STATE.json monthly roll to outbox/archive/STATE_ARCHIVE_2026-09.md is DUE and not done - September started and the August notes are still inline; file is ~30 KB, read in full at every session start.

THE EVENING PASS HAS NOT DELIVERED A BRIEF SINCE 2026-08-31 (57h, three consecutive failures): 09-01 died writing a header only, 09-02 never fired at all (no log exists), and the 09-03 01:35 catch-up was KILLED mid-pass (LastTaskResult 0xC000013A) after it had already refreshed sheet_mirror.json and written a research ledger line but before it wrote a brief or updated STATE.json. Self-retry covers session-limit deaths only - not auth deaths (08-19, 09-01) and not kills. With watchers removed 2026-08-30 there is nothing that notices; the 36h rule at the next live pass is the only detector, and it has now caught the same class of failure three times.

WEEKLY_GOALS.md 'THIS WEEK' is still 'Week of 2026-08-23 (Sun-Sat)', a window that ended 08-29 - 5 days of passes have ranked off an expired header. The 08-29 'Signal - week' task DID fire at 21:14 and exited 1 (saturday_week last_ok still 2026-08-22, 12 days). Raised to him as Q1 in the 08-31 evening brief; no answer, and no evening brief has reached him since, so it was never re-asked. Signal never writes a goal - this needs his ~2 min by voice. Next scheduled week pass: 09-05 21:03.

open_flags entries MUST be dicts {raised, flag} - hud.py is their only reader and calls .get() on each (g_state line 83, renderer 850/852). The 2026-09-02 pass appended bare strings and silently killed the HUD for 2d 9h: index.html froze at 08-31 22:16 while every brief kept closing with the hud link. Normalised 2026-09-03 morning and verified `hud.py --print` exit 0. Nothing alerts on a crashed hud.py since the watchers were removed 2026-08-30, so this recurs invisibly if a pass appends a string again.

SIGNAL'S ENTIRE EVIDENCE BASE IS ON A MACHINE HE MAY NO LONGER USE. At 2026-09-05 20:20 he said the MacBook is now his main laptop. Every scheduled Signal task, all four reporters, the inbound relay daemon, the HUD generator and every tracked artifact run on the Dell, off this disk. Consequence measured today, not predicted: App/ shows 0 files touched for three straight days and usmle 11, and neither zero can be distinguished from invisibility - two reporters hit this wall independently on 09-05 and the app reporter hit it again today. It also silently invalidates day items: the Stitch packet (26 files) and the master-doc W2 dispatch are both Dell-only and cannot be started from the Mac, so any pass that ranks them first is issuing a defective-at-issue item, the 08-31 failure class. Asked once in the 09-05 evening brief, unanswered, NOT re-asked this morning per the no-questions-in-the-morning law. Until he rules, every zero in every brief must be labelled invisible-not-idle, and machine-independent items outrank Dell-bound ones.

RESIDUE OF THE DELETED 09-04 FLAG, carried forward because the fix was never written: the inbound relay's liveness is proven by a HEARTBEAT FILE, not by the running process's path. The 09-04 outage happened because two orphaned daemons from a deleted pre-reorg path kept %APPDATA%/Signal/inbox_daemon_alive.txt fresh, so daemon_alive() returned True and pull_inbox.py stayed hands-off, exiting 0 while doing nothing. Any future move of inbox_daemon.py reproduces it exactly. A path check in daemon_alive() closes it; not written - he is the architect. Verified healthy again this pass by checking the running process's CommandLine (PID 39308, canonical path), which is the check the code lacks.

THIRD DISTINCT PASS-DEATH CAUSE, and it is the one nothing on file predicted: THE DELL SIMPLY SLEEPS. Power-Troubleshooter logged 'returned from a low power state' at 2026-09-08 12:32:51 with no reboot since 2026-08-15; the 09-06 evening pass, both 09-07 passes and the 09-08 morning/midday slots never fired, then Task Scheduler dumped all three of 09-08's passes at 12:38 at once as missed-start catch-up (three concurrent Claude sessions racing over STATE.json, sheet_mirror.json and the ledgers - the 09-08 evening pass had already written the mirror before the morning pass finished). This is NOT the OAuth cause in the 09-02/09-03 flags and NOT the ENOTFOUND cause found 09-04/09-05. Two fixes are needed and neither exists: (1) the pass tasks are not configured to wake the machine, so ~47h of silence produced no alert of any kind - the watchers that used to text him about a dead pass were removed 2026-08-30; (2) catch-up should run at most ONE pass, not three concurrently. Directly downstream of the 2026-09-06 MacBook flag: a laptop he no longer opens is a laptop that sleeps.

