docs(debug-journal): fold in the 2026-09-28 queue slice (15 candidates); mark GH#5217 fixed by PR #5218, correct entries overtaken by PR #5245/#5267/#5147

[springfall2008]

Automated journal flush (bot run 2026-09-28 00:15 BST): folds the queue slice into tools/debug-journal.md, verifying each candidate against current main (05639a37) first. 15 candidates processed (5213–5237); 8 left queued (5238, 5241, 5242, 5243, 5252, 5259, 5261, 5269) for the next slice.

Per candidate

• GH#5213 (folded, 3 places). Replay limits → replay section: --redo resets the load model and the live ML forecast cannot be replayed from the dump, so a replay differing from the live plan in exactly the optimiser's marginal ~10p slots is expected — the decisive check is whether the same version's own plan in the attached log contains the slot; plus the temporary-TEST_REGISTRY-test replay recipe. Version-A/B trap → Load ML row: the Pi 5 reporter's "reverting fixed the CPU" was an observation-window artefact (identical fine-tune shape on both versions; short version windows prove nothing against the 2h retrain boundary). Downgrade ABI crash → "Version drift" trap: rolling back to ≤v9.0.x with v9.1.0+ state crashes the first plan cycle (TypeError: must be real number, not list in double_array, kernel ABI 5→7 via PR refactor(export): replace the packed export-limit float with a (mode, target, power) tuple #5047), self-recovers next cycle.
• GH#5214 (folded into the HA write/verify row). SA slot selects can report unknown permanently (the GH#5009 degraded-read shape, persistent), and a user edit of INVERTER_DEF charge_time_format to anything ≠ exactly "HH:MM:SS" is a masking workaround that silently severs SA window control via Predbat dummy entities — with the three tells and the pre-feat: Extend Solar Assistant (Growatt) functionality - Schedules and target SoC #3533 "S" note. Verified on main: the != "HH:MM:SS" branch is at inverter.py:737.
• GH#5215 (folded into the Octopus row). The daily cheap-slot cap is derived per account (get_octopus_slot_max() from the tariff code via has_six_hour_cap()) but enforced per car (fresh slots_per_day locals per car in rate_add_io_slots()/load_octopus_slots()), semantics genuinely open (Octopus's own blog says per-car) — verified live on main; do not assume the docs' per-car sentence is wrong.
• GH#5216 adoption-pass rule (both queue files folded into the GivTCP row). During PR fix(givtcp): fill per-inverter arg slots by REST endpoint index, not discovery order (#5209) #5216's development the re-probe adoption pass judged published_discovery gates before the adopted endpoint had published; the merged PR re-publishes after a fleet-growing rediscover() (verified on main at givtcp.py:529-553) and adds a claim hand-back (claimed_from). Recorded as a design rule for that path, plus the suspected pre-fix(givtcp): fill per-inverter arg slots by REST endpoint index, not discovery order (#5209) #5216 residue (a late-adopted tail endpoint never claims its discovery keys).
• GH#5217 (rewritten as fixed). PR GE Cloud: parse option validation text with bracketed labels without crashing (#5217) #5218 merged 2026-09-26 — the paren-crash-loop now has parse_validation_options() (gecloud.py:405, verified on main), with the paren-less guard the candidate insisted on built in. Entry keeps the pre-GE Cloud: parse option validation text with bracketed labels without crashing (#5217) #5218 signature (too many values to unpack every cycle) and notes the triggering setting was never captured.
• GH#5219 (folded into the charge/discharge-curve row). find_charge_curve() divides by max_power = int(battery_rate_max_charge * MINUTE_WATT) with no zero guard (numerator guard exists, verified on main at inverter.py:1391); trigger is Predbat's own freeze-export 0 W write mirrored into a live BMS rate sensor; second crash site adjust_battery_target_multi() (execute.py:1045) means a guard only in find_charge_curve() is not enough. Rewrote the candidate's "Inverter.init re-runs every 5-minute cycle" general fact: that is pre-fix(inverter): Persist inverters across planning cycles #5126 only — on current main the objects persist and refresh_config() re-reads live config per cycle (verified: inverter.py:580 recomputes battery_rate_max_charge in refresh_config()), which still closes the feedback loop.
• GH#5221 (new symptom row). The export-side ratchet latches the threshold at the lowest = 99 sentinel (fetch.py:2432, verified), not the event price — the dump's stored rate_export_cost_threshold: 120 is a display copy from publish_rate_and_threshold(); includes the counterfactual replay that settles it and the PR fix(fetch): exclude saving-session/Axle boosted minutes from automatic rate thresholds (#5050) #5163 (open) caveat.
• GH#5222 (new symptom row). "Sensor disappears after restart, planning still works" → battery_size_tracking()'s only publish site gated behind if today_key not in existing_history: (inverter.py:796, verified), so a mid-day HA/AppDaemon restart skips the publish until midnight; the trimmed mean is recovered separately. Includes the test-masking trap (the regression test never asserts re-publish).
• GH#5224 (folded into the Solcast row). The Open-Meteo API has no horizon parameter — shading is client-side (shading_factors via gti_hourly_to_period_kwh(), solar_model.py:127, verified); the suspected-not-verified compatibility caveat about the rany2 HA integration's entity shape is kept marked as suspected.
• GH#5225 (folded into the Teslemetry row). Freeze export is disabled for TESLA at exactly one flag (support_discharge_freeze: False, verified at config.py:2266ff), and flipping it alone is not enough — the device-side hold mapping is the second half (non-TBC path writes the raw 99 target into the 81-99 snap band; TBC path has no hold), generalised to every cloud-emulator type.
• GH#5227 (folded into the Gateway MQTT row). Two probed gaps: a sole serial="" slot is bound as the control target (_needs_reconfigure() treats "" as new, verified at gateway.py:1174; firmware rejects it — the 2026-09-16 fleet incident), and if len(status.inverters) == 0: return (gateway.py:815, verified) drops EV data and staleness updates wholesale on a fully-withheld hub. Plus the topology-probe trap (battery-presence filter hides data-less slots).
• GH#5230 (folded into the Web/Chat row). The picker's Default row is a silent no-op after any manual pick (web_chat.py:434 calls set_selected_model() only behind if model_id:, verified; set_selected_model() already pops on falsy id, verified chat_store.py:546) — live on main, draft PR fix(chat): let the Default model be reselected after picking another (#5230) #5274 open; apps.yaml edits do not work around it.
• GH#5232 (folded into the "Reserve keeps ratcheting" row). "Reserve stuck at N%" → first knob is set_reserve_min (default 4); battery_min_soc is a raise-only floor on it; scalars parse fine (resolve_arg() extracts-by-index only for lists). Resolution set_reserve_min: 0.
• GH#5237 (folded into the GH#5050 row + the premise trap). The import-side scan overwrite (rate_import_cost_threshold = highest, verified fetch.py:1307) widens the cheap band under a session stamp (rate_max inflated → plain peak admitted into low_rates, which feeds car planning and the charge_window_best seed); fix direction rate_max_base. Finding 2 → premise-trap bullet: check an item's enable: field in config.py before believing an "expert-gated" claim (the candidate's premise was false on both v9.0.3 and main). Note: the candidate's car_charging_plan_smart: false (default) reference was overtaken by PR fix(car): default car_charging_plan_smart to on (#5237) #5251's default flip — the folded text avoids stating the stale default.
• All 15 candidates were verified before folding; none dropped as duplicates — the journal had no prior coverage of any of them (the processed archive ends at 5211).

