Apology for the #211 co-author trailer; fix crashes, silent errors and API input handling (AI agent)

[salihyesil59]

Note about #211 and an apology

First, an apology. The commit merged from #211 (56a95f7, "Fix off-by-one abscissae in non-linear lensing source scaling spline") was prepared with an AI agent, and its message still ends with a co-author trailer naming that AI agent. Because of that trailer GitHub now lists the AI agent as a contributor to CAMB. This does not follow the AI agent guidelines in CONTRIBUTING.md, and I should have removed it before opening that PR. I'm sorry for the trouble.

If you would prefer the AI agent not to appear as a contributor, you could force-push master back to the previous version, before 56a95f7. This PR also contains the #211 spline fix itself, without the co-author trailer, so after that you can merge this PR to get the #211 fix back together with the fixes described below. It's entirely your call, and if you'd rather leave history as it is, that's completely fine too. The commit in this PR has no co-author trailer and uses the "AI agent" author name and commit-message tag as described in CONTRIBUTING.md.

Summary

This PR fixes bugs found while reviewing the Python wrapper and parts of the Fortran code with an AI agent. Each bug was reproduced before it was fixed, and regression tests were added for most of them. Everything is in a single commit.

Settings that killed the Python process

These cases used to end the whole Python process through ERROR STOP, a segfault, heap corruption or a hang, so they could not be caught and would kill e.g. a notebook kernel. They now raise CAMBError or CAMBValueError:

• mathutils.threej / threej_pt with |m| > l. threej_pt is documented to return 0 for these inputs.
• get_background_redshift_evolution / get_background_time_evolution and get_time_evolution after get_background(..., no_thermo=True). The on-demand ThermoData%Init could never work there, because InitVars had not set dtaurec, so TRanges_Add_delta stopped.
• get_time_evolution with k <= 0, times later than tau0, or before any background has been calculated.
• Negative matter power redshifts (segfault).
• min_l other than 1 or 2 (double free).
• Vector modes with massive neutrinos.
• Source windows or custom sources in non-flat models.
• GaussianSourceWindow with sigma <= 0.
• SplinedInitialPower without the needed scalar/tensor table. Fortran now raises a global error instead of using an unallocated spline. Also log-regular tables with fewer than two points or kmin >= kmax.
• Dark energy w(a) tables where a is not increasing.
• ExternalNonLinearRatio grids with fewer than two points or non-increasing values, and use before set_ratio (now a global error rather than error stop).
• EarlyQuintessence with zc earlier than the start of integration (1/astart - 1); the minimizer previously hung or crashed.

Silent wrong results

• Non-linear lensing source scaling spline (same fix as Fix off-by-one abscissae in non-linear lensing source scaling spline #211). State%transfer_times has bounds 0:num_transfer_redshifts+1, but the whole array was passed as the spline knots for scaling(1:n). The knots were therefore shifted by one, and the first knot was uninitialised memory. If that memory held a NaN, the lensed spectra were returned all-NaN without an error; otherwise C_phi had ~5e-5 errors at high L. This PR includes the fix so that it is complete on its own; it merges cleanly with the current master, which already has it.
• Global lensing_method reset. An ini file without lensing_method set the global lensing method to 1. So read_ini, and write_ini which calls it to validate, switched every later Python calculation away from the optimized default (5). This changes BB when AccurateBB=T. The ini reader now keeps the current value when the key is absent. The sample ini files all set lensing_method explicitly, so their results are unchanged.
• get_zre_from_tau for tau < 0.001. Reionization is switched off below this optical depth, and the function returned the unrelated stored redshift (10 by default). For example tau=0.0009 gave zre=10 while tau=0.005 gave zre=1.1. It now returns 0. Also get_zre(tau=0.0) no longer ignores the given tau.
• set_initial_power_table(k, pk=None) passed an empty P(k) array with n=len(k), so Fortran read beyond the array. Spline tables with unsorted or non-positive k were silently accepted and gave garbage interpolation; they are now rejected. Empty tables still clear the spectrum.
• Precision. The lensing potential L-scaling factors returned to Python were computed with default-kind real(), giving ~2e-7 relative errors in C_L^phiphi. The A10 constant was also a single-precision literal.
• set_H0_for_theta did not apply the final H0 through a custom setter_H0, so model state updated by the setter could be inconsistent with the solved H0.

