lifecycle: add orphans, record the M1 rehearsal task, file the triage

complete/2026/06/build-testpypi-rehearsal-mode.md

@@ -0,0 +1,140 @@
+# build-testpypi-rehearsal-mode
+
+- shipped: 2026-06-30
+- pr: https://github.com/PyAutoLabs/PyAutoHands/pull/111
+- commit: `71936fb47c19fd6ba7b64fa9283bf7231e47b165`
+- repos:
+  - PyAutoHands (the prompt targets "PyAutoBuild", this repo's former name)
+- milestone: M1 — prerequisite for the Heart release-validation gate (M2)
+
+## RECORD RECONSTRUCTED AFTER THE FACT (2026-08-08)
+
+This record was **not** written at ship time by `ship_library`. The task shipped
+on 2026-06-30 and nothing in the Mind was updated: the prompt stayed in
+`active/`, and `planned.md` went on listing it as `status: planned` for six
+weeks. It was found during the registry-integrity audit
+(`draft/maintenance/pyautomind/registry_integrity_check.md`) and reconstructed
+from upstream git evidence, which is why it carries less detail than a record
+written by the ship path — there is no contemporaneous trap/finding log.
+
+Everything below is verified against PyAutoHands `main`, not inferred.
+
+## Summary
+
+Added a TestPyPI-only `rehearsal` mode to `release.yml` so the release pipeline
+can build and publish the current source to TestPyPI and **stop** — no PyPI
+upload, no git tag, no notebook generation, no Colab-URL bumps. That is what
+lets the Heart install and validate the actual wheels before any PyPI
+promotion, which is the only way packaging-layer bugs (bad `MANIFEST`, missing
+data file, direct-URL dependency) get caught.
+
+## Evidence it shipped
+
+In `PyAutoHands/.github/workflows/release.yml` on `main`:
+
+- a `rehearsal` `workflow_dispatch` input, described in-file as "the one mode
+  switch: false (default) = a full real release";
+- a `resolve_mode` job publishing `rehearsal` as a job output — the single
+  source of truth for rehearsal-vs-live;
+- a unique PEP 440 dev segment appended per rehearsal run, so repeated
+  dispatches never collide, plus `--skip-existing` on the upload;
+- the PyPI/tag/notebook jobs gated `if: needs.resolve_mode.outputs.rehearsal != 'true'`.
+
+The merge commit body states the intent directly: *"to unblock the Heart
+wheel-based release-validation gate (M2)"*, and notes it was "Validated live on
+run #645".
+
+It has been in production use since: the 2026-08-07 release drive ran Stage 2
+as rehearsal run 31192317261 → `testpypi 2026.8.7.1.dev70601`.
+
+## Scope delivered beyond the prompt
+
+The same PR carried a dispatch-surface cleanup the prompt did not ask for:
+removed the dead `skip_scripts` / `skip_notebooks` knobs, the redundant
+`skip_release` dry-run (rehearsal supersedes it) and the defunct
+`update_notebook_visualisations` job, leaving the dispatch surface as
+`minor_version + rehearsal`. It also added a pre-upload direct-reference-URL
+guard — the `[nss]` `git+` footgun named in the prompt's motivation — and
+renamed Pulse→Heart, PyAutoAgent→PyAutoBrain.
+
+## The milestone chain this belonged to
+
+M1 of four, all four now shipped and all four discovered stale in `planned.md`
+on the same day: M0 `heart-ci-linkage` (PyAutoHeart `heart/checks/ci_status.*`),
+M1 this task, M2 `heart-release-validation` (`pyauto-heart validate --ingest`),
