prompt: retire two already-delivered PyAutoBrain bug prompts

complete/2026/07/version-drift-stamp-consistency.md

@@ -0,0 +1,123 @@
+- issue: none (the prompt was never issued — the fix shipped inside a `/wake_up` session's branch)
+- delivered-by: **PyAutoBrain#144** — commit `e26ab17` *"fix: version_drift checks stamp consistency, not equality-to-release-tag"* (2026-07-19, branch `claude/wake-up-u53v8z`)
+- classification: library (PyAutoBrain) — bug
+- LEDGER BACKFILL, not new work: this record was written 2026-08-08 when the prompt was
+  picked up for development and found already delivered. No code was written for it.
+  PyAutoBrain#144 had **no** completion record until this one, so the ledger carried a gap.
+- prompt Status was stale: it read `fix-implemented-on-branch (claude/wake-up-u53v8z) —
+  pending review/merge`. That branch merged 2026-07-19; the status was never advanced,
+  which is what made the prompt look open.
+
+## What was wrong
+
+`bin/version_drift.sh` compared each library/workspace's committed source version stamp
+(`*/__init__.py` `__version__`, workspace `version.txt`) against the latest PyAutoLens
+**release tag** and flagged any mismatch as drift. The release design deliberately
+abandoned that invariant: `release.yml` stamps `__version__` into the wheel build tree
+only and does **not** commit it back, because daily "Update version to X" commits to
+every library main were the noise engine behind the June/July 2026 accidental-release
+cascade (PyAutoBuild#118 / #120). pip users get the version from the stamped wheel;
+source checkouts stay frozen.
+
+So the check reported drift after **every** release, indefinitely — pure `/wake_up`
+noise, and misleading, since it read as a release defect when the release was fine.
+
+## What shipped
+
+**Option 2 of the three the prompt offered** — honour the freeze, flag only
+*non-uniform* stamps:
+
+- The invariant is now "every coupled repo carries the SAME stamp as its siblings";
+  a repo out of step with the consensus is flagged. The latest release tag is shown
+  for context only, with a consensus trailing the tag reported as the expected freeze
+  rather than drift.
+- `gh` became optional — the reference tag is informational, so the consensus check
+  runs on local stamps without it, which is what makes the script usable in web/CI
+  sessions.
+- Fixed a stale stamp path left by the organ rename: `PyAutoConf/autoconf` →
+  `PyAutoNerves/autonerves`.
+- Both documentation updates the prompt required: the `version_drift.sh` header comment
+  (which still described the tag-equality model) and the `/wake_up` skill step-5 wording.
+
+Verified in that PR: uniform frozen stamps read clean; a single out-of-step stamp is
+still flagged — i.e. the reframe did not simply disable the check.
+
+## Verification done before retiring the prompt (2026-08-08)
+
+- `git merge-base --is-ancestor e26ab17 origin/main` → true, so the fix is on main.
+- Read the shipped `bin/version_drift.sh`: header now documents the freeze and the
+  consensus invariant; no `== latest release tag` comparison survives.
+- Read `skills/wake_up/wake_up.md` step 5: reworded to "version-stamp *consistency
+  across the coupled libs + workspaces* … the latest release tag is shown for context
+  only". Both doc requirements satisfied.
+
+## Note
+
+Option 3 (retire the script in favour of Heart's `version_skew` check) was not taken
+and remains a live option — `version_drift.sh` still exists. Nothing depends on
+revisiting it; recorded only so a future reader knows it was considered and left open
+rather than overlooked.
+
+## Original prompt
+
+# version_drift.sh reports permanent false drift — asserts an abandoned invariant
+
+Type: bug
+Target: pyautobrain
+Repos:
+- PyAutoBrain
+Difficulty: small
+Autonomy: safe
+Priority: normal
+Status: fix-implemented-on-branch (claude/wake-up-u53v8z, PyAutoBrain) — pending review/merge
+
+`PyAutoBrain/bin/version_drift.sh` compares each library/workspace's committed
+source version stamp (`*/__init__.py` `__version__`, workspace `version.txt`)
+against the latest PyAutoLens **release tag**, and flags any mismatch as drift.
+That invariant (committed source stamp == latest release tag) was **deliberately
+abandoned** by the release design.
+
+Evidence (2026-07-19 wake-up): reference tag `2026.7.19.1` (published 13:52),
+but every stamp — PyAutoNerves / PyAutoArray / PyAutoFit / PyAutoGalaxy /
+PyAutoLens and the autofit/autogalaxy/autolens workspace `version.txt` — reads
+`2026.7.9.1`. Verified against GitHub `main` **and** the `2026.7.19.1` tag
+commit (identical blob SHA `19b7d518`): the source stamp is frozen there too, so
+this is not container/mirror staleness — the stamp genuinely is not bumped on
+release.
+
+Root cause (working as intended, NOT a release bug): `PyAutoHands/.github/
+workflows/release.yml` (step "Stamp version in build tree", ~L394-403) seds
+`__version__` into the **wheel build tree only** and does not commit it back.
+The inline comment is explicit: *"the stamp is NOT committed back to the library
+repo. Daily 'Update version to X' commits to every library main were the
+stale-CI/noise engine behind the June/July 2026 accidental-release cascade
+(PyAutoBuild#118 / #120); pip users get __version__ from the stamped wheel,
+source checkouts use PYAUTO_SKIP_WORKSPACE_VERSION_CHECK."* So PyPI wheels are
+correctly versioned; the committed source stamp is intentionally frozen.
+
+Consequently `version_drift.sh` will report drift after **every** release,
+indefinitely (the source stamp is never advanced again) — pure noise in the
+`/wake_up` digest, and misleading: it reads as a release defect when the release
+is fine.
+
+Fix options (pick during triage):
+1. **Re-point the reference to what pip users actually get** — compare the
+   published PyPI / wheel version per package against the latest release tag
+   (the meaningful post-cascade invariant), not the deliberately-frozen source
+   stamp.
+2. **Honor the documented skip semantics** — treat a uniformly-frozen source
+   stamp as expected (mirror `PYAUTO_SKIP_WORKSPACE_VERSION_CHECK`), only
+   flagging *non-uniform* stamps (one repo out of step with the others), which
+   is the real drift worth catching.
+3. **Retire the script** if Heart's `version_skew` check (pinned-dep skew,
+   already green) covers the remaining need.
+
+Whichever is chosen, update the `version_drift.sh` header comment (it still
+describes the tag-equality model) and the `/wake_up` skill step 5 wording.
+
+Note: this is the same class of defect as the origin drift-check false-positive
+fixed on the `claude/wake-up-u53v8z` branch (repos_sync.py `normalize_remote`
+made host-agnostic) — a wake-up check encoding an invariant that reality no
+longer holds. Found during the 2026-07-19 `/wake_up` version-pin investigation.
+
+<!-- formalised from a /wake_up investigation, 2026-07-19 -->

