[Tier 2] Quiet-weekend wins: LP fidelity compounding gains

[albinati]

Tracker for medium-effort improvements that compound LP accuracy over time but don't move savings noticeably on day 1.

Issues

#	Story	Effort	Why it's Tier 2 not Tier 1
#196	V11-C DHW draw learning — rolling 14-day prior	half day	Annual value £30–60 but DHW physics tuning needs careful tests
#238	Daikin physics calibration via 2-hourly Onecta consumption	1 day	Compounds COP accuracy month-over-month, but the auto-tune loop has feedback-loop failure modes that need bounds

Why these two

• Data capture is done for both (`pv_realtime_history`, `daikin_telemetry`, `appliance_jobs.actual_kwh`). The auto-tune feedback loops are what's missing.
• Both have larger blast radius than Tier 1: a bad DHW prior under-prepares Friday-evening showers; a bad COP calibration mistimes Daikin grid-charge windows. Need integration tests + bounded clamps before deploy.
• Neither is urgent — current physics models work, just aren't household-tuned.

Definition of done for the epic

Both child issues closed. After both ship, the LP should:

1. Adjust DHW pre-charge volume per slot based on the trailing 14-day usage prior for that day-of-week × hour-of-day cell ([V11-C] DHW draw learning — rolling 14-day prior #196)
2. Auto-adjust the Daikin COP curve from observed kWh-vs-prediction residuals, bounded by a sanity clamp (Daikin physics calibration via 2-hourly Onecta consumption #238)

Re-evaluate in 30 days post-deploy via LP scorecard (`get_lp_scorecard`): does the forecast_accuracy.load section improve? Does dispatch_accuracy.import % go up?