fix: don't end a steady-current run on a ramp-up overshoot

[zebraengine]

Problem

The first live calibration probe (session 80, 2026-09-14) held 32.5 A for forty minutes and its window was never fitted — even after a forced refit. The car's ramp-up read 6 → 15 → 24 → 31 → 32.6, 32.7, 32.9, then 37.6, 43.3 (two samples overshooting toward the 48 A it was headed for, in the seconds before the probe cap took effect at +37 s), then 32.5 A for the rest of the hold. _steady_current_prefix treats the first out-of-band sample after a run has started as "the steady run ended", so it kept a 4-sample, 6-second window that failed the very next gate (span >= 1.8 τ).

Every probe lands mid-ramp (the controller caps a few seconds after charging starts), so every probe would have been lost the same way — which defeats both the degradation watch's fixed-condition comparison and the current-exponent fit that #31 added.

Code touched

• wallmonitor/thermal.py: PREFIX_SETTLE_S = 60. In _steady_current_prefix, an out-of-band sample within the first minute of a run discards the run and restarts it instead of ending it. A current change after that still ends it. Docstring + constant comment explain the incident.
• tests/test_wallmonitor.py: direct test on the prefix finder with the exact ramp from session 80 (600-sample window recovered, starting at 32.5 A), and a settled run that a later current change still ends.

Left alone: i_ref (median of the first 10 min), the band, the span cap, sag / free-plateau detection.

Risk

• Fits can appear for sessions that were previously rejected on a ramp wobble. On production that is two: session 80 (the probe, free-running) and session 41 (a 35.5 A regulated window from July — excluded from the drift watch by free_plateau, admitted to the τ/rise medians, where it moves rise_ref_c 36.3 → 36.1).
• The current exponent moves 1.20 → 1.56 ± 0.09 on production, because the probe is its first 32 A point and has the most leverage. The 32.5 A and 44.4 A fits independently agree on n ≈ 1.7; the two 39.6 A fits (both on the two hottest consecutive evenings) read ~3.5 °C above that line, so the regression splits the difference. Monthly probes will pin it. At the operating point: sustainable at 29.5 °C ambient is 45 A (44 under 1.2, 46 under I²).
• Drift watch: verdict unchanged (delta +0.13, CI [−2.83, 3.10], not drifting, no lead). Residual sd rises 0.85 → 1.44 because the 39.6 A pair no longer lines up; ambient coefficient 0.35 ± 0.24 — not significant, compromised_by empty.

Verification

• uv run pytest: 152 passed (one new).
• Replayed on the production DB (read-only scratch worktree on the mini-PC): session 80 fits at 32.5 A, 18.7 °C raw rise, τ 13.75 min, RMSE 0.26, free-running, ambient bracketed 30.98 → 29.05 (sensor). Model: n 1.56, rise_ref 36.1, 19 rise fits.
• Gate-by-gate diagnostic of the original failure: prefix = 4 samples over 6 s, current med 32.7 → span >= 1.8 τ False. Raw vitals dump confirmed the overshoot pattern above.

Deploy: pull + sudo systemctl restart wallmonitor.service. The probe's fit appears retroactively — no new charging session needed.

🤖 Generated with Claude Code