fix: Vlasov1d vth conventions

[physicistphil]

(Claude-generated, with some light edits)

Fix vlasov1d normalization convention (√2 in v0) and related conventions audit findings

Summary

Two independent conventions audits (2026-07-16_VLASOV1D_CONVENTIONS_AUDIT.md and 2026-07-16_ADEPT_CONVENTIONS_AUDIT.md, now removed) found that electron_debye_normalization in adept/normalization.py used the most-probable-speed convention v0 = sqrt(2*T0/m_e), while the entire _vlasov1d engine (initializer, collision operators, gold tests, all numeric configs) is built on the RMS convention v0 = sqrt(T0/m_e). This branch fixes that root cause plus the other confirmed findings from the audits, scoped to _vlasov1d and _tf1d.

The dimensionless dynamics were never wrong — the solver core never sees v0, so numeric-input runs are byte-for-byte unchanged. What was wrong is the physical interpretation bolted on at the unit boundary:

• A run with normalizing_temperature: 2000eV and T0: 1.0 was physically a 4000 eV plasma.
• L0 was sqrt(2)·λ_De instead of λ_De, so every dimensional-string input (box sizes in µm, gradient scale lengths, laser k0) was off by √2, and every logged unit (v0, x0, c_light, box_length) was √2-wrong.
• A separate sign error in the NRL Coulomb-log expression gave logLambda_ee = -11.8 (and a negative logged nuee) at 2000 eV / 1.5e21 cm⁻³; corrected to +7.22 and a positive rate.

Changes

Core fix

• adept/normalization.py: v0 = sqrt(T0/m_e) (RMS convention), so x0 is the true Debye length; fixed the Coulomb-log sign (log(sqrt(n0) * T0_eV^-1.25)). vth_norm() is deliberately untouched — its only callers are in vfp1d, which is self-consistently built on sqrt(2T/m) (sentinel: test_kappa_eh).
• adept/_tf1d/modules.py: same two fixes in tf1d's private inline copy of write_units (diagnostics-only; tf1d dynamics don't consume these).

Collision operators (audit F4, F5)

• Dougherty compute_vbar was returning n·ū (missing 1/n) — momentum was not conserved wherever n(x) ≠ 1. Now sum(v·f)/sum(f).
• The Krook target Maxwellian was hard-coded to exp(-v²/2), dragging any T0 ≠ 1 species toward T=1. It now uses the species T0/mass (threaded through cfg.grid.species_params). This path was latent (is_on: False in all shipped configs).

Diagnostics (F6, F7)

• Field moments: the quantity saved as "v" was actually the flux ∫v f dv; it is now saved honestly as "j", and "v" holds the true mean velocity j/n. Pressure and heat-flux moments are now centered on ū, not n·ū.
• The field-moment -flogf entropy had the wrong sign (disagreed with the scalar version); fixed.
• New scalars: nuee_norm (code-unit collision rate, the reference for baseline), mean_kinetic_energy, mean_field_energy, mean_total_energy (electrostatic energy-conservation monitor).

Grid (F11, latent)

• dx = (xmax - xmin)/nx and the semi-Lagrangian interpolation period is xmax - xmin; both previously ignored xmin. No shipped config uses xmin ≠ 0.

Config / cleanup (F3, F12)

• Removed the dead GridConfig.c_light knob (always derived from the normalization; only wavepacket.yaml set it, never read).
• Retuned the srs-debug-small.yaml numeric ey drivers to the corrected ĉ = 15.984: pump k0 = 0.162949, seed k0 = -0.086080, w0 = 1.700942, EPW at kλ_De = 0.249, with the backscatter matching condition documented inline. (Under the old ĉ = 11.30 the triad was internally inconsistent between the ES and EM branches.)
• Fixed the AmpereSolver docstring (claimed a spurious 1/m_s in the current).

Docs & tests

• docs/source/solvers/vlasov1d/config.md now states the normalization convention explicitly (v0 = √(T0/mₑ), L0 = λ_De, Maxwellian exp(-v²/2) at T=1, Bohm–Gross ω² = 1 + 3k², ĉ ≈ 15.98 at 2 keV) and documents the unit traps: species v0/T0 are code-units-only, baseline is a rate in units of ω_p0 (no 2π) with nuee_norm as reference, and the super-Gaussian α pitfall (variance = T0/mass only at m=2; ×1.24 at m=3, ×1.37 at m=4).
• New tests: test_units_boundary.py (dimensional round-trip against CODATA + NRL — the class of test whose absence let the √2 bug survive) and test_fp_momentum_conservation.py (isolated Dougherty operator with 50% density modulation + drift; per-cell n/momentum/energy conservation).
• Regenerated 9 config regression fixtures (resonance, fokker_planck_conservation, multispecies_ion_acoustic) under the corrected normalization, e.g. c_light 11.302 → 15.984, x0 12.14 → 8.584 nm, logLambda_ee -11.78 → +7.22.

Breaking changes / migration

• Reinterpretation of past runs: numeric-input dynamics are unchanged, but every previously logged physical unit was √2-wrong, and a "2000eV" run was really 4000 eV. Anyone who back-computed physical parameters from old runs, or tuned normalizing_temperature to hit a target, must re-interpret.
• Dimensional-string configs (e.g. srs.yaml with xmax: 100um): these now map to the correct code values, so results will differ from pre-fix runs.
• Analysis scripts reading the field moment "v" as a flux must switch to "j"; the -flogf sign also flipped.
• Configs setting grid.c_light must remove it (knob deleted; it was dead).
• User copies of srs-debug-small.yaml or other numeric EM drivers tuned against ĉ = 11.30 need retuning to ĉ = 15.984.
• Downstream consumers of the logged nuee should note the sign/magnitude correction.

Out of scope

The audits' findings for other solvers are not in this branch: _vlasov2d Krook target, _pic1d (inherits F10/F11), vfp1d inverse bremsstrahlung, lpse2d, and the orphaned vlasov1d2v. vth_norm() is intentionally unchanged to protect vfp1d.

Test plan

Run locally on 2026-07-17 after the branch commits. Only pre-existing failures unrelated to this branch remained (3 tf1d resonance tests broken on main via KeyError: 'ax', plus osiris deck-roundtrip tests parametrized with paths from another machine).

• pytest tests/test_vlasov1d/test_units_boundary.py — dimensional round-trip vs CODATA/NRL
• pytest tests/test_vlasov1d/test_fp_momentum_conservation.py — Dougherty conservation with n(x) ≠ 1
• Existing vlasov1d gold tests (Landau resonance, ion-acoustic) — unchanged dimensionless physics
• Config regression fixtures regenerated and passing
• vfp1d test_kappa_eh still passes (guards the untouched vth_norm())