docs(debug-journal): fold in the 2026-09-22 queue slice; correct the component-thread and shared-fixture entries

[springfall2008]

Automated journal flush — the triage bot folding the day's queued findings into tools/debug-journal.md and re-checking existing entries against what has merged since (verified against main 8b9fe029). A human merge is the review gate. 55 candidates remain queued.

Folded in (10):

• GH#4993 follow-up (SolaX 10402 token thrash) — rewrote the SolaX row. 10402 repeating every cycle with data still flowing is server-side token invalidation between cycles (same token succeeds mid-cycle and is rejected next cycle; stated ~30-day expires_in vs real ~1-cycle token lifetime); prime suspect a second client sharing the same client_id/secret — kept marked suspected, not proven. Each 10402 also inflates the component's lifetime error counter. Secondary finding: "No solar forecast data was returned from HA sensors" on the SolaX Cloud template = no PV forecast source configured at all (the template's pv_forecast_* regexes match nothing without the Solcast HA integration). Code claims re-verified on 8b9fe029: the solax.py auth-error branch clears the token and request_wrapper() retries in-request; the solcast.py HA-sensor fall-through branch; the templates/solax_cloud.yaml regexes.

• GH#5087 (Solis API codes misclassified) — fixed on main since the candidate was written (PR fix(solis): stop retries draining the SolisCloud daily API allowance #5089, merged 2026-09-14): B0115/R0000 are now classified (SOLIS_API_CODES, SOLIS_API_CODES_NO_RETRY) with a per-inverter datalogger-offline cooldown, and automatic_config() now retries via automatic_config_done instead of the first-cycle-only gate. Folded as a fixed entry with the pre-fix(solis): stop retries draining the SolisCloud daily API allowance #5089 mechanism kept (B0115 was treated as rate limiting, R-series fell through to "Unknown code" with the full retry ladder, draining the 200/day allowance).

• GH#5090 (Solis amp↔watt voltage) — fixed on main since the candidate was written (PR fix(solis): use a stable voltage for amp<->watt conversions #5095, merged 2026-09-15): get_nominal_voltage() priority is now solis_nominal_voltage → HV live (>60V) → BMS charge-voltage classification (16S/15S) → 48V fallback. The pre-fix diagnostic (failed-write readbacks at a constant ratio of the target) is kept for older logs; pre-fix(solis): use a stable voltage for amp<->watt conversions #5095 there was no workaround, since solis_nominal_voltage fed only the capacity path.

• GH#5091 (auto-config first-cycle gate) — fixed by the same PR fix(solis): stop retries draining the SolisCloud daily API allowance #5089 (automatic_config_done retry). Folded with Solis: one unreachable inverter keeps the whole component in startup backoff #5093's premise correction (auto-config actually ran at the end of every completed first attempt regardless of poll_success — now moot since the gate is gone). GE Cloud still runs its auto-config under if first: only, noted so its self-heal after an outage is described as the slow one.

• GH#5093 (per-inverter failures folded into one poll_success) — still live on main; folded as still-live with the workaround (list only working SNs in solis_inverter_sn) and the contained fix shape (per-inverter booleans already exist at every poll_success site). PR Solis: don't let a refused startup register read abort startup, and back off B0600 #5185 (open, unmerged) noted as the neighbouring per-inverter startup_reset_registers() backoff.

• GH#5092 (bare ValueError for a missing load source) — new symptom-table row: fetch_sensor_data()'s bare raise ValueError unwinds the whole cycle (the fetch-before-inverter-read ordering is what makes "SoC stops updating"); cloud components set load_today at runtime, so "fix apps.yaml" is the wrong pointer during a component startup delay; no fallback exists before a first successful fetch.

• GH#5094 (Tesla discharge-stop fallback) — appended to the HA write/verify row: adjust_charge_immediate()'s fallback to charge_stop_service with domain="discharge" lands on active discharge to load on Tesla, so freeze/hold cycles issue a discharge command then a hold ~2s later. Re-verified on 8b9fe029 (inverter.py:3089-3090); workaround plus the repeat: False caveat kept.

