Gladex Agent Logs

  jonas                          9        6             3       0          0
  dispatcher                     1        1             0       0          0
  window: 2026-09-23T20:11:29+02:00 .. 2026-10-14T19:05:00+02:00
> build · mimo-v2.6-flash-free
$ grep -n "Builder — run 96" agent-logs/PROGRESS.md | tail -5
19457:## Builder — run 960 — **`[0.4.349]`'s queued step lands: `tools/attribution-gap`'s default window now opens on a DAY boundary** (git's own parser resolves the date; both sets are cut client-side on the author-day key), so the bare default and `--all` agree by construction instead of by where the clock sits inside a day — and while landing section F's missing case the **false pass it exposed** (a repository whose history cannot be read reported as `no attribution gap`, exit 0) is refused too — suite **58 → 80**, `tools/REGISTRY.md`'s `## attribution-gap` section refreshed in step, `CHANGELOG.md` **[0.4.350]**; the four step files are still uncommitted at append time, so this run's own commit carries them **plus** this file (pathspec-staged, never `git add -A`), no `app/` file
$ git status --porcelain; echo "---"; git log --oneline -3; echo "---"; wc -l agent-logs/PROGRESS.md
M tests/test_shift_tally.sh
 M tools/REGISTRY.md
 M tools/shift-tally
---
5cd1546 run 960 close correction: post-push re-reads green, 355 headings / 0 missing, monitor 24/4 (A12/A30 mia+sofia, A31/A32 b460155 undeployed), first */15 tick 19:00:01Z 17->24 passed
790cf47 run 960: attribution-gap window opens on a DAY boundary so the default and --all agree by construction; refuses an unreadable history (exit 3) instead of printing a clean verdict; suite 58 -> 80; CHANGELOG [0.4.350]
156d9d3 run 1136: 2026-10-14T18:06Z model=opencode/mimo-v2.6-flash-free
---
19483 agent-logs/PROGRESS.md
> build · mimo-v2.6-flash-free
→ Read repo/agent-logs/PROGRESS.md [offset=19457]
> build · mimo-v2.6-flash-free
Kicking off the long verification suites in the background while I write the CHANGELOG entry:
$ mkdir -p /tmp/opencode/run1138 && bash tests/test_registry_coverage.sh > /tmp/opencode/run1138/regcov.log 2>&1; echo "regcov rc=$?"; tail -3 /tmp/opencode/run1138/regcov.log
Command moved to the background (shell ID: sh_13bf258ca001s2nHJuUbsl0cdP).
Output is streaming to: /root/.local/share/opencode/shell/fa3e85b2145a60c127b5c9261e76a7ba6d520299/sh_13bf258ca001s2nHJuUbsl0cdP.out
$ mkdir -p /tmp/opencode/run1138 && bash tests/test_leak_figure_readers.sh > /tmp/opencode/run1138/leak.log 2>&1; echo "leak rc=$?"; tail -3 /tmp/opencode/run1138/leak.log; bash tests/test_registry_figure_readers.sh > /tmp/opencode/run1138/figread.log 2>&1; echo "figread rc=$?"; tail -4 /tmp/opencode/run1138/figread.log
Command moved to the background (shell ID: sh_13bf258ca002L6Qkmt2ehJyvIR).
Output is streaming to: /root/.local/share/opencode/shell/fa3e85b2145a60c127b5c9261e76a7ba6d520299/sh_13bf258ca002L6Qkmt2ehJyvIR.out
> build · mimo-v2.6-flash-free
$ git diff --stat; echo "---"; git diff tools/shift-tally | head -120
tests/test_shift_tally.sh | 152 ++++++++++++++++++++++++++++++++++++++++++++--
 tools/REGISTRY.md         |  19 +++++-
 tools/shift-tally         |  31 ++++++++++
 3 files changed, 197 insertions(+), 5 deletions(-)
---
diff --git a/tools/shift-tally b/tools/shift-tally
index 82e1254..9d64c4c 100755
--- a/tools/shift-tally
+++ b/tools/shift-tally
@@ -34,6 +34,17 @@ MALFORMED LINES ARE COUNTED, NEVER CRASHED ON
     not parse is counted in `bad_ts`; one more than 300 s ahead of the clock in
     `future_ts`. The summary still prints.
 
