Implement CPM-aware threshold tuning with fold-level statistic reuse

[psychelzh]

Objective

Implement a CPM-aware tuning path that reuses fold-level computation across threshold candidates.

Why

Threshold sweeps are a core CPM workload. The current generic tuning path recomputes a full fold fit for each threshold value and pays substantial framework overhead. CPM should be able to reuse fold-local statistics and make threshold tuning much cheaper.

Scope

• Design a tuning runner that stays inside cpmr rather than routing through tune::tune_grid().
• Reuse per-fold artifacts wherever valid, for example edge-behavior association statistics computed from the analysis split.
• Support multiple thresh_level values under a fixed thresh_method.
• Keep leakage safety explicit and testable.
• Benchmark against:
  • repeated native fits without reuse
  • the current tidymodels tuning path
• Ensure the design composes cleanly with nested CV work tracked in Follow-up: Nested CV for model selection safety #31.

Open questions

• How much reuse is possible for alpha and sparsity thresholds respectively?
• Which preprocessing choices must invalidate cached fold artifacts?
• Should the first implementation focus only on threshold tuning, or also prepare for broader CPM strategy comparisons?

Non-goals

• A general-purpose hyperparameter optimization framework.
• Mirroring every feature of tune.
• Optimizing the tidymodels adapter enough to erase framework overhead.

Exit criteria

• A design for reusable fold artifacts is accepted.
• Benchmarks show a meaningful speedup on threshold grids.
• Tests/documentation cover the leakage-safe reuse assumptions.

[psychelzh]

Status after the 2026-03-16 cleanup in main:

• tuning can now target one native surface only: cpm_spec() plus fit_resamples();
• the wrapper layer is gone;
• the resample helper layout is simpler.

The main dependency is still #16: the tuning objective should follow the metric policy chosen there, especially for LOO/LOSO-style settings.

[psychelzh]

After #54, any CPM-aware tuning work should target the newer helper-based spec surface (screen, weighting, model) and focus on reusing Stage 1 fold-local artifacts without mixing in the broader Stage 2 model-family follow-up now tracked in #55.