• GH#5082 (GEC 3-phase freeze export) — appended to the GE Cloud row: inv_support_discharge_freeze is read statically from INVERTER_DEF while inv_has_timed_pause is runtime-probed, so set_export_freeze is never force-disabled for the eco-less 3-phase class; the "grep only for the eco warning" trap noted; the GIV-3HY charge_power_rate-limits-AC-only point kept as hardware-reported; fix precedent PR feat(fox): select Feed-in First for a Fox Cloud freeze export #5038.

• GH#5083 (hold hysteresis unreachable) — new symptom row: the hold gate has no tolerance, the nested 1% hysteresis is unreachable and its else re-enables the window, so ±1% SoC quantisation wobble thrashes charge/hold writes; confirmation recipe and "don't hunt for a v8.53 regression" kept. Structure re-verified on 8b9fe029 (line numbers drifted, structure identical).

• GH#5079 pointers — folded into the rewritten Test-order section: hand-rolled args snapshot/restore remains in test_inverter.py/test_agent_tools.py (≥9 modules at write time — the others were cleaned up); soc_kwh_history = {} restore-to-default in test_calculate_yesterday.py remains latent-fragile; the runner still breaks on first failure.

Existing entries corrected (2 areas):

• Component health row — the entry said ComponentBase.start() runs every component's run() loop "on the main asyncio event loop" and concluded a blocking sync call stalls shutdown, HA callbacks and every other component. Wrong on both counts (candidate 5075, from the PR Run ML training off the event loop and let a shutdown abandon it #5112 review). Re-verified on 8b9fe029: Components.start() starts each component through base.create_task(), which spawns a real non-daemon OS thread per component running its own asyncio.run() — a sync CPU-bound call blocks only that component's loop; the real cost is shutdown, where stop_all() does t.join(5 * 60) per thread after terminate() and the interpreter's non-daemon-thread shutdown then waits again unbounded. PR Run ML training off the event loop and let a shutdown abandon it #5112 (open draft, unmerged) is the fix; the misleading "Lock released" comment is still on main.

• Test-order pollution section + Test-ordering trap bullet — the section said all tests share one PredBat fixture; PR test: give each test its own PredBat rather than sharing one (#5079) #5102 (merged 2026-09-15) gave every registry test its own instance and create_predbat() now pins minutes_now/now_utc from FIXTURE_MINUTES_NOW, closing the GH#5026 wall-clock class the entry had anticipated. References to tools/shuffle_fuzz_known_pairs.txt removed (the file no longer exists). Historical mechanism kept for older logs; the candidate's residue pointers folded in.

Dropped (4) — already covered by existing entries:

• 5071 (debug snapshot ~5MB + unbounded download-all): both halves already in the journal (the debug_history_count max raise with storage arithmetic, and the web download route capped at 16 by PR Raise debug_history_count maximum to 500 for multi-week history (#5070) #5071).

• 5072 (retrain interval vs the 48h staleness cap): already in the Load ML row's "frequency is a separate axis" paragraph.

• 5073 (SolarEdge preserve-charge refusals): core already in the symptom row (idle-reset writes; no failure dedup in write_and_poll_value()); only two case notes folded in (SE preserve flag invisible in dumps; unexplained second execute pass ~46s later).

• 5074 (chat OpenRouter-first error wording): already in the Web/MCP/Chat row.

Quality-gate note: the cspell/markdownlint/black hooks pass (&& gates the quick suite on them); the quick suite then aborted at test_teslemetry_local_weekday_follows_the_base_clock at 00:09 BST — the journal's own documented local-midnight clock-window failure (docs-only tree, second sighting), unrelated to these docs-only changes.

🤖 Generated with Claude Code

[Copilot]

Copilot wasn't able to review any files in this pull request.

---

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