Add signal-decomposition degradation analysis

[bmeyers]

Ports the CVXPY seasonal-trend decomposition prototype into a new rdtools.signal_decomposition module and wires it end-to-end through TrendAnalysis. The implementation now provides a unified framework for conventional degradation analysis, expanded lifecycle trend analysis, and SD++ soiling analysis.

New module (rdtools/signal_decomposition.py):

• Public degradation() entry point returning (Rd_pct, Rd_CI, sd_trend_results)
• Three trend types: linear, year-1 piecewise-linear breakpoint, and nonlinear monotone
• Four loss functions: L2, L1, Huber, and conditional quantile
• Optional natural-log transform; log-Huber is the recommended and default configuration
• Missing data handled natively through a masked equality constraint
• Moving-block bootstrap confidence intervals for overall and lifecycle-dependent degradation rates
• Stability and conditional-quantile sweeps for diagnosing model sensitivity and degradation patterns
• Plotting/reporting: plot_decomposition, plot_trend, plot_stability, animate_degradation, and format_degradation_report

SD++ soiling:

• Adds a nonpositive soiling component for dry, accumulation-and-recovery behavior
• Solves a deterministic regularization path and applies an explicit structural detection rule
• Returns a true no-soiling model when no coherent component is detected
• Freezes the detection decision before applying two interval-weighted IRL1 refinement steps to reduce shrinkage
• Reports time-averaged and quarterly loss, insolation-weighted loss when available, local soiling-rate intervals, cleaning-event strength, bootstrap uncertainty, and selector diagnostics
• Adds soiling decomposition plots and formatted reports

TrendAnalysis wiring (rdtools/analysis_chains.py):

• _signal_decomposition_degradation private integration method
• sensor_analysis and clearsky_analysis accept signal_decomposition in the analyses list and pass configuration through sd_kwargs
• plot_signal_decomposition_summary method
• Existing preprocessing can be reused explicitly across repeated analyses

Dependencies and minimum environment:

• Python 3.11 is now the minimum supported version; Python 3.10 was removed from CI and Pixi environments
• CVXPY >= 1.5.3 and SPCQE >= 0.0.3 are hard runtime dependencies
• SciPy minimum raised from 1.8.1 to 1.13.0
• pvlib minimum lowered to the tested 0.11.0 floor
• CVXPY and SPCQE are included in the exact minimum-dependency environment
• check-deps verifies that all direct dependency pins match their declared lower bounds
• test-min builds the Python 3.11 lower-bound environment and runs the complete test suite
• Pixi is pinned in CI for reproducible lockfile behavior

Tests and documentation:

• Expanded signal_decomposition_test.py coverage for trend models, losses, bootstrap uncertainty, structural soiling selection, IRL1 refinement, null behavior, metrics, plots, and reports
• Integration coverage in analysis_chains_test.py
• Complete minimum-environment suite passes with 491 tests
• Added a full signal-decomposition guide covering basic degradation, expanded trend analysis, diagnostic sweeps, and SD++ soiling
• Added public PVDAQ examples with synthetic year-1 breakpoint and dry-soiling signals
• Added generated decomposition and soiling figures
• Updated api.rst and the v3.3.0 changelog

Todo list:

• Code changes are covered by tests
• Code changes have been evaluated for compatibility/integration with TrendAnalysis
• [ ] New functions added to __init__.py
  This is intentionally not being done. The function should be imported as rdtools.signal_decomposition.degradation or from rdtools.signal_decomposition import degradation to avoid clashing with the existing rdtools.degradation module.
• api.rst and the other Sphinx documentation pages are up to date
• Example notebooks have been rerun and result differences scrutinized
• Changelog updated

[bmeyers]

Just finishing up examples and package dependency testing. Ready for general review. Open item: minimum scipy version support.

[codecov-commenter]

Codecov Report

❌ Patch coverage is 91.24886% with 96 lines in your changes missing coverage. Please review.
✅ Project coverage is 94.99%. Comparing base (b5eb6e2) to head (d414a75).

Files with missing lines	Patch %	Lines
rdtools/signal_decomposition.py	91.48%	72 Missing ⚠️
rdtools/degradation.py	84.55%	21 Missing ⚠️
rdtools/plotting.py	95.31%	3 Missing ⚠️

Additional details and impacted files

@@               Coverage Diff               @@
##           development     #518      +/-   ##
===============================================
- Coverage        96.79%   94.99%   -1.81%     
===============================================
  Files               12       13       +1     
  Lines             2342     3437    +1095     
===============================================
+ Hits              2267     3265     +998     
- Misses              75      172      +97     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[bmeyers]

We updated the minimum supported environment to Python 3.11 and defined dev-min as a coherent, tested lower-bound stack. Each direct runtime dependency is pinned to its declared minimum, check-deps verifies those pins remain synchronized, and test-min builds the environment and runs the full test suite. Pixi is also pinned in CI for reproducibility.

The newest package in this minimum environment is CVXPY 1.7.5, released December 5, 2025. As of August 25, 2026, the environment is therefore about 8 months and 20 days old. Most other minimum dependencies date from 2022–2024, with a few newer requirements introduced by the signal-decomposition stack. This gives us a realistic, reproducible minimum environment rather than a collection of independently old version floors that may not build or work together.

[bmeyers]

I am currently working on raising the "age" of the minimum environment, which is currently about 8 months. The new target is 2 years. Cvxpy has been downgraded successfully, and I'm now testing additional binding packages.

[bmeyers]

Follow-up: The minimum-version review is complete. We lowered the tested floors to CVXPY 1.5.3, pvlib 0.11.0, and SPCQE 0.0.3, and regenerated the Pixi lockfile. The complete dev-min suite passes with 491 tests.

The newest signal-decomposition dependency floor is now SPCQE 0.0.3, released October 28, 2024—1 year, 9 months, and 28 days old as of August 25, 2026. This is close to the two-year target while retaining the first stable SPCQE release; the older pre-cutoff release is an alpha with an incomplete public API.

The complete environment still contains one newer direct requirement: bayesian-filters 1.4.5, released October 20, 2025. This is the maintained drop-in replacement for the unmaintained FilterPy package and has no earlier PyPI release, so reverting it solely to improve the environment-age statistic would undo an existing maintenance decision. Subject to that exception, the newly introduced signal-decomposition stack now has a practical minimum age of approximately two years.

[bmeyers]

Last remaining open item is to update the changelog, per the PR template

[bmeyers]

I have no more planned commits for this PR. Ready for final review!