Smooth-gate transport-species screening for trace species

[djkees]

Summary

Transport-species screening in compute_transport_properties used a hardcoded, always-hard-zero cutoff to decide whether to revive a trace gas species for the transport-species list, ignoring smooth_truncation entirely and drifting out of sync with a custom trace value. This replaces it with the same tsize/esize threshold and compute_nj_effective smooth gate used everywhere else in the solve.

djkees#11

Changes

• source/equilibrium.f90: replaced the hardcoded transport_log_cutoff = 25.328436d0 constant and its hard nj(i) <= 0.0d0 step check with gas_amount_ln_threshold (ion-aware, using tsize/esize) and compute_nj_effective (respects smooth_truncation/truncation_width), matching the pattern used at every other activity-threshold check in this file. Added a comment noting why this loop rarely has anything to revive in the common converged path (species are already screened by the far more permissive log_min floor before this loop runs) and when it actually matters (the singular-matrix fallback path, or a future tighter/configurable reporting floor).
• source/equilibrium_test.pf: strengthened the existing test_equilibrium_transport_ch4_o2_smooth_truncation stub with transport-property assertions; added test_equilibrium_transport_ionized_smooth_truncation (ions + transport + smooth_truncation together, not otherwise covered); added test_transport_screening_smooth_gate_at_threshold, which isolates the changed loop directly by calling compute_transport_properties twice on the same converged solution with one species pinned exactly at its activity threshold — once with smooth_truncation off, once on.

Testing

• cmake --build build-dev --target cea_core_test -j4 && ./build-dev/source/cea_core_test.exe — full 124-test suite (122 pre-existing + 2 new) passes.
• Verified test_transport_screening_smooth_gate_at_threshold actually isolates this change: temporarily reverted just equilibrium.f90 (keeping the new test) and reran — that one test failed as expected (equilibrium_test.pf:4184, the smooth-mode revival assertion), with every other test still passing. Restored the fix and reconfirmed all 124 pass.
• Ran the same CH4/O2 and ionized Li/F2 chamber-state cases end-to-end with smooth_truncation on vs. off through the Python bindings: converged results agree to 1e-8–1e-16 relative (ordinary solve-path noise, not a discrepancy), with a ~4-45% per-solve timing overhead for smooth mode depending on mixture size, consistent with it being an opt-in, numerically-smoother reformulation rather than a different physical model.

Compatibility / Numerical behavior

• No expected changes to numerical results
• Expected changes (explain and provide validation)

For default settings (no custom trace, smooth_truncation off), behavior is unchanged: tsize equals the old hardcoded constant's value by coincidence at convergence, and the full pre-existing test suite (including several exact-reference-value regression tests) passes unmodified. Real behavior changes are scoped to: (1) a custom trace value, which the old code silently ignored for this one loop; (2) ionized mixtures, which now get the same looser esize threshold used everywhere else instead of a flat cutoff; (3) smooth_truncation=True runs, where deep-trace species (near the log_min reporting floor) are now revived to a small nonzero amount instead of an exact hard zero — validated directly by the new isolated test, and shown not to meaningfully affect aggregate transport properties in two realistic end-to-end cases.

---

Drafted with Claude's assistance.

• The root cause (a hardcoded constant that happens to equal the default xsize, and a smooth_truncation flag ignored entirely by this one loop) was found by direct reading of equilibrium.f90 and tracing where eq_soln%nj is populated before this loop runs (compute_reported_nj / log_min), not taken from the issue text alone.
• Numerical claims were independently verified: full existing test suite reruns, an isolated Fortran-level test that was checked to fail against the pre-fix code and pass against the fix, and end-to-end Python-binding comparisons (CH4/O2 and ionized Li/F2 cases) with on/off timing and per-species nj diffs, not just aggregate outputs.

Suggested changelog entry (for CHANGELOG.md, left to the maintainer to place)

Fixed

• Transport-species screening in compute_transport_properties now uses the same tsize/esize activity threshold and smooth_truncation gate as the rest of the equilibrium solve, instead of a hardcoded cutoff that only coincidentally matched the default trace setting and always applied a hard zero/nonzero step. Results are unchanged for default settings; a custom trace value or ionized mixtures may now include slightly different trace species in transport-property output, and smooth_truncation=True runs revive trace species continuously instead of via a step.