Other fixes

• scalar_coupling_matrix and pcl_coupling_matrix failed with a broadcasting error for mask power spectra longer than 2*lmax+1, the usual case.
• chi_squared accepts lists and non-C-ordered or integer arrays.
• write_ini now writes all AccuracyParams boosts that camb.f90 reads, so e.g. TimeStepBoost or LensingBoost no longer fail the round-trip validation. It also writes vector_mode, which is required when reading, so vector-mode parameters previously always failed.
• read_ini from a URL raises on HTTP errors and always deletes its temporary file. A requests.Response is falsy for error status codes, so the file was leaked.
• IniFile.original_filename is now the top-level file rather than the last INCLUDEd one, so saveFile() and relativeFileName() use the right path.
• symbolic compiles generated code in a private temporary directory with absolute paths. Before, it used shared predictable file names in the system temp directory and os.chdir, so concurrent processes could clash. The gfortran check on Windows now only runs on the first compile, as intended.
• Unknown names for named integer fields now raise ValueError instead of KeyError. Strings assigned to boolean fields raise TypeError instead of IndexError; note that bool("F") is True.
• Plain lists and integer arrays are accepted by:
  • comoving_radial_distance, redshift_at_comoving_radial_distance
  • get_background_densities, get_dark_energy_rho_w, get_Omega
  • get_background_time_evolution, which also fixes get_background_redshift_evolution for scalar z
  • get_lensed_cls_with_spectrum, get_lensed_gradient_cls, get_partially_lensed_cls
  • set_dark_energy_w_a, SplinedSourceWindow bias_kz, ExternalNonLinearRatio.set_ratio
• get_BAO's params argument is now optional, as documented.
• Unknown dotted set_params names raise CAMBUnknownArgumentError.
• CAMBparams.diff aligns differing lines, so extra lines in one structure (e.g. source windows) no longer shift all later fields.
• Fixed a continued character constant in the SP(k) out-of-range warning that inserted a long run of spaces. Docstring fixes and a lensing_method option comment in params.ini.

Behaviour changes to be aware of

• An ini file without lensing_method now uses the current global value (the optimized default 5) instead of forcing 1. This only changes results for command-line runs with accurate_BB = T that did not set lensing_method.
• Unsorted or non-positive k / a spline tables now raise instead of being silently accepted.

Testing

• python -m unittest camb.tests.camb_test: all 22 tests pass, including the new regression tests.
• python -m unittest camb.tests.mathutils_test and camb.tests.spk_test pass.
• python fortran/tests/run_tests.py against CAMB_test_outputs (v2): 136 runs, 0 errors, 0 numerical accuracy mismatches, 0 extra/missing files.
• pre-commit run --all-files is clean.
• The branch merges cleanly with current master.

[salihyesil59]

Thank you so much for taking the time to improve this.

Just a small heads-up: c8fc8b7 has a Co-Authored-By: Claude line. Last time with #211 that made Claude show up as a contributor, so I wanted to mention it in case it's not intended.

Have a great day.

[cmbant]

Many thanks for your PR. Claude made two review iterations on it, I 've not reviewed the code yet.
(For future PRs would help good idea to split into smaller items, e.g. important possibly result-related bugs (array indices), and error reporting nits/other unrelated groups of changes separately).

[salihyesil59]

Many thanks for the review and the helpful feedback. I really appreciate the suggestions and will keep them in mind for future PRs.

[cmbant]

Thanks for the PR, wil merge now!
(if you used forformat for the fortran format, please star https://github.com/cmbant/forformat - can't get it in vscode until it gets 5 stars).

[salihyesil59]

Thank you very much for merging my PR. I really appreciate your time and the opportunity to contribute to this project. I've also starred forformat, happy to support your work. Thanks again for maintaining such a valuable scientific tool.