THE 2026-09-06 RE-RANK RESTED ON A FALSE PREMISE, corrected 09-08: it ranked research to item 1 because 'both app items are DELL-ONLY' and the stitch packet was 'unstartable from the Mac by construction'. A full workspace snapshot for the Mac has existed since 2026-09-04 at D:\FOR-MACBOOK\workspace\main-workspace\ - counted 09-08: App/design-mockups/ complete, App/master-doc/place_in/ 140 files, place_out/ 5, stitch-upload/ 26 files. No app item was ever Dell-only. THE LIVE PROBLEM IS THE OPPOSITE AND IT IS WORSE: the copy is strictly ONE-WAY - every mtime under it matches the Dell exactly, so nothing has ever been written back from the Mac. Any work he does there is permanently invisible to every reporter, every mtime scan and the ledger Stop hook. Until he rules on where Signal lives, a zero on any disk-based track must be reported as UNOBSERVED, never as a measured zero. === SUPERSEDED 2026-09-16: this flag's own correction is now wrong. D:\FOR-MACBOOK\ does not exist - no D: volume is mounted on this machine (Get-PSDrive / Win32_LogicalDisk return C: only). The snapshot was on an external drive he has since unplugged. Do not cite that path as reachable. See the 2026-09-16 flag.

D:\FOR-MACBOOK\ IS GONE, AND THIS RETRACTS THE 2026-09-08 CORRECTION THAT WAS ITSELF A RETRACTION. The 09-08 morning brief told him, in its second line, that it had 'ranked around a blocker that was not there' because a full 09-04 workspace snapshot existed at D:\FOR-MACBOOK\workspace\main-workspace\. Verified today: there is no D: volume on this machine. Get-PSDrive -PSProvider FileSystem returns C: only; Win32_LogicalDisk returns C: only; Test-Path D:\ is False. It was an EXTERNAL DRIVE and he has unplugged it - matching his own 2026-09-06T09:58:27 dictation, the last one he ever recorded on this machine: 'I plugged in the hard drive, so I need you to do the read up for the hard drive' and then 'So everything is reorganized and I can wipe the stuff off?'. CONSEQUENCE: any pass that ranks an item as reachable because 'the snapshot is on D:' is ranking off a path that does not resolve. Check Test-Path before citing it, every time. Twice now this system has stated the Mac-reachability of his work as fact and been wrong in opposite directions - 09-06 said unreachable when a copy existed, 09-08 said reachable when the copy was about to vanish. The honest standing answer is that Signal cannot see the Mac at all and must say so rather than guess. === HARDENED SAME DAY, 18:2x, by the app reporter's exhaustive background scan (the one that had timed out at 120s when the flag was first written - it completed, exit 0, and found nothing). Three things now pinned down, so this is not 'I could not check': (1) a -Depth 2 search for 'FOR-MACBOOK' under C:\, C:\Users\BYJar, C:\Users\BYJar\Desktop and C:\Temp completed with ZERO hits; (2) there are NO mapped or network drives at all, so it is not behind a disconnected share; (3) the only other volumes on this machine are Image, WINRETOOLS and DELLSUPPORT - Dell OEM recovery partitions with no drive letter - so there is no unmounted data volume waiting to be assigned D:. CONCLUSION, and it is the strongest the evidence supports: no copy of that snapshot exists on this machine's reachable storage, and nothing from the Mac has EVER been synced back here. Note the precise limit - this proves the Dell has no record of his Mac work, NOT that he did or did not do any. Whether he worked on the MacBook stays cannot-tell from this host; Drive metadata is the only channel that answered that question today.

THE 09-06 MACBOOK FLAG IS NO LONGER A RISK, IT IS THE MEASURED STATE - and it now has a number. Signal was dark 2026-09-08 12:47 -> 2026-09-16 18:10 (8 days, ~32 missed pass slots, zero briefs delivered; the last evening brief that actually reached him is 2026-09-05). Across that exact window this disk recorded 0 human file writes, 0 dictations, 0 Telegram messages, 0 Sheet writes and 0 HUD checks - while Google Drive recorded him working on 09-10, 09-12, 09-13, 09-14, 09-15 and 09-16, including ~5 hours on an AI Studio prompt named 'Review Of Software Flowmap' on 09-13 that is unmistakably app work. The evidence base and the worker are now on different machines. Until that is resolved every reporter zero is an artefact of the observer, not a fact about his day, and the reporters cannot tell the difference. THE DECISION IS HIS AND IT IS THE ONLY QUESTION THAT MATTERS: either Signal moves to the Mac (scheduled tasks, reporters, relay, HUD generator all re-homed) or the Dell stays powered as the evidence machine and he accepts that it only ever sees what he syncs to it. Asked tonight as Q1. Do not re-rank anything off Dell mtimes until he answers. NOTE the cheap partial fix that needs no ruling: Drive metadata reads see his Mac work from any machine, so a Drive-backed reporter would have caught all six of those days.