Existing entries corrected (merges since the last flush)

• GH#5146 entry (car-charging row), two corrections. (1) "a numeric power sensor reads as 'not charging'" is pre-car_charging_now holds the battery for the car, never adds a car slot #5245 only — since PR car_charging_now holds the battery for the car, never adds a car slot #5245/feat(ohme,myenergi): bind car_charging_now to the charger power sensor #5267 car_charging_now can be a charging power sensor (number in watts counts as charging from CAR_CHARGING_NOW_POWER_W, car_charging_now_value() verified at fetch.py:2576-2596). (2) "plan-table display should match prediction.py's condition" is now implemented: PR fix(output): show a car icon and "Hold for car" reason when discharge is held for a charging car #5147 made the plan table take hold_for_car from the prediction's own record (predict_car_hold_best), shown for at least half a slot on Demand rows.
• GH#5050 row extended (not invalidated): the import-side scan-overwrite distortion from GH#5237 (above).
• Checked the other merges since yesterday's flush (fix(triage): allow printf so the review flow can pipe comments into gh api #5257, fix(octopus): request a replan when Intelligent dispatches change #5250, feat(solis): freeze charge, freeze export and holds on FB00 via the Energy Storage Control Switch #5258, feat(solis): SolisCloud freeze charge and holds via the Battery Reserve #5265, feat(ohme,myenergi): bind car_charging_now to the charger power sensor #5267, triage-bot: ask for evidence/splits before triaging, and wake waiting_for_user on reporter reply #5264, triage-bot: check "me too" comments, trusting them unless clearly unrelated #5268, fix(gecloud): re-assert the AC charge gate when force charging, not just every 24h #5270): no other journal entry invalidated — feat(solis): SolisCloud freeze charge and holds via the Battery Reserve #5265/feat(solis): freeze charge, freeze export and holds on FB00 via the Energy Storage Control Switch #5258 were already documented when they merged, and fix(gecloud): re-assert the AC charge gate when force charging, not just every 24h #5270 (GH#5269) has its candidate still queued for tomorrow.

Gate

./run_pre_commit (from coverage/): all hooks passed (cspell, markdownlint, ruff, black, dictionary sorter — kellerza/rany2 added to the dictionary as genuine identifiers). The quick suite ran and aborted at the documented near-midnight trap: test_teslemetry_local_weekday_follows_the_base_clock (test_teslemetry.py:2490) asserts against the host's datetime.now().weekday() at 00:11 BST, the 00:00–01:00 window the journal itself records as crashing that module (see the "Build test dates from Predbat's clock" trap). Pre-existing test defect, unrelated to this docs-only diff; modules after teslemetry did not run this pass.