docs(debug-journal): fold in 2026-09-23 queue slice; correct entries for the #5100/#5102 merge batch

[springfall2008]

Automated journal flush from the triage bot's 2026-09-23 run — per-candidate outcomes below; the daemon opened this PR from branch bot/debug-journal-2026-09-23.

Took the first 15 candidates from the queue in plain filename sort order: 2513, 2689, 2716, 5099, 5100, 5101, 5103, 5105, 5106, 5107, 5112, 5116, 5121, 5122, 5125. 47 candidates remain queued (62 at the top level minus this slice); a slice drains per day.

Every candidate's claims were re-verified against current main (8b9fe029, v9.1.0-58) before folding: each cited symbol was grepped/read on the working tree, and PR merge states were checked via git log --grep.

Folded in (11)

• 2513 — no charge-start hysteresis exists: the execute-side hold band is gated inv_target_soc_percent < 100.0, so a 100% target re-commands the window at SoC 99% (execute.py hysteresis verified on main). Folded into the keep-mechanisms paragraph, with battery_charge_power_curve's model-only scope and the Improvement Proposal: SoC Buffer Handling in Predbat #2628 deadband link.
• 2689 — find_charge_curve IndexError guard fixed in PR Fixes for discharge/charge curve calc #2814 (first in v8.26.2); the four-way guard verified correct at inverter.py on main. Folded into the Charge/discharge curve row with the stale-version triage rule.
• 5099 — two facts: (a) Inverter objects are rebuilt every 5-minute cycle (fetch_inverter_data() defaults create=True; verified execute.py + predbat.py call sites on main) — folded as a correction to the self.base.args trap entry, which said rebuilds happened only on the balance path (which no longer constructs anything since Fold balance_inverters into execute: one source of inverter rate writes #5149); (b) battery_calibration literal-value semantics vs GivTCP's != "Off" detection (re-verified both sides on main) — folded into the GivTCP REST row. The candidate's pre-Fold balance_inverters into execute: one source of inverter rate writes #5149 balance-path claim was dropped as overtaken.
• 5100 — storage expiry is judged by the real clock; save_plan() fixed by PR fix(plan): save_plan() expiry computed from the wrong clock (#5079) #5100. Folded as a new Storage row, including the counterweight rule against "build test dates from Predbat's clock" and the run-the-victim-alone-at-two-times-of-day fuzzer tip.
• 5101 — non-string givtcp_rest entry crashes the v9 poll loop anonymously; re-verified the unvalidated InverterRestState construction (givtcp.py) and the rest_api + "/" + api crash line (givtcp_rest.py) on main. Folded into the GivTCP REST row with the log signature (no discovered N inverter(s) line anywhere; malformed entry produces no Errno-22 retries of its own).
• 5103 — GE Cloud EMS: /settings re-read only on the first tick once an EMS sets polling_mode = False; re-verified the gate and the settings_from_cache startup skip on main (line drift 1808→1978 noted; symbols named). Folded into the GE Cloud row.
• 5105 — the three independent manual-dropdown builders and manual_select()'s inability to drive manual_api; re-verified the three builders and their values not in time_values tails on main. New Manual override selects row.
• 5107 — marginal-rate binary sensors are plan-relative what-if predictions; verified marginal.py exists on main with the same MARGINAL_TIME_OFFSETS/threshold structure and the plan.py call site. New Marginal rate sensors row.
• 5112 — watchdog timeout path cancels the awaiter, not an asyncio.to_thread worker, and never sets api_stop; the wait_for/shield/cancel() shape re-verified on component_base.py on main. Folded into the Component health row, kept marked as a finding on the still-open PR Run ML training off the event loop and let a shutdown abandon it #5112.
• 5116 + 5121 — PV calibration self-reference. Rewritten rather than folded as written: the candidates predate the rest of the fix — main has since merged PR fix(solar): calibrate PV against the raw provider forecast, not Predbat's own calibrated output #5121 plus follow-ups (d9592461, 91b2d8a7) that went further than the candidate's fix idea: a dedicated sensor.<prefix>_pv_forecast_h0_uncalibrated published unconditionally, per-point fallback_to_state, and the h0-now-attribute/state legacy fallback — all verified on main (solcast.py, utils.py, db_engine.py). Folded into the Solcast row as fixed-since with the mechanism kept, plus the surviving history_attribute/db_engine facts and the calibration-test fidelity trap.
• 2716 — the "hold charges at peak" three-path map (plan-side price gate exempts levelled windows; "Hold charging" can be the armed-window fallback; the pre-window freeze write programs 100%); the key execute.py/plan.py lines re-verified on main. New symptom-table row, including the Hold charging target X%-Y% = SoC–target-pair decoder and the unresolved SoC-rising-past-target branch. The candidate's suspected items (what selected markturner's 17:30 window) were dropped as issue-specific.
• 5122 — ECO-branch structural gap ("import instead of discharge" is not a modelled action outside selected windows); verified the ECO Mode branch at prediction.py on main. New symptom-table row.
• 5125 — split in two: the header-icon wiring (published entities, three render sites, CSS in get_header_html) folded into the Web row; the vacuous-red-gate trap (a test importing a not-yet-added constant fails at import and skips every test in the module) added as a new bullet under "Traps when investigating".

