host function; native forms them itself, which is what makes
  the comparison parity.
- The general modal module described its leaves as refused until a backend
  executes them; it now runs the route on the reference and holds native's
  ten steps to it.
- Five catalogue row modules cited test files that no longer exist.
- The cycle-rates test checks that the last periodic row is the saved
  plane's scan: its name now says so.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
A bare approx(x) keeps pytest's rel=1e-6 beside abs=1e-12, so for a
nonzero number below 1e-6 the absolute term decides it: in the default
lane 7 such sites accepted any result (a beta2 of 1e-26 s^2/m, a time
of 2e-14 s, an area of 3e-24 m^2) and 5 more were loosened to 2-37 %.
The wrapper now passes rel=1e-6, abs=0 for a bare approx of a nonzero
real or complex number; a bare approx of zero, a sequence, an array or
a mapping keeps pytest's defaults. The GUI suite, which compares time
limits of 1e-13 s bare, gets the same wrapper.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The test compared refractive_index_minus_one at 1e-6 bar with its 1 bar
value scaled by the density, at rel=1e-9; pytest's default absolute
tolerance made that pass. The density scales chi1 = n^2 - 1 exactly,
not n - 1, which departs from proportion by chi1/4 (3.5e-4 for SF6), so
the comparison now goes through (n - 1)(n + 1).

It still fails for every gas, because the function is not what it says:
it forms chi1 as (1 + chi1) - 1, so n - 1 carries a relative error of
about eps/chi1, 1e-12 for helium at 1 bar and 8e-9 to 3.4e-7 at 1e-6
bar, while chi1 itself scales with the density to the last bit. The
test is marked xfail(strict=True) until the source is fixed.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
With the absolute tolerances gone, 22 more tests failed in the default lane.
Each was read against what it claims:

- Tests whose expected value was wrong: the coherence time of a Gaussian
  is sqrt(2) tau, not tau/sqrt(2); the Raman record's window is the
  grid's realised 512 fs, not the 400 fs asked; the mass flow was
  compared with a density at 294 K while it used 293.15 K; the fused
  silica GDD is 36.23 fs^2/mm once Malitson's air-relative expansion is
  made absolute, as optmat makes it, not the textbook 36.16.
- Tests that left out a known effect: the coherent capillary's weak
  collar takes 2.2e-7 of a single bin's amplitude (now run without it);
  the flux-path case's 16 radial nodes lose 6e-6 of the energy even in
  vacuum (now 32, which conserve to rounding); the first radial node is
  off axis by a Gaussian factor of 1.007 (now corrected for); the
  interface's normal-incidence factor gains 1.0e-7 on oblique columns;
  a Planck edge rounds to one over its first 3 %.
- Tolerances below the arithmetic: a sum of 33 weights, axis
  differences a thousand spacings from zero, linspace's own spacing,
  exp(log W) deep in a collar, and the gas step's Fresnel closed form.
- Quantities that are differences at rounding level, now held at the
  scale they are formed from: a lossless transfer's loss, an overlap of
  orthogonal zeros, a Gram defect, a shift between 2 ps delays, an
  error norm beside the state's, and 1 - f for fractions near one.
- Negative assertions given a scale: the shock term moves the field by
  1e-6 of its peak, so it is held at 1e-9 of it with rtol=0; the alarm
  weights differ elementwise, atol=0.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
FieldGrid.from_limits admitted ffac >= 2, but below 2.2 the fine Nyquist
frequency ffac omega_max / 2 falls short of the collar's foot at
1.1 omega_max, so the coarse Nyquist bin stayed inside the band with a
non-zero window. That broke the energy identities that oversample,
spectral_density, node_spectrum and the grids guide state for every grid
solprop builds (2 % at ffac = 2).

