fix(calibration): report calibration and evaluation period metrics separately

[DarriEy]

Problem

run_final_evaluation returned only final_metrics, so the results saver's final_result.get('calibration_metrics', {}) wrote empty dicts. Every HEC-HMS run persisted:

"calibration_metrics": {},
"evaluation_metrics": {},

…even though the log showed a computed KGE. The scores were only recoverable from best_score in the separate best-params file.

The metrics were also computed over the whole post-warmup series as a single block rather than per period, so the evaluation period was never scored at all. Every run logged EVALUATION PERIOD: No evaluation period configured despite the config setting one — silently removing HEC-HMS from any calibration-vs-evaluation generalisation comparison.

Found while reproducing the SYMFLUENCE Paper 3 experiments, where all 17 HEC-HMS calibration cells had empty metrics.

Fix

Split the periods the way the sibling HBV plugin already does, reusing its warmup semantics:

• drop warmup from the full simulation first, then filter to the period — matching how each calibration iteration was scored (run 2002–2009 → drop 365 warmup days → filter to 2004–2007)
• the evaluation period already starts after warmup, so it takes no further skip
• return calibration_metrics and evaluation_metrics alongside final_metrics, keeping unprefixed aliases in final_metrics for existing readers

Verification

Re-ran a HEC-HMS calibration on Bow at Banff with the patched package:

	before	after
calibration_metrics	{}	17 metrics, Calib_KGE 0.851
evaluation_metrics	{}	17 metrics, Eval_KGE 0.658

Calib_KGE 0.851 matches the optimiser's own best_score for that run — an independent check that the calibration-period split is correct. Eval_KGE is information that previously did not exist.

Existing test suite passes (3 passed); ruff clean.

🤖 Generated with Claude Code