complete/2026/08/hygiene-gone-ref-counter.md

@@ -0,0 +1,138 @@
+- issue: none (the prompt was never issued — the defects were fixed under other issues, see below)
+- delivered-by: **PyAutoBrain#205** (squashed `7ad1e43`, 2026-08-07) and **PyAutoBrain#200** (squashed `5cb1c73`, 2026-08-05, issue PyAutoBrain#197)
+- classification: library (PyAutoBrain) — bug, infrastructure
+- LEDGER BACKFILL, not new work: this record was written 2026-08-08 when the prompt
+  was picked up for development and found already delivered. No code was written for
+  it. Defect 2's delivery is separately recorded in
+  `complete/2026/08/hygiene-coverage-drift.md`; Defect 1 (#205) had **no** record
+  until this one, so the ledger carried a gap.
+
+## What the prompt claimed
+
+`hygiene tidy` under-reported git debris by roughly an order of magnitude — it
+reported "12 stale branches across 9 repos, 0 stashes, 0 [gone] refs, 0 dirty
+checkouts" when the real state was 91 branches across 28 repos, 4 stashes, dozens of
+`[gone]` refs and 3 dirty checkouts. Two independent defects.
+
+## Both defects were already fixed
+
+**Defect 1 — the `[gone]` counter could never be non-zero.** `prescan_tidy` counted
+with `git branch -vv | grep -c '\[gone\]'`, but porcelain prints the upstream as
+`[origin/<branch>: gone]`, never the bare `[gone]`, so the pattern matched nothing in
+every repo, always. Fixed by **PyAutoBrain#205** (2026-08-07), which switched to
+`for-each-ref --format='%(upstream:track)'` — the idiom
+`enumerate_condemn_candidates` and `repo_cleanup` already used — and added a
+regression test building a fixture with a genuinely gone upstream, verified to fail on
+the unfixed counter.
+
+**Defect 2 — the scan covered 9 repos, not the organism.** `prescan_tidy` iterated
+hardcoded `LIB_REPOS` + `ORG_REPOS` literals, so PyAutoCTI, PyAutoReduce, PyAutoMemory,
+every workspace, every HowTo and every assistant were invisible; the "0 stashes" and
+"0 dirty checkouts" figures were scope artifacts, not detector bugs. Fixed by
+**PyAutoBrain#200** (2026-08-05): `LIB_REPOS`/`ORG_REPOS`/`WS_REPOS` are now derived
+from the body map via `_hygiene_repos.py` reading `repos.yaml`, with a
+`repos_sync.py` coverage check that fails if they drift from the map or if a repo name
+is written back into an array literal. That is exactly the resolution the prompt asked
+someone to decide on, and it names the same `_hygiene_extras.py`/PR#193 precedent.
+
+## Verification done before retiring the prompt (2026-08-08)
+
+The prompt specified its own control: *"Before the fix, `prescan_tidy` must reproduce
+`0 [gone] refs` on a checkout known to have some. After, the count must be non-zero on
+that same tree. A green run on a tree with zero `[gone]` refs proves nothing — pick the
+tree first."* That control was run rather than assumed:
+
+- Built a throwaway repo with a real bare remote, pushed a branch, deleted it on the
+  remote, and `fetch --prune`d so the checkout genuinely carried a `[gone]` ref.
+- On that same tree: the old form `git branch -vv | grep -c '\[gone\]'` → **0**;
+  the shipped form `for-each-ref '%(upstream:track)' | grep -c '\[gone\]'` → **1**.
+- Confirmed no repo-name array literal survives anywhere in `hygiene.sh` — every loop
+  runs over the derived `CODE_REPOS`/`SCAN_REPOS`, including `run_tidy` and
+  `enumerate_condemn_candidates`, the two the prompt named as sharing the limitation.
+
+## Trap worth keeping
+
+Establishing "already fixed" required deepening the clone. Cloud-session checkouts are
+**shallow** (`git rev-parse --is-shallow-repository` → true), and a shallow PyAutoBrain