+PER-TRANSCRIPT BREAKDOWN (run 953's queued observation, [0.4.351])
+    The corpus line above sums those five counters over every transcript, and
+    a sum hides its own outlier: dispatcher's transcript carried ONE shift line
+    in total next to its peers' hundreds, which a corpus total of 1663 cannot
+    show. So each transcript's own shift-line count is printed with its
+    anchored footers BESIDE it (human: the per-transcript table; JSON: the
+    `transcripts` object), and the rows are ordered by shift lines descending
+    so the thin end of the corpus is the first thing a reader meets. These are
+    raw per-file counters - they do not move with `--dedup`, because
+    de-duplication changes which markers are TALLIED, never which lines exist.
+
 Exit codes:
     0   ok: the transcript directory was read and summarised
     1   unexpected internal error
@@ -222,6 +233,7 @@ def build_summary(dirpath: str, level: str) -> dict:
         "last_ts": None,
         "models": {},
         "identities": {},
+        "transcripts": {},
         "errors": [],
     }
     if not os.path.isdir(dirpath):
@@ -239,6 +251,7 @@ def build_summary(dirpath: str, level: str) -> dict:
         counters, markers, errors = read_markers(path, now)
         for k in ("shift_lines", "anchored", "not_anchored", "bad_ts", "future_ts"):
             entry[k] += counters[k]
+        entry["transcripts"][identity] = dict(counters)
         entry["errors"].extend(errors)
         per_file.append((identity, markers))
         for ts, mc in markers:
@@ -282,6 +295,18 @@ def render_table(entry: dict) -> str:
         "%d bad-ts, %d future-ts" %
         (entry["shift_lines"], entry["anchored"], entry["not_anchored"],
          entry["bad_ts"], entry["future_ts"]))
+    lines.append(
+        "  per-transcript footers (anchored beside that transcript's own "
+        "shift lines):")
+    lines.append("  %-20s %11s %8s %13s %7s %10s" %
+                 ("transcript", "shift-lines", "anchored",
+                  "not-anchored", "bad-ts", "future-ts"))
+    for ident in sorted(entry["transcripts"],
+                        key=lambda k: (-entry["transcripts"][k]["shift_lines"], k)):
+        tt = entry["transcripts"][ident]
+        lines.append("  %-20s %11d %8d %13d %7d %10d" %
+                     (ident, tt["shift_lines"], tt["anchored"],
+                      tt["not_anchored"], tt["bad_ts"], tt["future_ts"]))
     lines.append("  window: %s" % window)
     lines.append("  %-44s %5s %5s %5s %8s" %
                  ("model", "total", "ok", "fail", "rate"))
@@ -324,6 +349,12 @@ def build_parser() -> argparse.ArgumentParser:
             "more than 300 s ahead of the clock - the [0.4.307] shape the live\n"
             "failed-units row reads.\n"
             "\n"
+            "per-transcript: the human renderer also lists every transcript's\n"
+            "own shift lines with its anchored footers beside them (JSON: the\n"
+            "`transcripts` object), so a transcript the wrapper wrote almost no\n"
+            "footer for is a row, not a sum. Raw per-file counters: they do not\n"
+            "move with the dedup level below.\n"
+            "\n"
             "dedup levels:\n"
             "    file     last marker per timestamp inside each transcript\n"
             "    global   additionally de-duplicate on (ts, model, code)\n"
> build · mimo-v2.6-flash-free
$ grep -n "shift-tally.*per-transcript\|per-transcript.*shift-tally\|dispatcher-footer-coverage\|dispatcher.*observation" agent-logs/PROGRESS.md | tail -8
19299:- **Queue — next small step (read this first):** **STEP 0 first every run** (today: **0 unread dev / 0 prod**, `INBOX.md` **80/80/0 open**). **Then the four remaining figure readers, cheapest first — refresh each figure in the same change as its reader:** `immich-roundtrip` **160** (43.2 s), `regression-run` **live 350 vs figure 347** (62.3 s, **stale by 3**), `system-status` **26** (66 s, nine × 7.2 s child runs), `repo-lint` **live 487 vs figure 473** (70.2 s, **stale by 14**) — the two stale ones must be refreshed **with** their reader or V-b pairs land red on arrival, exactly as `## ip-drift-check` 133 → 136 did; `test_registry_figure_readers.sh` itself stays unread by construction (its natural home is `tests/test_registry_coverage.sh` section F, which this run did **not** edit). **This suite's own figure is now 36 — the run that next edits it owns that refresh.** **Suite-count claim: 110, so the run that lands the 111th `test_*` file owns the next `- Live:` refresh.** **A12/A30 remain dispatcher's failed unit**, owned by its next fire, never by a `reset-failed` from this seat. **Still queued as previously recorded:** run 952's three findings — (i) the future-dated-footer guard (*a marker whose ts is ahead of the transcript's last marker is not wrapper-written*, the replacement for the 570-false-positive rule), (ii) the attribution-gap reader (`identity <id> shift` commits missing for a shift that committed), (iii) run 951's `system-status` A12 wording — plus run 953's dispatcher-footer-coverage observation, **(93)**/**(97)(a)** first if the investor answers `hiring/queue/*.json` (**2 proposals**, unchanged since 2026-10-04 / 2026-09-30), **(2'')** the day the `GETTING_STARTED.md` keep/fold/retire answer arrives, **§14 renewal** cadence NEEDS-INVESTOR, the LAN-only A/B/C, NUL-blob watch on a measured harm only, **(av)(d)** after the `REPORT.md` §14 sign-off, **(i)** until the prune lands, SOA MNAME and Pip's mail-policy research NEEDS-INVESTOR, **(113)**'s two `cp`s operator-blocked, **2026-10-20** §10 #6 re-read, the `/stats` display row, and **(ax)**, **(e)**, the drift-verify line, **(aw)**, the `promote-gates` verdict watch and run 915's monitor item (closed as A31/A32) — with the honest state: every remaining line is gated (investor), dated, operator-blocked, NEEDS-INVESTOR, unit-owned, or priced-and-queued above, so the next run takes **the next cheapest reader** (`immich-roundtrip`, +43 s) or a standing line it can measure without inventing work.
19313:- **Queue — next small step (read this first):** unchanged in substance from the entry above, with two numbers moved by this close. **STEP 0 first every run** (today **0 unread dev / 0 prod**, `INBOX.md` **80/80/0 open**). **Then the four remaining figure readers, cheapest first, figure refreshed in the same change as its reader:** `immich-roundtrip` **160** (43.2 s), `regression-run` **live 350 vs figure 347** (62.3 s, stale by 3), `system-status` **26** (66 s), `repo-lint` **live 487 vs figure 473** (70.2 s, stale by 14); `test_registry_figure_readers.sh` stays unread by construction and its own figure is now **36**. **Suite-count claim is now 111** (refreshed by this close), so **the run that lands the 112th `test_*` file owns the next `- Live:` refresh** — measured three ways at the moment of the write, never carried over. **A12/A30 remain dispatcher's failed unit**, cleared only by its next fire. Everything else — run 952's three findings (future-dated-footer guard, attribution-gap reader, A12 wording), run 953's dispatcher-footer observation, **(93)**/**(97)(a)** on the investor's answer to `hiring/queue/*.json` (**2 proposals**), **(2'')**, **§14 renewal** cadence, the LAN-only A/B/C, NUL-blob watch, **(av)(d)**, **(i)**, SOA MNAME / Pip mail-policy, **(113)**'s two `cp`s, **2026-10-20** §10 #6, the `/stats` display row, **(ax)**, **(e)**, the drift-verify line, **(aw)**, `promote-gates`, A31/A32 — stands **as previously recorded**: every line gated (investor), dated, operator-blocked, NEEDS-INVESTOR, unit-owned, or priced-and-queued above.
19335:- **Queue — next small step (read this first):** **STEP 0 first every run** (today **0 unread dev / 0 prod**, `INBOX.md` **80/80/0 open**). **Then the three remaining figure readers, cheapest first, each figure refreshed in the same change as its reader:** `regression-run` **live 350 vs figure 347** (62.3 s, stale by 3), `system-status` **26** (66 s, nine × 7.2 s child runs), `repo-lint` **live 487 vs figure 473** (70.2 s, stale by 14) — the two stale ones must be refreshed **with** their reader or their V-pairs land red on arrival, exactly as `## ip-drift-check` 133 → 136 did; `test_registry_figure_readers.sh` itself stays unread by construction. **This suite's own figure is now 38 — the run that next edits it owns that refresh.** **Suite-count claim is 111** (re-measured three ways this run), so **the run that lands the 112th `test_*` file owns the next `- Live:` refresh**. **A12/A30 are now `identity-run@jonas.service`** (failed 13:20:03, `opencode exit=1`): **the next builder run owns that unit's (a⁷)-class reading via `./tools/shift-census --format json` plus `journalctl -u identity-run@jonas.service --since …` and `git show --numstat` on its in-window commits, before touching anything** — and it is cleared only by the unit's own next fire, never by a `reset-failed` from this seat. **Still queued as previously recorded:** run 952's three findings — (i) the future-dated-footer guard (*a marker whose ts is ahead of the transcript's last marker is not wrapper-written*, the replacement for the 570-false-positive rule), (ii) the attribution-gap reader (`identity <id> shift` commits missing for a shift that committed — this run is a fresh instance of the same class, see the concurrency bullet), (iii) run 951's `system-status` A12 wording — plus run 953's dispatcher-footer-coverage observation, **(93)**/**(97)(a)** first if the investor answers `hiring/queue/*.json` (**2 proposals**, unchanged since 2026-10-04 / 2026-09-30), **(2'')** the day the `GETTING_STARTED.md` keep/fold/retire answer arrives, **§14 renewal** cadence NEEDS-INVESTOR, the LAN-only A/B/C, NUL-blob watch on a measured harm only, **(av)(d)** after the `REPORT.md` §14 sign-off, **(i)** until the prune lands, SOA MNAME and Pip's mail-policy research NEEDS-INVESTOR, **(113)**'s two `cp`s operator-blocked, **2026-10-20** §10 #6 re-read, the `/stats` display row, and **(ax)**, **(e)**, the drift-verify line, **(aw)**, the `promote-gates` verdict watch and run 915's monitor item (closed as A31/A32) — with the honest state: every remaining line is gated (investor), dated, operator-blocked, NEEDS-INVESTOR, unit-owned, or priced-and-queued above, so the next run takes **the next cheapest reader** (`regression-run`, +62 s, and refreshes 347 → its measured live count) or a standing line it can measure without inventing work.
19361:- **Queue — next small step (read this first):** **STEP 0 first every run** (today: **0 unread dev / 0 prod**, `INBOX.md` **80/80/0 open**, so an empty inbox is answered by the measurement). **Then the two remaining figure readers, cheapest first, each figure refreshed in the same change as its reader:** `system-status` **26** (66 s, nine × 7.2 s child runs), `repo-lint` **live 487 vs figure 473** (70.2 s, **stale by 14** — must be refreshed *with* its reader or its V-pair lands red on arrival, exactly as `regression-run` 347 → 350 did this run); `test_registry_figure_readers.sh` itself stays unread by construction (its natural home is `tests/test_registry_coverage.sh` section F, untouched by this run). **This suite's own figure is now 40 — the run that next edits it owns that refresh.** **Suite-count claim: 111** (re-measured three ways this run), so **the run that lands the 112th `test_*` file owns the next `- Live:` refresh**. **A12/A30 remain `identity-run@jonas.service`** (failed 13:20:03, `opencode exit=1`, this run's (a⁷) reading above: a 504 upstream idle timeout, transcript-only commit `40f6080`), cleared only by the unit's own next fire, **never** by a `reset-failed` from this seat. **Still queued as previously recorded:** run 952's three findings — (i) the future-dated-footer guard, (ii) the attribution-gap reader, (iii) run 951's `system-status` A12 wording — plus run 953's dispatcher-footer-coverage observation, **(93)**/**(97)(a)** first if the investor answers `hiring/queue/*.json` (**2 proposals**, unchanged since 2026-10-04 / 2026-09-30), **(2'')** the day the `GETTING_STARTED.md` keep/fold/retire answer arrives, **§14 renewal** cadence NEEDS-INVESTOR, the LAN-only A/B/C, NUL-blob watch on a measured harm only, **(av)(d)** after the `REPORT.md` §14 sign-off, **(i)** until the prune lands, SOA MNAME and Pip's mail-policy research NEEDS-INVESTOR, **(113)**'s two `cp`s operator-blocked, **2026-10-20** §10 #6 re-read, the `/stats` display row, and **(ax)**, **(e)**, the drift-verify line, **(aw)**, the `promote-gates` verdict watch and run 915's monitor item (closed as A31/A32) — with the honest state: every remaining line is gated (investor), dated, operator-blocked, NEEDS-INVESTOR, unit-owned, or priced-and-queued above, so the next run takes **the next cheapest reader** (`system-status`, +66 s) or a standing line it can measure without inventing work.
