docs(debug-journal): recover the stalled backlog and correct five entries

[springfall2008]

Recovers the debug-journal backlog that the flush could not land. Rebased onto main now that #4979 has merged, so the diff is a single file: tools/debug-journal.md.

This is the work the 2026-09-06 18:20 and 2026-09-07 00:00 runs did and then discarded: both verified their candidates in full, hit the .claude/ write block, explained themselves and exited 0. Every claim below has been re-verified against main at f0cb3f0d, not carried over on trust.

Folded in

GH#4967 — the consider_full clamp has a twin in the C++ kernel. Extends the car-charging row with the three things the issue does not mention: the clamp is duplicated verbatim at prediction.py:799 and prediction_kernel.cpp:827 (fix one and the plans diverge), octopus_intelligent_consider_full never reaches predict() at all, and car_charging_limit is held by reference and read by execute.py for a real control decision — so a fix must assign a separate model-facing value, not mutate it. Also records the fix-order dependency: that clamp is currently the only thing bounding GH#4952's overestimate.

GH#4975 — new symptom row for "low power mode does nothing". The PV-overlap guard in find_charge_rate() uses a fixed 0.1 kWh threshold (const.py:83), not one scaled to system size, and the GH#4577 dawn split only protects the pre-dawn part of a window — so a window lying entirely in daylight can never engage low power. That is by design (PR #4373), which makes this class of report an enhancement rather than a defect. Worth having as a row because it is a repeat: GH#3311, #4557, #4699, #4975.

GH#4965 — new trap, and the second half is the useful one. The test_control_ledger.py stub trap is minor. The finding worth keeping is that the harness cannot reproduce the crash the issue describes: the clock-skew read sits behind if "discharge_start_time" in self.base.args:, which the mock inverter never satisfies, so an attribute-deletion reproduction passes whether or not the bug is fixed. Verified by stashing the fix and re-running the suite. Assert the guard is present instead.

(Note: the candidate located this behind if self.discharge_enable_time:. That gate closes two lines earlier; the real one is the args check. The observation held, the explanation did not — corrected before folding.)

Rewritten rather than folded

GH#4973 claimed gh search issues was silently dead on the runner. Re-probing dissolved the premise: the search worked, and the "must-hit" query missed because the run had silently corrected a typo in the issue title while retyping it. Folding the original claim would have told future runs to distrust a working tool. The salvageable half — treat an empty duplicate-search result with the same suspicion as a confident one — went into the existing bullet.

Existing entries corrected

Entry	Invalidated by	Change
Octopus row, GH#4946	PR #4957	Fixed; dispatch_billed_off_peak() prices completed dispatches off-peak regardless of location
Fox row, GH#4939	PR #4962	The half the reporter actually hit was event routing on a literal predbat prefix. The zero-caller set_battery_charging_time() half is still live
Carbon row, half (a)	PR #4958	Fixed; one ISO8601 call built in UTC
Carbon row, half (b)	—	Narrowed: the zero-scoring is still in the code, carbon_replicate() only reduces how many minutes reach it
GivTCP row	PR #4864, #4959	The inverter.py:1435 citation had drifted onto unrelated code after REST handling moved to givtcp_rest.py
Two-scans trap	PR #4943	Both halves fixed; was written in present tense as if still live

Fixed entries keep their mechanism rather than being deleted — a reader holding an older log still needs it.

Nothing dropped

All four candidates produced something. Also checked and found already current: the GE Cloud row (corrected in #4969), Sigenergy #4926, Solis #4875/#4923, the memory row, and the still-unmerged Solis guards.

Second commit: a sweep of the run logs

The user asked for the bot's logs to be reviewed for anything that never reached the queue. Three more findings, all verified against main:

• Ohme row, GH#4952 (still live). slot_list() derives energy as watts * hours (ohme.py:124) while the observed 16A→10A→6A taper shows watts is a per-slot cap — 2.55x over on the reporter's data. Ohme publishes estimatedSoc, which would give the true delta, and ohme.py never reads it. The same slots publish as AT_HOME dispatches with no source (ohme.py:670), so the whole plug-in-to-target block is stamped at the low rate. Cross-referenced to the car-charging row, because the predict() clamp is currently the only thing bounding it.
• Octopus row, GH#4950 (fixed, PR fix(octopus): don't replenish the IO low-rate slot cap mid-window at midday #4951). The low-rate slot budget was keyed on the loop minute rather than the slot start, so a noon-straddling window drew a fresh budget at midday. It had no entry at all, and "a cheap slot at an hour Octopus never offered" now has three distinct documented causes in that row.
• New trap: midnight_utc is local midnight, and now_utc is local time. Both names are wrong in the same direction — update_time() sets now_utc_real = datetime.now(self.local_tz) (predbat.py:654). This already sent one PR review hunting for a timezone frame mismatch that does not exist.

Deliberately not folded in: PR #4957's review flagged "three unshared copies of the dispatch predicate". The merge it reviewed has since addressed that — dispatch_billed_off_peak() is a shared helper and the one remaining plain location test carries an explicit #4946 cross-reference. Folding it would have added a stale warning to the file.

Note on spelling

The GH#4973 entry needs to talk about a typo in an issue title. I reworded around it rather than adding a misspelling to .cspell/custom-dictionary-workspace.txt — the dictionary is for real vendor terms.

run_pre_commit: clean.

🤖 Generated with Claude Code

[Copilot]

🔵 Needs a closer look

journal_pr_opened() can crash the daemon on invalid/non-JSON gh output because json.loads() isn’t guarded, undermining the intended “fail closed and preserve the queue” behavior.

Pull request overview

This PR recovers and corrects the triage debug-journal backlog, and hardens the /journal-update flush so verified findings aren’t silently lost when the run can’t actually write/open its PR. It also updates internal agent/skill docs to reference the new journal location.

Changes:

• Move the debug journal reference to tools/debug-journal.md and update the daemon + skill/docs pointers accordingly.
• Gate queue archiving on a PR existing for today’s journal branch (protects against “blocked but exit 0” flushes).
• Fold in multiple debug-journal corrections/additions (Octopus, Fox, Carbon, GivTCP, low-power charging symptom row, etc.) and add tests covering the new flush semantics.

File summaries

File	Description
tools/triage_daemon.py	Moves journal path and changes flush semantics to archive only after a PR exists (adds journal_pr_opened() and passes today).
tools/test_triage_daemon.py	Updates/extends tests for new journal path constraint and “archive only if PR exists” behavior.
tools/debug-journal.md	Recovers backlog and corrects multiple entries; adds new symptom/trap notes and marks fixed items as fixed while keeping mechanism.
CLAUDE.md	Updates debug-journal path reference.
AGENTS.md	Updates debug-journal path reference.
.claude/skills/journal-update/SKILL.md	Updates maintained journal path and git add instructions for the new location.
.claude/skills/issue-triage/SKILL.md	Updates links to the debug journal.
.claude/skills/issue-triage-followup/SKILL.md	Updates links to the debug journal.
.claude/skills/issue-pr/SKILL.md	Updates links to the debug journal.

Review details

• Files reviewed: 1/1 changed files
• Comments generated: 0
• Review effort level: Lite

---

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