+makes `git diff origin/main...<branch>` fail with `no merge base` and makes
+`git log origin/main..<branch>` list commits that ARE on main. Both look like evidence
+of divergence and are artifacts of the graft boundary. `git fetch --deepen=500` first,
+then test ancestry with `git merge-base --is-ancestor`.
+
+## Original prompt
+
+# The hygiene tidy pre-scan reports 0 [gone] refs unconditionally, and scans only 9 of ~28 repos
+
+Type: bug
+Target: PyAutoBrain
+Repos:
+- PyAutoBrain
+Difficulty: small
+Autonomy: supervised
+Priority: normal
+Status: formalised
+
+Found during a `/repo_cleanup` sweep on 2026-08-04. `hygiene tidy` reported
+"12 stale branches across 9 repos, 0 stashes, 0 [gone] refs, 0 dirty checkouts".
+The real state at that moment was 91 branches across 28 repos, 4 stashes,
+dozens of [gone] refs and 3 dirty checkouts.
+
+Two independent defects produce that gap.
+
+## Defect 1 — the `[gone]` counter can never be non-zero
+
+`agents/conductors/hygiene/hygiene.sh:129`:
+
+    g=$(git -C "$dir" branch -vv 2>/dev/null | grep -c '\[gone\]' || true)
+
+`git branch -vv` prints the upstream as `[origin/<branch>: gone]`, never the
+bare `[gone]`. The pattern matches nothing, in every repo, always. Proof on
+PyAutoHands, which had three at the time:
+
+    git branch -vv | grep -c '\[gone\]'                        -> 0
+    for-each-ref --format='%(upstream:track)' refs/heads       -> [gone] x3
+
+`prescan_tidy` then folds `gone` into the `total` at line 133-134, so the
+conductor's prioritisable count for `tidy` is systematically understated and
+`tidy` under-ranks itself against the other hygiene modes.
+
+Fix: count via `for-each-ref '%(upstream:track)'` (already the idiom used in
+`repo_cleanup`'s own audit), or match `: gone]`.
+
+Note `enumerate_condemn_candidates` at line 419 does NOT share this bug — it
+uses `for-each-ref` correctly.
+
+## Defect 2 — the scan covers 9 repos, not the organism
+
+`prescan_tidy` (line 121) iterates `LIB_REPOS` + `ORG_REPOS` only:
+
+    LIB_REPOS=(PyAutoNerves PyAutoFit PyAutoArray PyAutoGalaxy PyAutoLens)
+    ORG_REPOS=(PyAutoBrain PyAutoHands PyAutoHeart PyAutoMind)
+
+That is the "9 repos" in the summary line. Invisible to it: PyAutoCTI,
+PyAutoReduce, PyAutoMemory, PyAutoGut, PyAutoScientist, every workspace, every
+HowTo, every assistant, autolens_profiling, admin_jammy. Both the "0 stashes"
+and "0 dirty checkouts" figures were scope artifacts, not detector bugs — all 4
+stashes and all 3 dirty trees lived in repos it never looks at.
+
+Decide whether the fix is to derive the repo set from `PyAutoMind/repos.yaml`
+(the body map) rather than re-listing it, which is the same hard-coding problem
+`_hygiene_extras.py` was just refactored to remove in PR #193. `run_tidy` /
+`enumerate_condemn_candidates` share the same two lists and the same limitation.
+
+## Why it matters
+
+`hygiene tidy` is the advertised front door for git debris and the thing a
+human reads before deciding whether a cleanup is worth running. Under-reporting
+by roughly an order of magnitude makes it read as "nothing much to do" when
+there is. A wrong-but-quiet number is worse than no number.
+
+## Control
+
+Before the fix, `prescan_tidy` must reproduce `0 [gone] refs` on a checkout
+known to have some (any repo with a merged-and-deleted upstream). After, the
+count must be non-zero on that same tree and match
+`for-each-ref | grep -c '\[gone\]'`. A green run on a tree with zero [gone] refs
+proves nothing — pick the tree first.
+
+<!-- raised from a /repo_cleanup sweep, 2026-08-04 -->

