Fix HyperTreeSTL short-horizon forecasts

[janrth]

Fixes #10.

Summary

• Cap the default STL smoothing window by the current sequence length so reflect padding is valid for short forecast horizons.
• Bypass smoothing for a one-step horizon, where reflect padding cannot be applied.
• Repeat reflected seasonal values when centering short horizons whose length is smaller than the STL period.
• Add regression coverage for fcst_h=3 and fcst_h=1.

Tests

• .venv/bin/python -m pytest tests/test_hypertree_stl.py -q
• .venv/bin/python -m pytest -q

[janrth]

@kashif

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 970ce1bac6

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".

[janrth]

@codex

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. 🚀

Reviewed commit: bb090810e1

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".

[StatMixedML]

Thanks a lot for the detailed report in #10 and for taking the time to put up a fix with tests, @janrth, really appreciated.

It turns out this same short-horizon crash was already addressed independently on our conformal branch back in 8c186ed (2026-06-10), as part of the STL rework. That branch applies the same three fixes you identified:

• caps the default smoothing window by the sequence length (K = min(max_w_model, 2*T - 1)) so reflect padding stays valid,
• bypasses smoothing for the one-step case (T ≤ 1),
• ping-pongs the reflected seasonal values when centering horizons shorter than the STL period.

Because the two fixes overlap, I'm going to close this PR as superseded to avoid a conflicting duplicate, but your end-to-end fcst_h=3 / fcst_h=1 regression tests are actually a nicer, more faithful guard than the unit-level test currently on conformal, so I'd like to encouruage you to open a new PR with the test only.