+M3 `heart-release-profile-wheel-integration` (the named `release` profile in
+`heart/validate.py`). The other three had no prompt file to fold and were
+removed from `planned.md` rather than recorded.
+
+## Original prompt
+
+# Add a TestPyPI-only "rehearsal" mode to release.yml
+
+Type: feature
+Target: PyAutoBuild
+Repos:
+- PyAutoBuild
+Status: planned
+Difficulty: too-large
+Autonomy: supervised
+Priority: normal
+Milestone: M1 — prerequisite for `feature/pyautoheart/release_validation.md` (M2)
+
+## Why
+
+The Heart-owned release-validation pipeline (see
+`feature/pyautoheart/release_validation.md`) needs to test the organism against
+**built wheels**, not source checkouts — that is the only thing that catches
+packaging-layer bugs (a bad `MANIFEST`, a missing data file, a too-loose or
+direct-URL dependency). Two real incidents in `complete.md` motivate this: the
+PyAutoFit `[nss]` `git+` direct URL silently broke every TestPyPI upload for
+weeks, and the nufftax/JAX dependency-floor mismatch produced broken installs —
+neither was caught by the source-based validation.
+
+`release.yml` already knows how to build and publish to TestPyPI (its
+`release_test_pypi` job), but that capability is **coupled to the full
+release** (TestPyPI → PyPI → tag → notebook commits in one flow). We need to be
+able to build + publish the current source to TestPyPI **and stop**, so Heart
+can install and validate those wheels before any PyPI promotion.
+
+## Task
+
+Add a `rehearsal` (TestPyPI-only) execution mode to `release.yml`:
+
+- A new `workflow_dispatch` input (e.g. `rehearsal: true`, or reuse/extend the
+  existing skip flags) that:
+  - builds every package from current source and **publishes to TestPyPI**,
+  - then **STOPS** — no PyPI upload, no git tag, no `tag_and_merge`, no notebook
+    generation/commit to workspaces, no Colab-URL bumps.
+- Emit the resolved TestPyPI version string as a workflow output / artifact so
+  the caller (Heart / Brain health agent) can install exactly those wheels.
+- Keep the existing full-release path untouched and default.
+
+This is intentionally small and isolated — it is the highest-value, lowest-risk
+piece and unblocks the rest of the redesign. It does not change the full release
+flow; it just exposes "build + TestPyPI, then halt" as a first-class mode.
+
+## Notes / footguns
+
+- Respect the "pure executor" boundary: `release.yml` runs no readiness checks;
+  it just builds/publishes. The gate lives in Heart.
+- Verify the TestPyPI upload step tolerates re-runs of the same version
+  (TestPyPI rejects duplicate filenames) — the rehearsal will be dispatched
+  repeatedly. Use a dev/local version suffix or `--skip-existing` semantics as
+  appropriate.
+- Confirm no `git+` direct URLs leak into any uploaded wheel's metadata
+  (the original `[nss]` failure) — a rehearsal that can't upload is the bug
+  this whole effort exists to surface early.
+
+## Validation
+
+- Dispatch the rehearsal mode manually; confirm wheels appear on TestPyPI, the
+  version string is emitted, and the workflow halts before PyPI/tag/notebook
+  steps.
+- `pytest` in PyAutoBuild stays green.
+
+## PR
+
+"PyAutoBuild: TestPyPI-only rehearsal mode in release.yml".
+
+<!-- formalised retroactively by the Intake (Conception) Agent on 2026-07-08 -->