complete/index.md

@@ -6,7 +6,7 @@ Token-light navigation over the finished-work records (schema:
 only then grep a dated bucket. Curators: edit the band between the CURATED
 markers; everything below GENERATED is rebuilt.
 
-935 records across 7 buckets.
+937 records across 7 buckets.
 
 <!-- CURATED:START -->
 ## Highlights
@@ -34,6 +34,7 @@ _(curate hard-won records here — survives regeneration.)_
 - [hst-dq-bits-dial](2026/08/hst-dq-bits-dial.md)
 - [hygiene-coverage-drift](2026/08/hygiene-coverage-drift.md)
 - [hygiene-detail-flag](2026/08/hygiene-detail-flag.md)
+- [hygiene-gone-ref-counter](2026/08/hygiene-gone-ref-counter.md) — the prompt was never issued — the defects were fixed under other issues, see below
 - [hygiene-howto-refs-docstrings](2026/08/hygiene-howto-refs-docstrings.md)
 - [interferometer-delaunay-flaky-fitexception](2026/08/interferometer-delaunay-flaky-fitexception.md)
 - [interferometer-start-here-integrate-oom](2026/08/interferometer-start-here-integrate-oom.md)
@@ -545,6 +546,7 @@ _(curate hard-won records here — survives regeneration.)_
 - [validation-searches-env-optax](2026/07/validation-searches-env-optax.md)
 - [verify-install-check-f-autosimulate-and-dep-pin](2026/07/verify-install-check-f-autosimulate-and-dep-pin.md)
 - [version-check-compat-floor](2026/07/version-check-compat-floor.md)
+- [version-drift-stamp-consistency](2026/07/version-drift-stamp-consistency.md) — the prompt was never issued — the fix shipped inside a `/wake_up` session's branch
 - [version-pinning-design-review](2026/07/version-pinning-design-review.md)
 - [viz-refactor-asserts-1280](2026/07/viz-refactor-asserts-1280.md) — Two visualization sweep-assert failures, both diagnosed as env-config (NOT code bugs); workspace-only, no libr…
 - [viz-render-gallery](2026/07/viz-render-gallery.md) — Eyes agent Phase 1