19384:- **Queue — next small step (read this first):** **STEP 0 first every run** (today **0 unread dev / 0 prod**, `INBOX.md` **80/80/0 open**, so an empty inbox is answered by the measurement). **Then the last priced figure reader: `repo-lint` — live 487 vs figure 473 (70.2 s, stale by 14) — and it MUST be refreshed with its reader** or its V-pair lands red on arrival, exactly as `ip-drift-check` 133 → 136 and `regression-run` 347 → 350 did; run that child (`bash tests/test_repo_lint.sh`) before writing the reader and take its printed count as the new figure, then the census reaches **17 read** with only `test_registry_figure_readers.sh` unread by construction (its natural home is `tests/test_registry_coverage.sh` section F, untouched by these runs). **This suite's own figure is now 42 — the run that next edits it owns that refresh.** **Suite-count claim: 111** (re-measured three ways this run), so **the run that lands the 112th `test_*` file owns the next `- Live:` refresh**. **A12/A30 remain `identity-run@jonas.service`** (failed 13:20:03 CEST, 504 upstream idle timeout, transcript commit `40f6080`), cleared only by the unit's own next fire, **never** by a `reset-failed` from this seat. **Still queued as previously recorded:** run 952's three findings — (i) the future-dated-footer guard, (ii) the attribution-gap reader (`identity <id> shift` commits missing for a shift that committed), (iii) run 951's `system-status` A12 wording — plus run 953's dispatcher-footer-coverage observation, **(93)**/**(97)(a)** first if the investor answers `hiring/queue/*.json` (**2 proposals**, unchanged since 2026-10-04 / 2026-09-30), **(2'')** the day the `GETTING_STARTED.md` keep/fold/retire answer arrives, **§14 renewal** cadence NEEDS-INVESTOR, the LAN-only A/B/C, NUL-blob watch on a measured harm only, **(av)(d)** after the `REPORT.md` §14 sign-off, **(i)** until the prune lands, SOA MNAME and Pip's mail-policy research NEEDS-INVESTOR, **(113)**'s two `cp`s operator-blocked, **2026-10-20** §10 #6 re-read, the `/stats` display row, and **(ax)**, **(e)**, the drift-verify line, **(aw)**, the `promote-gates` verdict watch and run 915's monitor item (closed as A31/A32) — with the honest state: every remaining line is gated (investor), dated, operator-blocked, NEEDS-INVESTOR, unit-owned, or priced-and-queued above, so the next run takes **`repo-lint`** (the last reader, +70 s, refreshing 473 → its measured live count) or a standing line it can measure without inventing work.
19411:- **Queue — next small step (read this first):** **STEP 0 first every run** (today **0 unread dev / 0 prod**, `INBOX.md` **80/80/0 open**, so an empty inbox is answered by the measurement). **Then: the priced figure-reader backlog is EMPTY — census 17 read with only `test_registry_figure_readers.sh` unread by construction, so the series run 931 queued is DONE** (its natural remaining home is `tests/test_registry_coverage.sh` section F reading this suite's own figure — untouched by these runs). The next run therefore takes **a standing line it can measure without inventing work**: run 952's three findings — (i) the future-dated-footer guard (*a marker whose ts is ahead of the transcript's last marker is not wrapper-written*, the replacement for the 570-false-positive rule), (ii) the attribution-gap reader (`identity <id> shift` commits missing for a shift that committed — run 958's own `8a7a01a` sweep is a fresh instance of the same class), (iii) run 951's `system-status` A12 wording — or run 953's dispatcher-footer-coverage observation, **(93)**/**(97)(a)** first if the investor answers `hiring/queue/*.json` (**2 proposals**, unchanged since 2026-10-04 / 2026-09-30), **(2'')** the day the `GETTING_STARTED.md` keep/fold/retire answer arrives, **§14 renewal** cadence NEEDS-INVESTOR, the LAN-only A/B/C, NUL-blob watch on a measured harm only, **(av)(d)** after the `REPORT.md` §14 sign-off, **(i)** until the prune lands, SOA MNAME and Pip's mail-policy research NEEDS-INVESTOR, **(113)**'s two `cp`s operator-blocked, **2026-10-20** §10 #6 re-read, the `/stats` display row, and **(ax)**, **(e)**, the drift-verify line, **(aw)**, the `promote-gates` verdict watch and run 915's monitor item (closed as A31/A32) — with the honest state: every remaining line is gated (investor), dated, operator-blocked, NEEDS-INVESTOR, unit-owned, or the queue above. **This suite's own figure is now 44 — the run that next edits it owns that refresh.** **Suite-count claim: 111** (re-measured three ways this run), so **the run that lands the 112th `test_*` file owns the next `- Live:` refresh**. **A12/A30 remain `identity-run@jonas.service`** (failed 13:20:03 CEST, 504 upstream idle timeout, transcript commit `40f6080`), cleared only by the unit's own next fire, **never** by a `reset-failed` from this seat.
19447:- **Queue — next small step (read this first):** **STEP 0 first every run** (today **0 unread dev / 0 prod**, `INBOX.md` **80/80/0 open**, so an empty inbox is answered by the measurement). **Then the fix this run's own tool needs — one small verifiable step, priced at ~15 lines:** make `attribution-gap`'s covering set and its touch set agree on the **same** timestamp field (either filter both client-side on the author-day key that is actually compared, or widen the git `--since` by one day of slack and *say so in the help text*), so that **bare `tools/attribution-gap` stops printing six gaps on a tree `--all` calls clean** — and land **section F's missing case with it** (a fixture whose touch and whose covering shift commit fall on the **same calendar day** either side of the cutoff, which is the middle of the window F1/F2 do not cover), plus the live assertion that `--all` and the default now agree on `/data/repo`. **After that:** run 952's **(i)** re-queued **only on a measured future-dated marker** (332 scanned, 0 found this run) and **(iii)** **only on a measured A12 misread** (the live row already names the units and the `work/ownlog/other` split), run 953's dispatcher-footer-coverage observation, **(93)**/**(97)(a)** first if the investor answers `hiring/queue/*.json` (**2 proposals**, unchanged since 2026-10-04 / 2026-09-30), **(2'')** the day the `GETTING_STARTED.md` keep/fold/retire answer arrives, **§14 renewal** cadence NEEDS-INVESTOR, the LAN-only A/B/C, NUL-blob watch on a measured harm only, **(av)(d)** after the `REPORT.md` §14 sign-off, **(i)** until the prune lands, SOA MNAME and Pip's mail-policy research NEEDS-INVESTOR, **(113)**'s two `cp`s operator-blocked, **2026-10-20** §10 #6 re-read, the `/stats` display row, and **(ax)**, **(e)**, the drift-verify line, **(aw)**, the `promote-gates` verdict watch and run 915's monitor item (closed as A31/A32, now waiting on `b460155`'s deploy) — with the honest state: every remaining line is gated (investor), dated, operator-blocked, NEEDS-INVESTOR, unit-owned, another desk's, or the priced queue above, so the next run takes **the `attribution-gap` window fix** (small, hermetic, and it reddens on the very command this entry quoted) or a standing line it can measure without inventing work. **This suite's (`tests/test_registry_coverage.sh`) own figure is 497 — the run that next edits it owns that refresh**, and `tests/test_attribution_gap.sh` is **58** on the same rule. **Suite-count claim: 114** (re-measured three ways this run), so **the run that lands the 115th `test_*` file owns the next `- Live:` refresh**. **A12/A30 are `identity-run@mia.service` + `identity-run@sofia.service`** (`Model unavailable: opencode/ling-3.1-flash-free`, class `other-error`), cleared only by those units' own next fire, **never** by a `reset-failed` from this seat.
19477:- **Queue — next small step (read this first):** **STEP 0 first every run** (today **0 unread dev / 0 prod**, `INBOX.md` **80/80/0 open** — an empty inbox is answered by the measurement). **This run closed `[0.4.349]`'s first priced item** (the window fix + section F's missing case + the live `--all`-vs-default agreement assertion), and in doing so **refused the false pass it exposed** (unreadable history → exit 3 in both modes). **NEW, re-queued unnumbered: the *hermetic* window is now day-granular, but the live default's `--since '7 days ago'` is still resolved by git's approxidate at call time** — so `window_open_day` can still shift across a midnight boundary *between* two runs of the same command (not a defect: both sets move together, and J5's subset assertion holds either way), and the honest next measurement is a **J5-shaped live assertion run twice ≥ 12 h apart** to confirm the day key moves without manufacturing a gap. **Also still open, unchanged:** run 952's **(i)** future-dated-footer guard (re-queued **only on a measured future-dated marker** — 332 scanned, 0 found) and **(iii)** run 951's `system-status` A12 wording (**only on a measured A12 misread** — the live row already names mia + sofia and the `work/ownlog/other` split); run 953's dispatcher-footer-coverage observation; **(93)**/**(97)(a)** first if the investor answers `hiring/queue/*.json` (**2 proposals**, unchanged since 2026-10-04 / 2026-09-30); **(2'')** the day the `GETTING_STARTED.md` keep/fold/retire answer arrives; **§14 renewal** cadence NEEDS-INVESTOR; the LAN-only A/B/C; NUL-blob watch on a measured harm only; **(av)(d)** after the `REPORT.md` §14 sign-off; **(i)** until the prune lands; SOA MNAME and Pip's mail-policy research NEEDS-INVESTOR; **(113)**'s two `cp`s operator-blocked; **2026-10-20** §10 #6 re-read; the `/stats` display row; and **(ax)**, **(e)**, the drift-verify line, **(aw)**, the `promote-gates` verdict watch, and run 915's monitor item — **A31/A32 now waiting on `b460155`'s `src/php/download.php` reaching dev and prod** (the dev-sync desk's, from `7f3166e`; this run promoted nothing). **V14b** (`shift-deploy` 132 vs 127) stays a **standing environmental red** owned by the stale-verdict gate, disclosed above.
> build · mimo-v2.6-flash-free
Now appending the CHANGELOG entry:
$ grep -c '^- \[' CHANGELOG.md; grep -n '^## \[0.4.350\]' CHANGELOG.md; ./tools/shift-tally --help | sed -n '1,60p' | grep -n "per-transcript" 
0
10759:## [0.4.350] - 2026-10-14 — run 960 takes run 959's priced queue: `attribution-gap`'s window now opens on a DAY boundary, so the default reading and `--all` agree by construction instead of by where the clock sits inside a day — suite **58 → 80**, and the false pass the new case exposed (an unreadable history reported as `no attribution gap`) is refused too
23:per-transcript: the human renderer also lists every transcript's
> build · mimo-v2.6-flash-free
$ cat >> CHANGELOG.md <<'EOF'

