feat(wandb): WandB plot quality overhaul — deep HPO and XAI insights

[Palamabron]

Summary

• Fix broken charts: prediction histogram reduced from 50k+ rows to 50 bins; per-era line charts now use numeric era_index x-axis (fixes alphabetical string sort)
• Add missing visualizations: drawdown curve, per-era correlation distribution histogram, bar charts for feature exposure, worst stability, and ensemble correlation matrix, line_series chart for feature importance over eras
• Fix WandB string metric bug: feature_importance_model_type moved from wandb.log() (silently coerced to NaN) to wandb.run.summary
• Expand HPO summary table: 18 → 30 columns — adds model_1/2/3_type (split for WandB parallel coordinates), XGBoost/LightGBM hyperparams, feature selection, noise injection, augmentation flags; enables answering "which hyperparams drive performance?"
• Add convergence chart: log_hpo_best_so_far() logs trial and running-best corr_sharpe per step to a search-convergence run, showing whether the HPO search is improving over time
• Remove duplicate metric: metric/corr_sharpe deduplicated in log_hpo_trial_metrics

[Palamabron]

Code review

Found 1 issue:

1. log_hpo_best_so_far calls wandb.init(reinit=True) + wandb.finish() on every trial, creating N independent 1-point WandB runs instead of one multi-step convergence run. The docstring says "Opens or reuses a 'search-convergence' run" but reinit=True always finishes the previous run and starts a fresh one — there is no resume. The convergence chart will never render as a single curve; WandB will show N separate single-datapoint runs in the group instead.

To fix, persist the run ID across calls (e.g. via wandb.init(id=run_id, resume="allow") with the ID stored in a file or passed as a parameter) or collect all points and log them in a single wandb.init block at the end of the HPO search.

https://github.com/Palamabron/AlphaPulse/blob/60f8dd473570c84c6a95f49760e00bd8b9ea85a6/src/alphapulse/logging_/wandb_utils.py#L299-L313

🤖 Generated with Claude Code

_{- If this code review was useful, please react with 👍. Otherwise, react with 👎.}

[Palamabron]

Code review

No issues found. Checked for bugs and CLAUDE.md compliance.

Reviewed the 10 changed files (1,453 additions, 92 deletions):

• shap_report.py — _collect_by_type correctly unifies MultiHeadPipeline and flat Pipeline traversal; XGB/LGBM/CatBoost/sklearn importance extractors are guarded with total > 0 checks; _aggregate_importance correctly averages across model types.
• wandb_diagnostics.py — Era-index fix for chronological ordering is correct; drawdown computation (np.maximum.accumulate) is correct; prediction histogram binning avoids logging 50k rows; _log_era_stratified_importance properly bounds runtime with max_eras=30.
• wandb_utils.py — log_hpo_summary_table column and data counts both increase by 12 (consistent); or-fallback for model_1_* keys is safe since learning rates are always positive; top_level_keys deduplication in log_hpo_trial_metrics correctly removes the metric/corr_sharpe duplicate.
• hpo_pipeline.py — _run_best_trial_diagnostics correctly uses feature_cols to select X_feat; load_train_only_frame is imported at the module level (line 26); 80/20 era split logic is sound.
• run_experiment.py — _build_wandb_config flat-key expansion is correct; is_multihead flag, per-model routing fields, and feature-group sizes are all properly extracted.
• tests/test_xai.py — Good coverage across XGBoost, LightGBM, CatBoost, RandomForest, Ridge (empty), and mixed pipelines.

🤖 Generated with Claude Code

_{- If this code review was useful, please react with 👍. Otherwise, react with 👎.}