from_limits now refuses ffac < 2 * WINDOW_FACTOR on a field grid, naming
the parameter and the reason. At exactly that value rounding can leave
the Nyquist bin an ulp below the foot, where the taper is about 1e-305
rather than zero; the foot is then moved onto the bin. FieldGrid's
validation checks that the collar is exactly zero at the zero-frequency
and Nyquist bins, so a hand-built or replaced grid cannot break the
invariant either. The recipe metadata's ffac floor follows, and the
docstrings, guide, primer and conventions page state the rule as
enforced. The band-plasma refusal test moves to ffac = 2.2, and the grid
property test samples ffac over [2.2, 6] and checks the end bins.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The function reads a file's resumable state into a ResumeState and continues
nothing; solprop.sol already dropped the name as misleading beside resume=True,
`solprop resume` and recipe.resume_file, which do continue a run. The new name
says what it does and sits beside read_snapshot. No alias is kept.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The field had one admissible value and nothing read it: each optmat rate
applies its own energy-loss convention and records it in its own ionisation
entry of the material manifest. Removing it drops the policies.energy_loss
entry from the hashed manifest; the run record's skeleton hash in test_display
moves for exactly that key (restoring it reproduces the former hash) and its
stored numbers are unchanged.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
SnapshotController refused a bad interval or capacity with OutputError, a
RuntimeError, and TDSEArchive a bad path or selection with bare TypeError or
ValueError, while the sibling run controls (progress, resume) raise
DriverError. A new OutputRequestError (SolpropError, ValueError) is now the
one refusal for both output requests; OutputError keeps the file failures.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The reference on-axis spectrum, node spectrum and aperture on-axis
spectrum took fourier.spectral_density of ones as their factor, the
interior weight 2 dt^2 / pi on every bin, while the native kernel gives
the zero-frequency and Nyquist bins of a field grid the analytic
signal's exact weight, a quarter of that. The two backends therefore
defined the statistic differently at the end bins, and the documented
identity (the sum times d_omega is the fluence) held only for native.

The Python sites now take fourier.energy_weights per d_omega, so both
backends define the statistic identically and the identity is exact.
On validated field grids the end bins carry no field, so results there
change only to rounding. New tests widen a grid's collar past its
Nyquist bin, which no validated grid allows, and check the identity
against spatial_fluence and the on-axis time integral, and native
against the definition at every bin.

Fixes #155

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Nothing in production built the old Results step or the views only it
used, the step-page registry, the geometry-page cards' gallery fields,
the start page's pictures, ParameterPage, PinField.set_roots, GUIDED or
EXPLORER_SETUP; tests and one type annotation kept them alive inside the
non-Qt coverage layer. Each was checked dead by grepping its imports and
references across gui/src, gui/tests, gui/tools, solprop.service and the
docs, and by an import graph from the entry points.

The route cards keep what card_of needs to name an Other-geometry recipe
(its only production use), so card_of loses the enabled-only path no
caller took. The plot options lose the helpers only the deleted panel
and Results step called. PinField's Roots button, whose only consumer was
set_roots, did nothing and goes with it. The hidden --workflow option,
accepted and ignored, is removed.

The tests that existed only for that code go with it. Kept and moved:
the card_of assertions (now by key), the Overview naming test (to
test_beam.py) and the modes handler's label and RDW-order test (to
tests/service/test_design.py, with its setup inline, marked medium).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The typeset bracket of Rosa, Melli and Vincetti (2021) eq. 10 is
3 + 2 * 10 wavelength / R_co, and Melli et al. (2025) eq. 7 writes
3 + 20 wavelength / R_co. The code used 10 while its docstring called
that the printed value, and the guide said the opposite. By the guide's
own rule (the printed form where the effect against the FEM data is
negligible) the coefficient is 20, which earlier implementations,
Luna among them, also carry.

The docstrings, the antiresonant guide and the Luna differences page
now state one value. The coefficient is no longer one of the
conventions in which solprop's form differs from Luna's, so the
evacuated production form reproduces Luna's stored value, and the
guide records the FEM comparison of 10 against 20 instead of a
pending decision. The symbolic witness is regenerated with 20, the
FEM regression bounds are re-measured with the same 15 % margin, and a
test pins the coefficient and the effective radius to eq. 10.

Fixes #160

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Home listed runs with the front end's own lister, synchronously on the
window's thread, recognising only folders holding run.h5 or recipe.toml,
while the Runs window and the start-up lookup used
solprop.service.runs.list_runs on a thread. The two could disagree about
what a run is, and a large or network-mounted workspace blocked the
window on every refresh.