## [0.4.351] - 2026-10-14 — run 961 lands run 953's queued observation: `tools/shift-tally` prints each transcript's anchored footers BESIDE that transcript's own shift lines (JSON `transcripts`, human per-transcript table), because the corpus line's sum hid dispatcher's one footer against its peers' hundreds — suite **203 → 232**, no new flag

### Added
- **The per-transcript breakdown, the cheapest durable form of run 953's queued observation.** The corpus line has always **summed** five counters over every transcript (`1663 shift line(s), 1211 anchored, 452 not-anchored, 0 bad-ts, 0 future-ts`), and a sum hides its own outlier: measured on the live tree before this change, **dispatcher's transcript holds ONE `===== shift ` line in total** next to lena's 360 and leon's 336 — while its prose carries **11 `## … Shift` sections** since 2026-10-11 and git holds **60 `identity dispatcher shift` commits**. Nothing in the summed figure could show that. So `shift-tally` now emits, beside the corpus line, one row per transcript carrying its own `shift_lines` / `anchored` / `not_anchored` / `bad_ts` / `future_ts` — **anchored next to that transcript's own shift lines**, which is the comparison the sum destroys (213 of dispatcher's peers' footers are anchored; dispatcher is 1/1).
- **Two channels, one shape.** JSON: a `transcripts` object keyed by identity name, inserted **between `identities` and `errors`** (the documented 21-key set, `tests/test_shift_tally.sh` E1/E2 pinned to the exact order). Human: a labelled table after the corpus line — `per-transcript footers (anchored beside that transcript's own shift lines):` with its own `shift-lines` header — sorted by shift lines **descending, thin end first**, so the outlier a reader needs is the last row, not a hunt. The `--help` epilog gained a `per-transcript:` block explaining the two channels; **no `--` token was added anywhere** (see Notes).
- **The rows are raw per-file counters, so they do not move with `--dedup`.** De-duplication changes what is *tallied*, never which lines *exist*: the fixture's `transcripts` object is byte-identical at `--dedup file` and `--dedup global` (P6), which is why the breakdown can be shown once per level without the reader wondering which level it came from. The rows sum back to the corpus counters exactly — nothing unattributed (P5/P16) — and a transcript with **no** shift line at all still gets an all-zero row (P12/P13), because a silent absence and a zero would look identical otherwise.
- **Live reading (2026-10-14, run 961), both levels agreeing:** lena **473 / 360** (shift lines / anchored), leon **424 / 336**, mia **359 / 235**, sofia **269 / 153**, aylin **128 / 120**, jonas **9 / 6**, dispatcher **1 / 1**, against a corpus of **1,663 / 1,211 / 452 / 0 / 0**, window `2026-09-23T20:11:29+02:00 .. 2026-10-14T19:05:00+02:00` — the table above renders identical at `dedup=file` and `dedup=global`, as the raw-counter rule requires.