THE DELL-SLEEPS CAUSE FROM 09-08 IS CONFIRMED AND UNMITIGATED, and it is now the dominant failure mode - it has cost more passes than the OAuth expiry and the ENOTFOUND deaths combined. Nothing was done about it on 09-08 beyond raising the flag, and the next 8 days produced not one pass. Four distinct causes are now on file (OAuth expiry 08-19/09-01, ENOTFOUND 09-04/09-05, sleep 09-06/09-08, sleep again 09-09..09-15) and THE SELF-RETRY COVERS NONE OF THEM: it only parses session-limit deaths. A sleeping machine cannot self-retry by construction - the fix has to be outside the session, either a wake timer on the pass tasks (Set-ScheduledTask -Settings (New-ScheduledTaskSettingsSet -WakeToRun -StartWhenAvailable)) or moving the schedule off this box entirely. NOTE the second-order failure that makes this worse than a missed brief: the watchers that used to text him when a pass died were removed 2026-08-30, so a dead Signal is now SILENT. He has no way to learn the system stopped except by opening the HUD and reading a stale timestamp. Whatever he rules on Q1, restore a death alert on some machine that stays awake. === ADDENDUM 2026-09-16 22:07: a second failure shape from the same cause - a pass that hibernates mid-run RESUMES on wake with its original prompt and date, then its runner sends. The 09-08 22:07 evening pass slept 8 days and delivered at 18:26 on 09-16 believing itself a manual catch-up. Any fix must also cover resumed stale sessions, not only missed starts.

THE OBSERVER PROBLEM IS NOW THE WHOLE PROBLEM, and this pass is the cleanest proof yet. Signal was dark 2026-09-16 22:25 -> 2026-09-23 19:20 (6d21h), the SECOND multi-day blackout in three weeks after the 8-day one that ended 09-16 - and in both blackouts he was working the entire time. Measured today: Drive activity on 09-17 (Genie mascot, 8 PNGs + a 3.78 MB animated web component), 09-18 (a complete 45-second launch-video script), 09-20 (4 Year-4 PBH lecture PDFs + the lecture-synthesis pipeline run + a gym session + ~40 PubMed topic screens) and 09-21 (the bladder write-up + a gym session + a waitlist signup) - against App/ cold 20 days, usmle-personal/ 23, landing/ 23, research/ 0 human writes. EVERY disk-based reporter is now structurally blind and has been for 3 weeks. Three separate corrections this pass all trace to the same root: signal-gym reported a 21-day zero that was really my frozen sheet_mirror.json; the abstract book was ranked item 1 five times while his research had moved to a different review; and the 08-31 HTML-chain reading (3 of 6 stages, router absent, h_tools.js absent) was false on all three counts. NONE of these were caught by a disk reporter - all three were caught by the Drive connector. THE STRUCTURAL FIX IS NOT MORE REPORTERS, it is that the Drive connector is the only live evidence channel left and it is available to a Claude session and NOT to the scheduled Python. His ruling asked for on 09-05, 09-08, 09-16 and again tonight - does Signal move to the Mac - is still unanswered and is now blocking the accuracy guarantee that is his stated acceptance test.

FIRST REAL WAITLIST SIGNUP, AND THE SYSTEM DID NOT TELL HIM. StudyFix Waitlist row 2 landed 2026-09-21T14:43:36Z, added to the sheet 3 seconds later - the first signup that is NOT his own NGU address (row 1 is betra.baher.2023@ngu.edu.eg from 08-11). It sat unseen for 2 days because the new-signup watcher was among the three deleted on 2026-08-30 on his own call, with no replacement. The landing page is noindex and has been cold 23 days, so this arrived without any promotion. Two things follow and neither is Signal's to decide: whether a signup alert comes back at all, and whether anyone replies to her. The address is PII - it is in the sheet only, and must never go over Telegram, into sheet_mirror.json, or into hud.py.

THE PIPELINE'S INPUT CHANGED AND NO DOC RECORDS IT. On 2026-09-20 18:32-18:33 he uploaded FOUR Year-4 Public Health lecture PDFs ([PH010] Intro to Scope of Public Health, [PH028] Intro to occupational medicine, [PH031] Physical hazards 1, [PH076] Intro to Research Methodology) and fed [PH031] - annotated in his own hand - into a new AI Studio prompt, 'Medical Lecture Note Synthesis Protocol'. Every doc on file still describes the pipeline's input as App/my-usmle-books/, and the workspace-root CLAUDE.md explicitly calls the user-uploads-a-lecture path 'a different, unbuilt thing'. He is building exactly that thing, on his own coursework, and WEEKLY_GOALS row 1 ('one topic end to end, flowchart -> blurt -> cards -> SBA') no longer names what he is actually running. DO NOT rewrite the goal - it is his. Batched to the evening as one question, together with (a) whether Year 4 has physically started, which reshapes every time estimate via the 2.5-3h daily commute, and (b) whether the YAP/TAZ abstract book is dead or paused now that bladder is the live review.