FROST PHASE 5 — REGIONAL BASELINES WITHOUT FROST

From a source checkout with the unchanged frost v1/phase4-v1 and parking v4 inputs:
python scripts/build-frost-phase5.py
python tests/frost-phase5.test.py

Requirements: numpy and scipy (versions in runtime.json).
The archive includes outputs, frozen design, coefficients, code and tests.
Original input paths and hashes are in source-register.json. It is not a
standalone raw-source bundle.

Primary comparison: same complete authorities, 2021–2023 development only.
Candidate selection combines temporal and whole-authority spatial validation.
All continuous feature transforms are learned inside each training fold.
The selected structure is fitted on 2021–2023 and frozen BEFORE 2024 outcomes
are evaluated. Source files are loaded for audit; test outcomes are not used in selection. 2024 dynamic controls and calendar effects carry 2023 values.
Realised 2024 miles are used only as the evaluation exposure offset.
No 2025 observations enter modelling. Frost means/bands are read only after
selection for contrast diagnostics; no frost effect is estimated.

Regional intercepts and shared slopes provide parsimonious regional baselines.
The regional_density candidate additionally allows London/SE density slopes.
London and South East are reviewed first; they are not separate arbitrary
models fitted to each held-out year. No authority fixed effects are used.

Interpret implementation completion separately from scientific calibration
and contrast gates. Failed diagnostics do not trigger hidden recalibration.