Home now shows solprop's summaries (those with a results file) from one
listing read on a thread and handed back through a signal; the window's
first listing also finds a run the last window left going, so opening
lists the workspace once instead of three times. A newer listing drops
an older one still in flight. model.workspace.list_runs and RunEntry go.
The window's runs_listed signal lets the tests wait for Home's listing
instead of asserting on it at once.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A run recorded the SLURM_JOB_ID it inherited as its own job, and took
itself for a Slurm run, even when the service had started it directly.
Cancel then ran scancel on that job: a direct run started from inside an
interactive allocation took the whole allocation, and everything else in
it, down with it. Following such a run also asked the scheduler about
the allocation instead of checking the run's process.

The batch script solprop submits now runs the job command with --slurm,
and only then does a run record its Slurm job. Reading a run's state, and
cancelling it, take the run's job only from submitted.json, the job sbatch
named for it, so a record that names an allocation cannot reach scancel
either. A direct run is signalled, wherever it sits.

The service tests now strip SLURM_* from the environment and put Slurm
commands that refuse first on PATH, so the suite can run inside a Slurm
job without touching the real scheduler.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
jtravs and others added 23 commits November 18, 2026 03:13
execute_scan gave the uncommitted claims back in its finally before
recording provenance and closing the writer. When the scan was failing
because the HDF5 file was, the release raised again, replaced the
scan's own error and skipped the close, so the scan file's lock stayed
held for the rest of the process. The release now runs inside a
try/finally that always records and closes, a release error during a
failing scan is logged while the scan's error is the one raised, and
the writer closes even if recording the provenance fails.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Shortening the fragment dropped the public names 0.2.0 had and this
release removes: COUPLED_SECTIONS and EXPONENTIAL_REPRESENTATIONS of
backends.exponential, the reference backend's helpers attribute,
backends.registry.statistics_refusal, NativeBackend.far_field_statistics
and solprop_native's *_FIELDS tuples (all checked against v0.2.0).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A wall's condition_limit admits exactly one, and a frozen-residual flow
refused at that limit computed the admitted step as zero and crashed on
log10(0) with a bare math domain error instead of raising its
TranslationError. The refusal now says that a limit of one admits no
positive step.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A compressor or transform_limit given an empty or out-of-range modes=
raises ModeError, which their docstrings did not say.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The order check of GaussianPulse and SpatialPulse asked for a Python
int, so an order from np.arange, which 0.2.0's GaussianPulse took, was
refused. The check now reads the order through operator.index, still
refusing a bool, a fraction or a value below one, and stores it as an
int.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The route-construction warning block holds a re-entrant lock while the
caller's code runs. Nothing waits on another thread inside it today; the
comment says why nothing must, so a later change does not deadlock.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… zero-dispersion search

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The native package raises only ImportError, ValueError and RuntimeError as
it loads; nothing on that path raises a SolpropError, so the clause that
kept one from being wrapped never ran. Any ValueError or RuntimeError is
now re-raised as a BackendError with its text and cause.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every record a resume can read is at the current output schema, written by
a summary() that stores quiet_checks and collar_quiet, so the branch for a
record without them was reachable from no file.

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

The same answer as keeping the best so far, ties included (the first
found), without a branch for a later zero that is no nearer, which needs
two zeros within one sample of the same distance. Test that a reference
inside a gas line's sampled bracket is blocked on neither side.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each is a path a user or a failing machine reaches: an unseeded run resumed
by its own Simulation keeps its realisation; channels from the numerical
aperture are one guide by cladding and aperture, and hash alike; an
execution status that is not a JSON object is refused naming the file; a
scan that ended normally reports a failed release and still closes its
writer; a new run refused its folder leaves the earlier run as it was; a
recipe that is not TOML fails as a plain run; and a record that cannot be
cut back after a short write keeps no more events.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The run is not the exact linear flow: the temporal absorber's in-stage
term is its only source, integrated under the step controller, so the
spectral phase is held only to n rtol |y| / |y_k| after n accepted steps
(1.2e-3 rad at the outermost bin held). The residual follows the step
sequence, which rounding moves: launches a few ulps apart measured 5e-12
to 4e-7 rad on both backends, the Linux aarch64 wheel 2.2e-6 rad, against
the fixed 1e-7 rad the test asserted. A run without the absorber is exact
to 1.5e-15 rad. The bound still fails a beta2 off by 1 % on 40 of 41 bins.

