feat: say on the dashboard when an install's fit lands far from the priors

[zebraengine]

Problem

DEFAULT_TAU_MIN = 12 and DEFAULT_RISE_REF_C = 36 come from one verified install. On a fresh install they govern the derate forecast and the amp controller until the first fits land, and once fitted the dashboard said only "fitted from N sessions" — a user whose charger has a 20-minute τ or a 25 °C rise had no way to know that the forecasts before that point were built on numbers that did not describe their hardware. Since #21, a fast-τ install also carries a standing cost that nothing surfaced: the fitter's identifiability gate is floored at the default τ, so short charges there never teach the model.

Item 3 of the "stop the priors from being per-one-install" follow-ups; independent of #22.

Code touched

• wallmonitor/thermal.py — ThermalParams.prior_deviation(): fitted τ and rise against the defaults as fractions, plus notable when either exceeds PRIOR_DEVIATION_FRAC = 0.30 (labelled a heuristic in the comment — it is roughly where the default-driven plateau error exceeds the fit's own noise, not a statistic). None until fitted. Included in as_dict() as model.prior_deviation, so /api/thermal and the SSE thermal frame carry it.
• wallmonitor/static/app.js — the dashboard model note appends, once fitted and notable, which parameter differs and by how much, states that earlier forecasts were rough, and for τ more than 30% under the default adds that only charges ≥ 1.8 × default τ (≈ 22 min) at steady current teach the model. The unfitted note now says the defaults are "from one verified install".
• docs/thermal-model.md — the model bullet says the same.
• tests/test_wallmonitor.py — test_thermal_params_report_prior_deviation: unfitted → None; near the defaults → reported, not notable; far off → notable with the sign the UI keys on; round-trips through as_dict().

Risk

• The note keys on 30% deviation, an arbitrary line. An install at 29% gets no note and one at 31% does; the numbers are in the API either way, and the note is informational — nothing else consumes notable.
• The "≈ 22 min" figure in the note is 1.8 × DEFAULT_TAU_MIN, computed in the frontend from the served default, so it tracks a constant change in thermal.py — but MIN_SPAN_TAU itself is not served; if that multiple changes, the frontend's 1.8 must follow. One-line coupling, commented.
• No behavioural change to forecasts, fits, or alerts. Purely an honesty surface.

Verification

• python -m pytest -q (venv): 112 passed (111 on main + 1 new).
• On the production install, prior_deviation reports τ 11.25 vs 12 (−6%) and rise 35.3 vs 36 (−2%): not notable, so this dashboard shows no note — correct, since the defaults are this install. A fresh install with τ = 20 would see "τ 20.0 min vs the 12 min default — forecasts before its first fitted session were rough"; one with τ = 6 additionally gets the short-charge caveat.

Deploy: restart the service; the note appears on the next dashboard load. No migration.