Existing entries corrected (3)

• "All tests share one PredBat/HA fixture" (GH#5079 section opening) — no longer true: PR test: give each test its own PredBat rather than sharing one (#5079) #5102 merged, unit_test.py now builds a create_predbat() per registry test; reworded, with the non-registry paths still sharing one instance and the known pairs marked as pre-test: give each test its own PredBat rather than sharing one (#5079) #5102 finds. This is the re-check candidate 5100 predicted would be needed once fix(plan): save_plan() expiry computed from the wrong clock (#5079) #5100/test: give each test its own PredBat rather than sharing one (#5079) #5102 merged.
• The minutes_now residue trap (GH#5026) — the fixture-level pin it said would close the class has landed: create_predbat() now pins minutes_now = FIXTURE_MINUTES_NOW (720) and now_utc from the pinned midnight_utc; updated to say the class is closed for registry tests on current main.
• The self.base.args trap — "inverters are rebuilt from scratch on the balance path and whenever num_inverters changes" corrected to "rebuilt every 5-minute cycle" (GH#5099's verification, against current execute.py), with the pre-Fold balance_inverters into execute: one source of inverter rate writes #5149 balance-path construction marked gone.

Dropped (0)

None — every candidate in the slice verified or was rewritten with its mechanism kept.

Notes

• One new cspell term: uncalibrated (introduced to the codebase by the merged fix(solar): calibrate PV against the raw provider forecast, not Predbat's own calibrated output #5121).
• Quality gate: run_pre_commit ran end to end; the pre-commit hooks (cspell, markdownlint, formatting) passed — the run then aborted at test_teslemetry_local_weekday_follows_the_base_clock at 00:09 BST, the known host-clock-window failure documented in this journal ("Build test dates from Predbat's clock": the test asserts its no-base fallback against the host's datetime.now().weekday() while the fallback is datetime.now(timezone.utc), so it crashes in the hour after local midnight until 01:00). Docs-only change; unrelated to this diff.
• Not merged here — a human merge is the review gate on this file.

[Copilot]

Copilot review overview

🔵 Needs a closer look

Four documentation corrections remain unresolved, warranting human review.

Review effort: Lite
Findings: 2

Open (2)

• Inaccurately claims battery rate curve only models predicted rates · New
• Incorrectly states startup blocks all components on the main event loop · New

What changed in this PR

Updates the debug journal with verified findings and corrections from the 2026-09-23 queue slice.

Changes:

• Adds troubleshooting guidance for battery control, integrations, storage, UI, and symptoms.
• Corrects test-isolation, clock, inverter-rebuild, and PV-calibration documentation.
• Adds uncalibrated to the cspell dictionary.

File	Description
tools/​debug-journal.md	Adds and corrects triage guidance and verified findings.
.cspell/​custom-dictionary-workspace.txt	Adds the uncalibrated term.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[springfall2008]

Automated reply from the triage bot.

Correct, and the entry now says so. Verified on main: execute.py:453 passes self.battery_charge_power_curve into find_charge_rate(), the first return value becomes new_charge_rate → charge_rate (execute.py:490) and is written by adjust_charge_rate() (execute.py:290), so "shapes only the modelled rate" was wrong.

Two qualifications went in with the correction, because both bear on whether the curve could be the mechanism here:

• The curve only selects the written rate when set_charge_low_power is on. With it off, find_charge_rate() returns max_rate unchanged (utils.py, the else at the end of the low-power branch) and the curve shapes only the second return value, which execute.py logs as real and never writes.
• Even in low-power mode the rate search floors at max(400W, battery_rate_min * MINUTE_WATT), so the 0.0 taper above ~97% that get_charge_rate_curve_cached() honours in the model (max(min(rate, max_charge_rate), battery_rate_min), utils.py:2395) never becomes a zero write.

The re-command/hysteresis gap is left as the separate mechanism it is — the curve cannot stop the window being re-commanded on either path.