[PROBE: energy] Detect cancelling errors in hourly surface energy budgets

[cehw]

Closes #15

What this probe asserts

A model with +20 W m-2 of surface-energy residual by day and -20 W m-2 by night passes the existing signed cumulative energy_closure. This probe catches that temporal cancellation by requiring every contiguous day/night block to satisfy

mean(abs(Rn - H - LE - G)) <= max(0.05 * mean(abs(Rn)), 2 W m-2).

The seeded hourly case contains two spinup days and fourteen scored days, starting at 06:00 local-solar time so all 28 scored blocks are complete. The new energy_closure_by_phase reuses the existing window/segment machinery; the cumulative criterion is unchanged. reference_coupled now declares its verified PT1H support, and reference_diurnal_bias changes only H using the current timestamp.

The boundary is a snow-free, zero-capacity bare skin with G at the actual soil surface. Negative turbulent fluxes are permitted. Boundary compatibility needs adapter documentation/review; passing does not establish realistic partitioning, thermal inertia, or component accuracy.

Discrimination

Reference model	Expected	Criterion that catches it
reference_coupled	PASS	n/a
reference_diurnal_bias	FAIL	energy_closure_by_phase

Both outcomes hold for the five declared gate seeds. An additional 20-seed check through the real adapter subprocess path found: positive control 20/20 passes, temporal-bias control 20/20 fails all 28 blocks, and the unchanged cumulative criterion still passes the bias control 20/20. Unit regressions also cover cancellation within a phase and weak-radiation allowances.

Validation

• ht validate: 17 probes, 30 models.
• pytest -q: 233 passed.
• ht gate: all 17 probes separate their declared references.
• ht verify-adapter --model <reference> --probe energy/surface-energy-closure: both references pass with 384 rows.
• New-probe gate on Windows restricted to two logical CPUs (processor affinity mask 3): 1.59 s including ten adapter runs; the per-run budget is 60 s. This is a local two-CPU measurement, not a claim about an already-completed GitHub runner.
• ht run --model <model> --probe energy/surface-energy-closure --gate-seeds --csv models/result.csv archived all five existing evaluated models as INCOMPLETE for missing hfls/hfss/hfg; Google and dhbv2 additionally declare no PT1H support. No physical-violation claim is made for them.

README, roadmap, contributor credit, the three-language site/flowchart and the probe-writing reference are synchronized. The focused tests verify deterministic forcing, complete blocks, annotation stripping, the minimum window, tolerance boundaries and spinup exclusion.

Checklist

• There is an accepted proposal issue and this PR closes it
• ht validate passes
• ht gate --probe energy/surface-energy-closure passes
• The generator is deterministic given a seed and commits no data
• The probe runs in under a minute on two CPUs (local affinity-constrained measurement above)
• Tolerance and denominator are justified in probe.yaml
• I constructed a model that passes cumulative closure with unphysical temporal errors, and the new criterion catches it