Revert "MAINT-28: during a real peak the maintenance charge stays paused (#247)"
checks / unit (pull_request) Successful in 27m32s
infra-pin / fresh (pull_request) Successful in 2s

Takes PR #247 (f3f9483) out of release/1.0 so the 19 Nov release ships
MAINT-24 + MAINT-25 without MAINT-28. The rig failed f3f9483
(reports/MAINT-28-test-report.md) and the fix-forward PR #248 review found
a High regression (reports/MAINT-28-fix-review-report.md). MAINT-28
continues on PR #248 after the release.

Co-Authored-By: Claude <noreply@anthropic.com>
This commit is contained in:
glenn schrooyen
2026-11-14 01:33:24 +01:00
co-authored by Claude
parent f3f9483d9c
commit 7bc8ad264b
6 changed files with 3 additions and 363 deletions
@@ -1,36 +0,0 @@
# MAINT-28 - during a real peak the maintenance charge stays paused
**Status: for review.** Tests: `test_maint28.py`. No option added; the peak aim, the month peak and the MAINT-13 numbers are unchanged. MAINT-24 follow-up (rig test finding F-1, `reports/MAINT-24-test-report.md`).
## What happened (rig x10, house 6.5 kW, month aim 5.95 kW, A in maintenance CHARGE)
`peakguard.shaping_suspended()` pauses the charge while the forecast is over the aim or the guard is engaged. MAINT-24 moved the aim from 3450 W to the month aim (SAFETY-80), so the pause no longer lasts the whole peak: the guard gives, its projection drops under 5950 W and it idles ("heading for 5 6xx W, under this month's peak"), the pause ends, the floor charges A 150-460 W, the grid goes to 6.3-6.9 kW, the guard trips again. A flipped charge/give 468 times in 45 min (base 02eac3b 11). The quarters stayed at the aim; the cost was a battery working against itself and a grid that swung 5.4-6.9 kW.
## The change
`peakguard.shaping_held(suspended, held, forecast, aim, house)`: `suspended` is the old `shaping_suspended()`, and once it was true the pause stays until the house (`grid + battery power`, the load without our own charge) is `MAINT_RESUME_UNDER_W` = 300 W under the aim. A guard that is engaged or a forecast over the aim still pauses at once. House unread: the forecast stands in; neither read: no pause (SAFETY-63). `ctl_cycle.py` keeps the flag in `_maint_peak_hold`. The hold is for N=2 and up, and only where MAINT-24 moved the aim (a banked month peak above `peak_cap_w`, so the aim is above cap minus margin). Elsewhere the old rule stays, byte for byte: SAFETY-87's N=1 digests (`n1_charge_low`) and SAFETY-119's N=2 charge at 5050 W over a 5000 W cap with no banked peak are pinned, and 1b5f7c9 behaves as 02eac3b there (a chase that was not a MAINT-24 regression).
Why the house and not the forecast: while A gives, the forecast sits just under the aim by design, so no forecast margin is both safe and reachable. The house is the meter reading plus the batteries' power (`grid + site_batt_w`): the net demand the meter would see with no battery at work. Solar is already in it, and that is right: the aim is a limit on what the meter takes, so a sunny hour that lowers the net demand is real room, and a cloud that raises it brings the hold back at once. It is not the gross load, and it would be wrong to add PV back (that would hold the pause on room the meter really has). It stays put when A switches, except for the age of the battery reading (up to 10 s on the rig); the 300 W line is far larger than that lag moves it.
Why 300 W: the house wanders +-50 W; once the charge resumes it settles at (aim - house) / 2, so 300 W gives at least 150 W, three times the noise, and charging less than that is not worth a restart.
The flag lives only while the charge-only branch is taken: every cycle that reaches the maintenance loop starts it at False, so a maintenance that ended, a mode change or a missing aim cannot carry a stale True into the next charge (test_maint28.py part 1b). `cycle()` has two returns before that loop, and the flag is kept through them on purpose: "inputs missing" and "readings late" (SAFETY-128) are holds of under `stale_after` seconds that write nothing and decide nothing, so the next full cycle sees the same charge and the same peak. A missed poll must not end a peak pause; a reset there would restart the charge on it (`_maint_shape_suspended` is kept through them the same way).
It only ever keeps the maintenance charge off. No cut-back is delayed or weakened: the guard's claim is filed as before. While the pause is held the loop works as at 02eac3b, so A also gives for the house (quarters then sit under the aim, A's energy goes to the load: 1.67 kWh in the 45 min below, 0.45 kWh on 1b5f7c9). Measured alternative, A capped at 0 W while held and the guard idle (review scratch, same 45 min run): A gives 0.44 kWh in the load instead of 1.66, quarters 5919/5915 W, but A switches give/idle 32 times in 40 min. The uncapped give costs ~1.2 kWh round trip per 45-min peak (a few cents) and ~50 min more maintenance charge, far inside the 36 h timeout, so the steady give is kept.
## Numbers (sim, real `cycle()`, A rated 2500 W)
House 3000 -> 6500 W for 45 min -> 3000 W, +-50 W every 5 s, banked month peak 6000 W (aim 5950 W), forecast = quarter average so far plus the rest at the present grid (`scratch/MAINT-28/dev/m28_repro.py`; `test_maint28.py` runs the same shape for 15 min to keep the suite short).
| 45 min load | base 02eac3b | 1b5f7c9 | MAINT-28 |
|---|---|---|---|
| A charge/give reversals | 0 | 52 | 1 (the first trip) |
| A charging seconds in the load (of 2700) | 0 | 367 | 301 (before the first trip) |
| quarter averages (W) | 5959 5950 5950 5438 5502 | 5356 5923 5924 5041 4479 | 5356 4012 4000 4226 4476 |
| A out / in (kWh) | 0.65 / 0 | 0.45 / 0.51 | 1.67 / 0.56 |
| charge back after the load | - | 3 s | 3 s |
Rig (MAINT-24 test report, F-1): 468 reversals in 45 min on 1b5f7c9, 11 on 02eac3b. Live shape (house 3000 W): no pause, A 1475 W, as MAINT-24.
## Tests and mutants
`test_maint28.py`: part 1 the real peak, 1b the stale hold, 2 `shaping_held()` edges, 3 the unchanged cases (digests equal 1b5f7c9). On 1b5f7c9 itself (hold stubbed out) part 1 and 1b fail. Mutants, each red: M1 hold state not kept (`held` always False: the 1b5f7c9 behaviour); M2 resume line `-` to `+`; M3 house not passed; M4 the per-cycle reset removed; M5 the N=1 gate removed (digest `n1_peak`: N=1, banked peak, house 6500 W); M6 the aim gate removed (digest `n2_nobank`).
## Not changed
MAINT-24's rate and its tests, N=1, the drain, modes off (digests of 1b5f7c9). The pause for a late or missing forecast.