### Tests
- **`bash tests/test_shift_tally.sh` → 232 passed / 0 failed / 0 skipped, rc 0** (was **203**; +29), the new **section P (P0–P17)** and **M16/M17**: P1 exactly the two fixture transcripts (the hyphenated report still excluded); **P2/P3 the row VALUES pinned, not just the key's presence** (`alpha 6 4 2 1 1`, `beta 2 2 0 0 0`); **P4 the row invariant** (`anchored <= shift_lines` per transcript); **P5 the rows sum to the corpus** (`8 6 2 1 1` — nothing unattributed); **P6 the breakdown is identical at `--dedup global`**; **P7a/P7b a missing dir and the report-only refusal fabricate no rows** (`{}`); **P8/P8b the table and its own `shift-lines` header render**; **P9/P10 the rendered rows** matched by their **five numeric fields** so the identity table's `alpha 2 1 1 50.00%` row — which also names alpha — cannot satisfy the regex (the suite's first draft matched on the identity name and was corrected to `{4}` numeric repetitions before any figure was quoted); **P11 descending order** (alpha 6 before beta 2); **P12/P13 the footer-less transcript gets an all-zero row under its own name**; **P14–P17 the live box, race-safe by shape** — one row per live transcript (`len(transcripts) == files`), rows summing to the live corpus (`1663/1211`), and the row invariant on every live transcript (never a count of the moving tree).
- **M16/M17, each precondition-asserted to occur exactly once, each must RUN and diverge, each surgical:** **M16** drops the store line (`entry["transcripts"][identity] = dict(counters)` → `pass`) — the mutant *is* the pre-`[0.4.351]` tool: `transcripts == {}` while the corpus still reads `8 6`, and the real tool's object is asserted **not** empty (so the assertion is reading the breakdown, not a constant); **M17** swaps the row's two columns (`shift_lines` ↔ `anchored`) — caught by the row's own invariant breaking (`False`) with the corpus path untouched (`8 6 2 1 1`), i.e. a green P2/P9 cannot be a row nobody read.
- **Cross-suite, figures deliberately NOT moved:** no `##` heading added, renamed or moved; `- Live:` suite figure **114** unchanged (`tools/regression-run --list` → 114, no `test_*` file added); `tests/test_registry_coverage.sh` C28g **30** and G38 **27** unchanged because **no option was added** — the option surface is still `--format {json,human}` + `--dedup {file,global,both}`; the REGISTRY `## shift-tally` section gained the `transcripts` key and the breakdown paragraph and **no `**Options**` row**.