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

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The evacuated fibre's beta2 ripples through zero twice within 3 nm near
455 nm. The zero-dispersion search classes both as fine structure and
refuses them, so the tutorial's loop over bracketed sign changes raised.
It now skips a refused bracket, and its committed page is regenerated.

Also name DEFAULT_ZDW_RANGE as a literal in the zdw statistic's
docstring: the cross-reference had no target and failed the docs build.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The native backend steps every run it admits in its own session and
refuses the rest by name, so nothing in solprop raises the warning any
more; the class survived only for filters that named it. Drop it and its
export, the tests' filters and catch-warnings guards for it, and the one
assertion that it was absent. The admission tests keep their checks of
which march stepped the run. The ADRs' Now notes record the removal; the
SEP 47 release note no longer says the class stays, and a removed
fragment says how to migrate.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
All four lessons and 40 tutorials re-run in one build on one Linux machine
(native backend, optmat 0.10.0), so every committed page comes from the
same code and machine. Beyond run times, interpreter versions and
object addresses: the antiresonant tutorial's figures and the interactive
session's RDW roots follow the effective-radius coefficient of 20; the
radial capillary coupling's budget follows the per-field leaky-field
discrepancy; the material-choices digests follow optmat 0.10.0's
equation-of-state and transport data; the rectangular warning prints
loss=model; and a few diagnostics move at rounding level.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Its printed energy losses move at rounding level (6.7e-10 to 6.9e-10 at
0.05 mm) with the mode-set overlap integration; its figures are unchanged.
The other worked examples were re-run in the same build and print the
same results with pixel-identical figures, so their committed pages stand.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…xample

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The free-space route now collects its default statistics when none are
named. The Kerr beams time four fixed steps, and those statistics at the
observed plane cost a quarter of that, so the rows read as a regression
against earlier releases although the propagation is as fast. The
builder names an empty set, which keeps the series measuring the steps.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
All four lessons and 40 tutorials re-run in one build on an Apple M5 Max
(native backend), so the run times they print are those of an unloaded
workstation rather than of a busy cluster node. Results are unchanged
beyond rounding: a few round-off diagnostics, the interactive session's
root-search evaluation counts and two figures' sub-pixel traces move;
the air-mixture tutorial's timing figure follows the new times.
Notebooks, archives and figures whose pixels are unchanged are kept as
committed.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
@jtravs
jtravs marked this pull request as ready for review November 18, 2026 12:58
@chatgpt-codex-connector

chatgpt-codex-connector Bot commented Nov 18, 2026 •

Copy link
Copy Markdown
Codex Review Summary: ✅ Completed

This comment shows the latest Codex review activity on this pull request.

Review Status Commit Review trigger
📝 Code Review ✅ Completed 2026-11-18T13:15:42.793378Z 6ffc6ac New commits
🔒 Security Review ✅ Completed 2026-11-18T13:17:07.795430Z 6ffc6ac New commits
ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

@chatgpt-codex-connector chatgpt-codex-connector Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: e096a8d8ca

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

Comment thread src/solprop/modes/base.py Resolved
The search took a sign change where the product of neighbouring signs is
negative. A sample of beta2 that is exactly zero makes both products
zero, so a zero on a sample was missed and the search reported none. A
zero sample now takes the sign of the last nonzero sample below it,
which puts the change in the interval above; a touch of zero that keeps
its sign is still no change.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
@jtravs
jtravs merged commit 2188a57 into main Nov 18, 2026
76 checks passed
@jtravs
jtravs deleted the chore/release-0.3-readiness branch November 18, 2026 17:05
@jtravs jtravs mentioned this pull request Nov 18, 2026
11 of 13 tasks
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment