Report an undefined control-chart scale rather than collapsing the limits (#557, #558); remove TODO.md

[kgdunn]

Summary

• ControlChart: NaN in the warm-up window silently yields zero-width control limits (s = 0.0) #557: a gap at the start of a series no longer collapses the control limits to zero width. ControlChart.calculate_limits computed its scale as np.sqrt(max(0.0, resids)). That guard was written to stop a negative radicand reaching sqrt, but max(0.0, nan) returns 0.0: nan > 0.0 is False, so max keeps its first argument. Missing values at or near the start propagate through the Holt-Winters recursion and leave every training-sample error NaN, so resids was NaN and the scale silently became zero. Two contributing paths are fixed alongside it, so no RuntimeWarning escapes calculate_limits any more.
• repeated_median_slope silently drops NaN pairs and leaks RuntimeWarnings #558: repeated_median_slope no longer drops missing observations silently. A NaN made the inner np.nanmedian receive an all-NaN list; it warned and returned NaN, which the outer median then discarded, so the slope came from whichever points happened to be clean. Non-finite pairs are now dropped pairwise and up front, the behaviour is documented, and fewer than three finite pairs raises ValueError.
• TODO.md removed. It was a migration index for issues Completed TODO.md items (MV plots, R2 attributes, VIP, jackknife CI, contribution plots) #188-219 and stated it would be deleted once those were triaged, which is now done.

Measured on 40 points of N(100, 2), seed 42:

Input	before	after
clean	s = 2.305649	s = 2.305649 (unchanged)
NaN at index 0	s = 2.302721 + RuntimeWarning	s = 2.302721, no warning
NaN in first 4	s = 0.0, limits 98.83071348008284 to 98.83071348008284	ValueError naming the cause

In the failing case sigma_0 was a healthy 1.092, so the existing zero-variance guard at _holt_winters_warmup_fit never fired. The collapse happened later, on a different quantity.

This is the pattern docs/development/error_handling.rst names explicitly: "Never silently substitute np.nan or 1.0 for an undefined statistic".

Test plan

• uv run pytest full suite: 3219 passed, 41 skipped (all skips are remote-dataset downloads blocked by the sandbox proxy, pre-existing).
• The 11 new tests were run against the unfixed source: 8 fail, 3 pass. The 3 that pass in both are the "clean input is unaffected" guards, which is what they are for. This rules out tests that cannot fail.
• uv run ruff check . and uv run ruff format --check . both pass.
• uv run mypy src/process_improve: no issues in 163 source files.
• Clean-data results verified identical to the last digit before and after (s = 2.305649, target = 99.933).

New coverage: TestControlChartMissingValues (4 tests) and TestRepeatedMedianSlopeMissingValues (7 tests), replacing the four # TODO markers that stood in for them at tests/test_monitoring.py:122-123 and tests/test_regression.py:24-25.

Checklist

• Version bumped in pyproject.toml (PATCH: 1.85.1 to 1.85.2; CITATION.cff kept in step in the same commit)
• Tests added or updated where relevant
• ruff check . passes
• CHANGELOG.md updated

Notes for review

Two things worth a second opinion:

1. Scope: fail loudly, not fix the modelling. The deeper cause is that the recursion's fallback for a missing error (the median of the last ten absolute errors) is undefined when the gap is at the very start, so error_i stays NaN and poisons alpha_hat / beta_hat / sigma_hat for the rest of the series. A chart could instead carry its state forward unchanged across a missing observation, which would produce usable limits rather than an exception. That is a modelling decision about what the chart should do with missing data, closer to Batch missing-data handling and smoothing filters #200, so this PR only stops the silent wrong answer and leaves that choice open.
2. Complexity budget. The added guards pushed _holt_winters_parameter_fit over the C901 / PLR0912 thresholds. Rather than add noqa ([ENG-25] Many noqa: C901 / PLR0912 / PLR0915 / PLR0913 suppressions (69 instances) #307 already tracks 69 such suppressions) the radicand computation was extracted into _training_error_radicand, now shared by the lambda grid search and the final estimate so the two cannot drift, with _tau_from_training_errors deciding what to do about an undefined result. Net complexity of that function is lower than before this PR.

Both issues were split out of #213, which listed them only as missing test cases. #197 and #198 were also reopened during this audit: they had been closed as completed while five of their listed items were still unimplemented.

🤖 Generated with Claude Code

https://claude.ai/code/session_0123j56Hk6jRz9zy91Doc1og

---

Generated by Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 77.77778% with 8 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/process_improve/monitoring/control_charts.py	71.42%	3 Missing and 3 partials ⚠️
...c/process_improve/regression/_robust_regression.py	86.66%	1 Missing and 1 partial ⚠️

📢 Thoughts on this report? Let us know!