### Notes
- **No `app/` file edited** → no reviewer gate, nothing to promote; dev and prod untouched at **0.4.28** (the investor's 2026-10-05 correction stands, the wrong figure quoted nowhere). STEP 0 paid first and owed nothing (0 unread `investor_to_agent` in **both** databases, `inbox-status` rc 0 *"nothing owed"*, last reply dev 146 / prod 111, `INBOX.md` **80 / 80 / 0 open**), so no `agent_to_investor` row was written. No live messages-DB write, no DNS write, no mail sent, no paid API, no API key configured; `hiring/queue/` untouched (**2 proposals `status=proposed`**); no unit or timer change, no `identity-run.sh` edit; `tests/test_registry_coverage.sh` untouched. The tool stayed **read-only by construction** (no write, no truncate, no unlink, no network, no child process — it never writes a transcript). Spend **0.00**, free `*-free` models only.
- **Why a JSON key and not a flag, recorded as the decision it was:** every new `--` token moves `tests/test_registry_coverage.sh` C28g (30) and G38 (27) and the REGISTRY `- Live:`/option figures with it, for a breakdown that needs no caller choice — the corpus counters were already computed per file, so the row is free at the same I/O. The env hook `GLADEX_SHIFT_TALLY_DIR` already existed for the fixture, so section P reaches the same code path the live read does.

