HFPropagationModel: negative slant TEC, and non-monotonic delay within a single mode (confirmed with real IRI)

[mijahauan]

HFPropagationModel yields negative slant TEC, and non-monotonic delay within a single propagation mode. Reproduced on AC0G-B4 with a real IRI-2020 ionosphere (source: iri), so this is not the parametric-fallback tier.

Reproduction

from datetime import datetime, timezone
from hamsci_dsp.propagation.model import HFPropagationModel
m = HFPropagationModel(receiver_lat=38.92, receiver_lon=-92.13)
t = datetime(2026, 3, 15, 18, 0, 0, tzinfo=timezone.utc)

predict('WWV', f, t):

f (MHz)	mode	primary_delay_ms
5	1E	4.410
10	1F	4.715
15	1F	6.933
20	1F	4.871

compute_differential_delay('WWV', a, b, t) → implied slant TEC:

pair	Δdelay (ms)	TEC (TECU)
5/10	−0.369	−9.15
10/20	−0.156	−15.45
5/15	−2.488	−52.06
10/15	−2.218	−296.93

Two distinct problems

1. Negative TEC. Slant TEC cannot be negative. All four pairs return negative values spanning a 30x range.
2. Non-monotonic delay within one mode. 5→10 MHz is explainable — those are different modes (1E vs 1F) and the F path is geometrically longer, so more delay at the higher frequency is fine. But 10/15/20 MHz are all 1F, same geometry, and the delays run 4.72 → 6.93 → 4.87. Ionospheric group delay goes as 1/f², so on a fixed path it must fall monotonically. 15 MHz is the outlier.

The implied TEC is meant to be a property of the ray path, not of the frequency pair used to probe it, so the pair-to-pair spread is the symptom that surfaced this.

Where it shows

hf-timestd's tests/test_propagation_model.py::TestHFPropagationModel::test_differential_delay_is_frequency_pair_independent asserts exactly the invariant that fails here. It is currently skipped behind a working-IRI predicate (hf-timestd 1ed8b1a) because the failure could not be attributed on a host where IRI would not build — that attribution is now done, and the skip should come off once this is fixed.

Note the environment trap while investigating: import iri2020 succeeds even when IRI is unusable — it builds its Fortran on first call, so a host without a compiler silently returns source='parametric'. Check point.source == 'iri', not the import.

Impact

Differential science that differences the bias away is unaffected. Absolute delay / slant-TEC consumers should not trust this model until resolved.