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.
 
-939 records across 7 buckets.
+940 records across 7 buckets.
 
 <!-- CURATED:START -->
 ## Highlights
@@ -576,6 +576,7 @@ _(curate hard-won records here — survives regeneration.)_
 - [autolens-jax-simulator-release-fixes](2026/06/autolens-jax-simulator-release-fixes.md)
 - [autolens-results-aggregator-valid-dataset](2026/06/autolens-results-aggregator-valid-dataset.md)
 - [bib-duplicate-keys](2026/06/bib-duplicate-keys.md)
+- [build-testpypi-rehearsal-mode](2026/06/build-testpypi-rehearsal-mode.md)
 - [dark-matter-potentials](2026/06/dark-matter-potentials.md)
 - [datacube-delaunay-release-memory](2026/06/datacube-delaunay-release-memory.md)
 - [datacube-shared-state](2026/06/datacube-shared-state.md)

draft/maintenance/pyautomind/active_prompt_orphan_triage.md

@@ -0,0 +1,90 @@
+# Triage the 8 orphaned active/ prompts, then make `orphans` a gate
+
+Type: maintenance
+Target: PyAutoMind
+Repos:
+- PyAutoMind
+Difficulty: medium
+Autonomy: supervised
+Priority: normal
+Status: formalised
+
+`scripts/lifecycle.py orphans` reports **8 of 10** prompts in `active/` that no
+registry entry claims. Each is work whose state nobody is tracking. Triage them,
+dispose of each, then flip the reporter into a gate.
+
+## Background
+
+`check` gained registry-integrity validation on 2026-08-08
+(`draft/maintenance/pyautomind/registry_integrity_check.md`), covering the
+direction **registry → prompt**: every `prompt:` path resolves, exactly, into
+the state folder its registry implies. The audit that motivated it found six
+`planned.md` entries for work that had already shipped — including the whole
+M0–M3 release-validation milestone chain.
+
+`orphans` is the mirror direction, **prompt → registry**, and it was added in
+the same pass but deliberately left **report-only**: it exits 0 unless given
+`--check`. Wiring it into `check` immediately would have left the tree red with
+8 findings, and a red gate nobody can turn green gets ignored or disabled. The
+condition is only worth gating once the backlog is at zero.
+
+## The inventory (2026-08-08)
+
+```
+active/benchmark_calibration_runs.md
+active/ep_optimise_expose_updater_delta.md
+active/euclid_eceb_editorial_revision.md
+active/fix_workspace_start_here_colab_links.md
+active/matplotlib_inline_standalones.md
+active/pyautoreduce_slacs1430_acs_comparison.md
+active/research_profiling_experiment_in_the_autolens_pr.md
+active/workspace_version.md
+```
+
+**One is already diagnosed.** `ep_optimise_expose_updater_delta.md` belongs to
+the `ep-optimise-updater` entry in `active.md`, which records
+`status: COMPLETE 2026-08-08` with both PRs merged (PyAutoFit #1457 →
+`3b960609`, autofit_workspace #136 → `cf8b4077`). That is finished work still
+sitting in `active.md`, so it wants the normal ship disposal:
+`lifecycle.py record ep-optimise-updater --date 2026-08-08 --from-file <body>
+--prompt ep_optimise_expose_updater_delta.md --apply`, which writes the record,
+folds the prompt and drops the `active.md` section in one step. The entry
+already carries enough substance for a rich record. It is an orphan only
+because the entry declares no `prompt:` field and the slug does not match the
+filename stem.
+
+The remaining seven need the same question answered individually: **shipped,
+still in flight, or abandoned?** Answer it against the upstream repo, not from
+the prompt's own text — that is the lesson of the M0–M3 chain, where every
+prompt still read as pending while the capability was live on `main`.
+
+## Disposal per outcome
+
+- **Shipped** → `lifecycle.py record <slug> --date <YYYY-MM-DD> --from-file
+  <body> --prompt <name> --apply`. If the record has to be reconstructed after
+  the fact, say so in it and cite the upstream commit/PR — see
+  `complete/2026/06/build-testpypi-rehearsal-mode.md` for the shape.
+- **In flight** → add the missing `active.md` entry, including a `prompt:`
+  field so the claim is by path rather than by slug coincidence.
+- **Abandoned** → `complete/archive/shelved/`, or `condemned.md` if it is
+  self-material for the Gut's transit-and-void lifecycle.
+
+## Then close the loop
+
+Once `lifecycle.py orphans` reports none:
+
+1. call `orphan_prompts(ROOT)` from `cmd_check` so the condition becomes drift,
+   and delete the report-only carve-out documented in `cmd_orphans`;
+2. move the three orphan tests in `tests/test_lifecycle_check.py` under the
+   `check` legs alongside the registry ones;
+3. consider requiring a `prompt:` field on every `active.md` entry that has a
+   prompt file. Slug-matching exists only because entries predate the
+   convention, and it is the weaker claim — `ep-optimise-updater` shows a real
+   entry and its real prompt failing to match on name alone.
+
+## Acceptance
+
+- `python3 scripts/lifecycle.py orphans` prints `none`.
+- Every disposal is evidenced: a `complete/` record with an upstream
+  commit/PR, a new `active.md` entry, or an archive/condemned move.
+- `orphans` is called from `cmd_check`, and `pytest tests/` is green.