### Queue
Authoritative next-candidate queue lives in `agent-logs/PROGRESS.md`, in the newest run entry's `Queue — next small step` bullet — it is not duplicated here, and has not been since `[0.4.67]`. To read the live queue, open `PROGRESS.md`; to know whether that is still true, run `tools/queue-source-check`.
EOF
grep -c '^## \[' CHANGELOG.md; tail -3 CHANGELOG.md
356

### Queue
Authoritative next-candidate queue lives in `agent-logs/PROGRESS.md`, in the newest run entry's `Queue — next small step` bullet — it is not duplicated here, and has not been since `[0.4.67]`. To read the live queue, open `PROGRESS.md`; to know whether that is still true, run `tools/queue-source-check`.
> build · mimo-v2.6-flash-free
$ sqlite3 /opt/startup/dev/data/messages.db "SELECT COUNT(*) FROM messages WHERE direction='investor_to_agent' AND read=0;" && sqlite3 /opt/startup/prod/data/messages.db "SELECT COUNT(*) FROM messages WHERE direction='investor_to_agent' AND read=0;"; ./tools/inbox-status; echo "rc=$?"
0
0
inbox-status — STEP 0 investor duty

  messages unread : dev 0 / prod 0 / total 0
  last reply      : dev 146 / prod 111
  INBOX.md        : 80 entries, 80 handled, 0 open

verdict: OK - nothing owed (0 unread, 0 open entries all replied)
rc=0
$ ./tools/version-check; ./tools/budget-show; ./tools/queue-source-check; echo "qsc rc=$?"; grep -c '~~HANDLED~~' INBOX.md; grep -c '^## ' INBOX.md
OK: dev=0.4.28 prod=0.4.28 (match)
=== Budget Status ===
Month:       2026-10
Allowance:   5.00 €
Spent:       0.00 € (no October charge in any file — MEASURED 2026-10-12: no public registry source can date a renewal (DENIC RDAP publishes only a "last changed" event, and does so for every .de domain; DENIC whois has no expiry field; the IANA RDAP bootstrap has no "de" entry), so the one open unknown is your billing cadence: monthly 1.50 as brief §4 states, or annual at the 2027-09-20 registration anniversary? Still NEEDS-INVESTOR, REPORT.md §14)
Remaining:   5.00 €

Log:
  - 2026-09-20: init, no domains ordered yet
  - 2026-09-20: investor APPROVED gladex.de (1.50/mo). Awaiting registrar purchase + DNS creds.
  - 2026-09-20: purchase completed same day — RDAP status active (changed 2026-09-20T11:10:57+02:00), DNS wired via pdns wrapper (A apex/dev/git/info/stats/log/team -> 77.90.15.49, MX 10 gladex.de), dig-verified; DOMAIN.md holds evidence.
  - 2026-09-24: ledger narrative corrected — spent-line still said "purchase pending at registrar" 4 days after registration (queued candidate #1); numbers unchanged (1.50/3.50, no money moved), log history above kept verbatim.
  - 2026-10-01: month rolled per brief §4 (balance resets to 5.00 on the 1st, unspent does NOT roll over). September history above is kept verbatim; September closed at spent 1.50 / remaining 3.50.
  - 2026-10-02: October opened at spent 0.00 / remaining 5.00 because no October charge exists in any file on this box — while gladex.de is a known 1.50/mo recurring cost with NO renewal date recorded anywhere. Figure to be corrected the moment the investor answers REPORT.md §14.
  - 2026-10-12: renewal date MEASURED instead of left open, figure unchanged (0.00 / 5.00, no money moved): (1) DENIC RDAP rdap.denic.de/domain/gladex.de returns exactly ONE event — "last changed 2026-09-20T11:10:57+02:00", i.e. the purchase day already on file — and NO registration/expiration event; (2) controls denic.de ("last changed 2024-12-19T13:33:43+01:00") and bund.de ("last changed 2026-06-17T13:11:54+02:00") return the same one-event shape, so this is how DENIC answers for EVERY .de domain, not a gap in ours; (3) DENIC whois (port 43) prints only "Domain: gladex.de" + "Status: connect" — no expiry field; (4) the IANA RDAP bootstrap (data.iana.org/rdap/dns.json, 592 services) contains no "de" entry and no "denic" service, so the standard bootstrap path rdap.org/domain/gladex.de answers 404 (control: example.com answers 302). Conclusion: NO public registry source can date a renewal — only the investor's registrar account can, so REPORT.md §14 stays open with its question narrowed to the billing cadence. Registry-side, the only anchor on file is registration 2026-09-20 (this log line of 2026-09-20, corroborated by the RDAP timestamp), whose first anniversary falls 2027-09-20 — so no registry-side event lands in October 2026.
queue-source-check: FAIL - 1 violation(s): [0.4.351] is not named in agent-logs/PROGRESS.md - the newest CHANGELOG entry has no record in the authoritative file
  - [0.4.351] is not named in agent-logs/PROGRESS.md - the newest CHANGELOG entry has no record in the authoritative file
qsc rc=1
0
80
> build · mimo-v2.6-flash-free
$ tail -6 /tmp/opencode/run1138/regcov.log 2>/dev/null; echo "=== leak/figread ==="; cat /tmp/opencode/run1138/leak.log 2>/dev/null | tail -4; tail -5 /tmp/opencode/run1138/figread.log 2>/dev/null
PASS: C16d plant: 999 is not the discovered count (or this is not a defect)
PASS: C16e a stale live figure -> exit 1
PASS: C16f VIOLATION figure_stale 999
PASS: C16g exactly one violation - the plant, and nothing else
PASS: C17a plant: the undated figure line is present
PASS: C17b plant: and THAT line carries no date
=== leak/figread ===

===============================
leak-figure-readers suite: 41 passed, 0 failed
===============================
PASS: V15a immich-roundtrip: figure 160 and suite tests/test_immich_roundtrip.sh both read off the section
PASS: V15b immich-roundtrip: the registry's figure equals what tests/test_immich_roundtrip.sh produced (child rc=0)
PASS: V16a regression-run: figure 350 and suite tests/test_regression_run.sh both read off the section
PASS: V16b regression-run: the registry's figure equals what tests/test_regression_run.sh produced (child rc=0)
PASS: V17a system-status: figure 26 and suite tests/test_system_status_unread.sh both read off the section
> build · mimo-v2.6-flash-free

Select another run log from the list above. Only files matching run-YYYYMMDD-HHMMSS-N.log are readable.

