v1.4.1: engine additions for probabilistic forecasting

[yallioux]

What and why

v1.4.1: engine additions for probabilistic forecasting. Patch release: new methods and options only, plus the rejection of one invalid call. Existing calls give the same forecasts (tests/regression/ unchanged apart from the text of one error message). Found while building a probabilistic benchmark harness on fev-bench.

• Fix: a distributional or mixture StaticTAM with the default log target rejects a training target with values <= 0 (they were clipped to 1e-12 before the log, so the fitted scale collapsed or exploded).
• Quantiles: AdaptiveTAM.predict_quantiles (scale renewed online from past standardized residuals), KalmanTAM.predict_quantiles (predictive variance of the filter), SafetyTAM.predict_quantiles (several levels, signed per-tail calibration), OperaTAM(loss_type='pinball', tau=...) and aggregate_quantiles.
• Distributions: target_transform "asinh" and "logit" for the distributional StaticTAM; loss="negative_binomial" and loss="tweedie" with discrete and compound Poisson-gamma quantiles, also for poisson.
• f(x, ..., period=None | 'auto' | (low, high)): the range of one cycle of a Fourier input. The default is unchanged; cyclic=True as a fixed domain was tried first and broke the regression results.
• Threads: limiting the CPU threads through OMP_NUM_THREADS / MKL_NUM_THREADS is documented and tested. A hang after torch.set_num_threads(n) seen on one PyTorch / oneMKL build reproduces without tam, so no engine change.
• Release commit: version 1.4.1, CHANGELOG [1.4.1] - 2026-10-08, version DOI 10.5281/zenodo.23248249 in the README BibTeX.
• Known limits, stated in the docs: the columns of AdaptiveTAM.predict_online carry the internal name __mu__ when the base is distributional; the Kalman variance after a long gap of missing targets is understated (inherited from the filter).

Checklist

Fill one row per line you add to CHANGELOG.md. "Not needed" is an answer only with its reason.

CHANGELOG entry	Test (name; fails before the change?)	math/ page	architecture/ page	User docs (README, THEORY, docstrings)	Callers checked	Regression
Fixed: log target rejects y <= 0	4 in tests/test_distributional.py (fail on 1.4.0)	02_distributional.md	02_distributional_code.md	docstring	mixture path	Unchanged
f(x, ..., period=...)	tests/test_cyclic_fixed_domain.py (13; fail on 1.4.0)	FOURIER.md	06_the_spectrum_api.md	docstring	StaticTAM, AdaptiveTAM windows	Unchanged (default None)
OperaTAM pinball loss, aggregate_quantiles	tests/test_opera_pinball.py (5; fail on 1.4.0)	05_opera_aggregation.md	05_opera_gpu.md	docstrings	square and absolute losses	Unchanged
SafetyTAM.predict_quantiles, calibrate(verbose=False)	5 in tests/test_safety.py (fail on 1.4.0)	04_risk_conformal_evt.md	04_risk_conformal_evt_code.md	docstrings	predict_intervals, ACI	Unchanged
target_transform asinh and logit	tests/test_target_transform.py (12; fail on 1.4.0)	02_distributional.md	02_distributional_code.md	StaticTAM docstring	mixture, conformal, CRPS, anomaly score	Unchanged
AdaptiveTAM.predict_quantiles	tests/test_adaptive_quantiles.py (12; fail on 1.4.0)	01_adaptive_online.md	01_adaptive_code.md	docstrings	plain base unchanged	Unchanged
KalmanTAM.predict_quantiles	tests/test_kalman_quantiles.py (9; fail on 1.4.0)	02_kalman_filter.md	02_kalman_torchscript.md	docstring	filter output bit-for-bit with and without the variance	Unchanged
negative_binomial, tweedie, predict_quantiles for counts	tests/test_count_families.py (16; 15 fail on 1.4.0)	01_reweighted_estimation.md	01_reweighted_estimation_code.md	docstrings, bibliography	other losses keep the same fit path	2 lines of RESULTS.txt: the supported-losses text of the unknown-loss error message
Limiting the CPU threads	tests/test_thread_limits.py (no code change)	Not needed: no theory	04_hardware_memory.md	README	Not needed: no code change	Unchanged
Release 1.4.1 (version, DOI)	Not needed: metadata only	Not needed	Not needed	README, CITATION.cff	Not needed	Unchanged

• Tests pass locally (pytest), and the new tests were written first and fail without the change.
• The documentation pair is updated (math/ for theory, architecture/ for the implementation).
• The docs build without warnings (build_docs.bat; the Documentation workflow runs on main after the merge).
• One CHANGELOG.md line per user-visible change, under [1.4.1] (was [Unreleased]).
• No frozen public API change: existing parameters keep their name, order and default (tests/test_api_contracts.py); new ones are keyword arguments with a default.
• No print() in new code.
• Data used by tests and examples is public.
• Every commit is signed off (git commit -s, see CONTRIBUTING.md).
• Every added file keeps its SPDX-FileCopyrightText and SPDX-License-Identifier header; its author is named as SPDX-FileContributor.

Pre-push verdict

READY TO PUSH against the v1.4.0 baseline, run before the release commit (which only changes versions and dates): tests pass, use cases ran, regression 0 fail / 0 warn. RESULTS.txt differs from 1.4.0 by the two error-message lines above; RESULTS_EXP.txt unchanged.