WP4: evidence scripts - timer baseline, constrained site, forecast, fit, multi-day

[Neal006]

WP4 of prompt.md — scripts only (solutions.md §0, §4, §6; ROADMAP.md §5). Plan: .docs/implementation-plan.md (deviation 9).

The numbers this PR puts on record (real day, python scripts/prove.py --policy timer)

                          kWh    $ bill    kg CO2   peak kW  clean-h %  overage $  short kWh
charge-immediately      378.9     67.42      23.0     101.0       86.3       0.80        3.0
dumb timer 09-14        350.6     91.51      18.6     139.7       89.2      32.88       31.3
noonshift               378.8     65.46       8.1     100.0       95.0       0.00        3.0
saving vs charge-now               2.9%     64.6%      1.0%
saving vs timer                   28.5%     56.2%     28.4%

The judge's kill question, answered: over a wall-clock timer the LP is −56 % CO₂, −28 % bill, −28 % peak, and the timer leaves 31 kWh undelivered while bursting the block. Stated-vs-actual departures: 29/36 left later than they said, 5 earlier (worst −158 min) — drivers under-state, they don't game.

Constrained site (--block 50 --feed 80): the feed binds all day, both policies peak at the 40 kW of headroom; Noonshift −23 % CO₂ with less shortfall (12.5 vs 14.6 kWh) but the bill is +7.4 % — deferred energy lands in the 16–21 h price peak. That is the carbon-vs-tariff conflict (ROADMAP §6.9, per-site W_CARBON), reported as measured, not hidden. The full rung there would sit at 60 × 0.667 kW = 40 kW.

What changed

• prove.py: --policy timer, --block/--feed, --forecast <file> (plans on it, scores on data/signal.json, prints mean |forecast − actual|); every run prints clean-hour share (metrics.md §3), block overage $, kWh short, stated-vs-actual. Gate unchanged.
• fetch_data.py signal --forecast → data/signal_forecast.json from /v3/forecast/historical (the 06:00 issue, horizon_hours=24); signal also stores health_damage ($/MWh) when the plan serves it. Not run here — no WattTime credentials on this machine. If Basic does not serve historical forecasts, the flag works on a live day via the same code path with /v3/forecast.
• scheduler.impact(..., health=True) → health_usd (None without an index), keyword-only; impact_detail carries it. One unit test.
• scripts/fit.py: cash tiers | perks only | not a fit from stated-slack median, tariff spread, demand charge, and --saving-rate (what prove.py measured — the slack rule alone would call Caltech "cash tiers"; 0.6 ¢/kWh measured makes it perks only, matching metrics.md §6b). Caltech stated slack median 3.8 h, distribution 4 / 15 / 17 = solutions.md Fact 1.
• scripts/replay_days.py --days …: min/median/max over data/days/<date>.json; names missing days. Only the committed day is on disk, so the range still needs signal files (fetch_data.py signal --day …).

70 tests pass; CI runs prove.py as before.

🤖 Generated with Claude Code

https://claude.ai/code/session_017thvbGo7rTVnG2f46cLMzU

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.