[Wed Oct 14 21:34:52 2026] 127.0.0.1:60106 Accepted
[Wed Oct 14 21:34:52 2026] 127.0.0.1:60106 Closing
[Wed Oct 14 21:35:00 2026] 127.0.0.1:34104 Accepted
[Wed Oct 14 21:35:00 2026] 127.0.0.1:34104 Closing
[Wed Oct 14 21:35:01 2026] 127.0.0.1:34106 Accepted
[Wed Oct 14 21:35:01 2026] 127.0.0.1:34106 Closing
[Wed Oct 14 21:35:03 2026] 127.0.0.1:34118 Accepted
[Wed Oct 14 21:35:03 2026] 127.0.0.1:34118 Closing
[Wed Oct 14 21:35:05 2026] 127.0.0.1:34054 Accepted
[Wed Oct 14 21:35:05 2026] 127.0.0.1:34054 Closing
[Wed Oct 14 21:45:24 2026] 127.0.0.1:57566 Accepted
[Wed Oct 14 21:45:24 2026] 127.0.0.1:57566 Closing
[Wed Oct 14 21:45:24 2026] 127.0.0.1:57570 Accepted
[Wed Oct 14 21:45:24 2026] 127.0.0.1:57570 Closing
[Wed Oct 14 21:45:24 2026] 127.0.0.1:57580 Accepted
[Wed Oct 14 21:45:24 2026] 127.0.0.1:57580 Closing
[Wed Oct 14 21:45:24 2026] 127.0.0.1:57592 Accepted
[Wed Oct 14 21:45:24 2026] 127.0.0.1:57592 Closing
[Wed Oct 14 21:45:24 2026] 127.0.0.1:57606 Accepted
[Wed Oct 14 21:45:24 2026] 127.0.0.1:57606 Closing
[Wed Oct 14 21:45:24 2026] 127.0.0.1:57622 Accepted
[Wed Oct 14 21:45:24 2026] 127.0.0.1:57622 Closing
[Wed Oct 14 21:45:24 2026] 127.0.0.1:33282 Accepted
[Wed Oct 14 21:45:25 2026] 127.0.0.1:33282 Closing
[Wed Oct 14 21:45:25 2026] 127.0.0.1:33284 Accepted
[Wed Oct 14 21:45:25 2026] 127.0.0.1:33284 Closing
[Wed Oct 14 21:45:25 2026] 127.0.0.1:33292 Accepted
[Wed Oct 14 21:45:25 2026] 127.0.0.1:33292 Closing
[Wed Oct 14 21:45:25 2026] 127.0.0.1:33296 Accepted
[Wed Oct 14 21:45:25 2026] 127.0.0.1:33296 Closing
[Wed Oct 14 21:46:36 2026] 127.0.0.1:38566 Accepted
[Wed Oct 14 21:46:36 2026] 127.0.0.1:38566 Closing
[Wed Oct 14 21:46:36 2026] 127.0.0.1:38576 Accepted
[Wed Oct 14 21:46:36 2026] 127.0.0.1:38576 Closing
[Wed Oct 14 21:46:36 2026] 127.0.0.1:38584 Accepted
[Wed Oct 14 21:46:36 2026] 127.0.0.1:38584 Closing
[Wed Oct 14 21:46:36 2026] 127.0.0.1:38598 Accepted
[Wed Oct 14 21:46:36 2026] 127.0.0.1:38598 Closing
[Wed Oct 14 21:46:36 2026] 127.0.0.1:38614 Accepted
[Wed Oct 14 21:46:36 2026] 127.0.0.1:38614 Closing
[Wed Oct 14 21:46:36 2026] 127.0.0.1:38628 Accepted
[Wed Oct 14 21:46:36 2026] 127.0.0.1:38628 Closing
[Wed Oct 14 21:46:36 2026] 127.0.0.1:38640 Accepted
[Wed Oct 14 21:46:36 2026] 127.0.0.1:38640 Closing
[Wed Oct 14 21:46:36 2026] 127.0.0.1:38654 Accepted
[Wed Oct 14 21:46:36 2026] 127.0.0.1:38654 Closing
[Wed Oct 14 21:46:36 2026] 127.0.0.1:38660 Accepted
[Wed Oct 14 21:46:37 2026] 127.0.0.1:38660 Closing
[Wed Oct 14 21:46:37 2026] 127.0.0.1:38670 Accepted
[Wed Oct 14 21:46:37 2026] 127.0.0.1:38670 Closing
[Wed Oct 14 21:46:44 2026] 127.0.0.1:38678 Accepted
[Wed Oct 14 21:46:44 2026] 127.0.0.1:38678 Closing
[Wed Oct 14 21:46:44 2026] 127.0.0.1:38692 Accepted
[Wed Oct 14 21:46:44 2026] 127.0.0.1:38692 Closing
[Wed Oct 14 21:47:39 2026] 127.0.0.1:38120 Accepted
[Wed Oct 14 21:47:39 2026] 127.0.0.1:38120 Closing
[Wed Oct 14 21:48:29 2026] 127.0.0.1:39550 Accepted
[Wed Oct 14 21:48:29 2026] 127.0.0.1:39550 Closing
[Wed Oct 14 21:48:29 2026] 127.0.0.1:39564 Accepted

Other app logs: downloads.log