refactor: rename the `multi` example package to `multi_dataset`

[Jammy2211]

Overview

The multi example package is renamed to multi_dataset across every workspace, doc and downstream consumer.

multi is ambiguous: it collides conceptually with the sibling multi_galaxy package, with the multi_gaussian_expansion feature folders, and with the general adjective "multi-" used throughout the prose ("multi-wavelength", "multi-plane", "multi-band"). multi_dataset names what the package actually is — the examples for fitting multiple datasets simultaneously via a factor graph.

Scope is the full rename, confirmed with the user: package directories and dataset paths and output path_prefix values. That last part is a breaking change for existing users' on-disk output/multi/ directories, and is accepted — it needs a release-note line.

Plan

• Phase 1 — user workspaces (autolens_workspace, autogalaxy_workspace): git mv the scripts/, notebooks/, markdown/ and dataset/ packages, rewrite every reference and the folder-naming prose, sweep the path-keyed sidecar configs, regenerate notebooks/markdown/indexes.
• Phase 2 — test / profiling / dev (autolens_workspace_test, autogalaxy_workspace_test, autolens_profiling, autolens_workspace_developer, HowToLens). Currently blocked — see the branch survey.
• Phase 3 — downstream reference consumers (PyAutoLens/docs, PyAutoGalaxy/docs, autolens_assistant, autolens_jax_joss). Gated on phase 1 merging, because these are GitHub blob/main URLs and paired_example pointers that 404 or dangle until the workspace move lands.
• Verify each repo by a post-move grep for the old path shape returning zero live hits, and by sidecar entry counts being preserved rather than merely changed.

Old	New
scripts/multi/	scripts/multi_dataset/
notebooks/multi/	notebooks/multi_dataset/
markdown/multi/	markdown/multi_dataset/
dataset/multi/	dataset/multi_dataset/
dataset_type = "multi"	dataset_type = "multi_dataset"
output/multi/... (path_prefix)	output/multi_dataset/...

Detailed implementation plan

Affected Repositories

• autolens_workspace (primary) — 89 tracked files under multi/, 40 files with references
• autogalaxy_workspace — 61 tracked files, 35 files with references
• autolens_workspace_test — 15 tracked files, 16 files with references
• autogalaxy_workspace_test — 11 tracked files, 12 files with references
• autolens_profiling — 3 tracked files, 7 files with references
• autolens_jax_joss — 8 files with references
• autolens_assistant — 5 files with references
• PyAutoGalaxy — 2 files (docs URLs)
• PyAutoLens — 1 file (docs URLs)
• autolens_workspace_developer — 2 files
• HowToLens — 1 comment

Branch Survey

Repository	Current Branch	Dirty?
./autolens_workspace	main	clean
./autogalaxy_workspace	main	clean
./autolens_workspace_test	main	clean
./autogalaxy_workspace_test	main	clean
./autolens_profiling	main	DIRTY (untracked dataset/imaging/jwst_lw/, unrelated)
./autolens_jax_joss	main	clean
./autolens_assistant	main	DIRTY (M README.md, unrelated)
./PyAutoLens	main	clean
./PyAutoGalaxy	main	clean
./autolens_workspace_developer	main	clean
./HowToLens	main	clean

Worktree conflict (hand-checked, not via worktree_check_conflict): autolens_workspace_test, autolens_profiling and autolens_workspace_developer are real worktrees (not symlinks) checked out on feature/point-source-chi-squared-variants under ~/Code/PyAutoLabs-wt/point-source-chi-squared-variants/, held by that task's in-flight phase 3. Phase 2 is blocked until it releases them. autogalaxy_workspace_test is free.

Suggested branch: feature/multi-package-rename-multi-dataset

Implementation Steps

Phase 1 — autolens_workspace + autogalaxy_workspace

1. git mv scripts/multi scripts/multi_dataset, and the same for notebooks/, markdown/, and (autolens only) dataset/multi.
2. Rewrite in-script path values: dataset_type = "multi" → "multi_dataset"; path_prefix=Path("multi", "modeling"); path_prefix=Path("multi") / "features"; the nested path_prefix=Path("slam", "multi", "simultaneous") (a leading-multi/ sweep misses this one); dataset_main_path = Path("dataset", "multi", "imaging", ...); dataset_path = Path("dataset") / "multi" / "rxj1131".
3. Rewrite cross-package references from sibling packages: interferometer/features/datacube, weak/features/strong_lensing, guides/modeling, guides/results/aggregator, imaging, point_source/start_here.py, misc/database/scrape, misc/searches, misc/simulators, gallery.
4. Rewrite folder-naming prose (no path-shaped regex catches these): README.md, scripts/README.md, AGENTS.md, root start_here.ipynb, the moved packages' own README.md and features/README.md.
5. Sweep the path-keyed sidecars by hand (see "Traps" below).
6. Regenerate notebooks + markdown + workspace_index.json + llms-full.txt.

Phase 2 — same mechanical pattern in the test/profiling/dev repos, plus their config/build/no_run.yaml SLOW and BOOTSTRAP-TARGET entries and smoke_tests.txt entries.

Phase 3 — GitHub URL rewrites in the two model_cookbook.md files and PyAutoGalaxy/docs/overview/overview_2_new_user_guide.md (which includes a version-pinned Colab URL, blob/2026.7.29.2/notebooks/multi/...), the */multi package prose in both overview_3_features.md, the autolens_assistant wiki/skills pointers, and the autolens_jax_joss paired_example fields in benchmarks/*.py, results/*.json, results/quick/*.json, results/RESULTS.md and README.md.

Traps

1. Path-keyed sidecars fail OPEN on a git mv — no error, just silent coverage loss. Sweep by hand; prove by entry count, not content:

• smoke_tests.txt — all four workspaces, including a commented-out # multi/start_here.py line in autolens_workspace
• config/build/no_run.yaml — both _test repos (SLOW + BOOTSTRAP-TARGET entries; one comment also names dataset/multi/lens_sersic)
• config/build/profile_release.yaml — autolens_workspace has - pattern: "multi/start_here", a pattern match, so it silently stops matching rather than erroring
• .script_sizes.json — both workspaces; edit by hand, do not run check_sizes.sh --update (it sweeps in unrelated repo-wide drift)
• .navigator_check_ignore — currently no multi entries, but re-check after the move
• autolens_workspace_developer/.gitignore — jax_profiling/dataset/multi/
• HowToLens/config/build/profile_smoke.yaml — a comment naming multi/start_here

2. multi/ is not always a path — do NOT blind-sed. In scripts/point_source/features/multiple_sources/ and its notebooks, "the multi/factor-graph API" is English meaning multi-dataset, not the folder. Those must stay (or be reworded to "multi-dataset/factor-graph"). Conversely, prose naming the folder — "Unlike other multi simulators", "The multi package extends…", "Checkout the autolens_workspace/*/multi package" — does need rewriting and no path-shaped regex catches it. The sweep is grep-assisted with eyes on each hit, not automated.

3. Out of scope, deliberately: PyAutoMind/complete/** and draft/** historical records (a log of what happened, not live references — do not rewrite history); generated .artifacts-* directories; and every multi token that is not this package — multi_galaxy, multi_gaussian_expansion, multipoles, multiplane, MultiStartProdigy, multiprocessing, and all prose uses of "multi-wavelength" / "multi-band".

Verification

• Post-move grep for the old path shape returns zero live hits per repo.
• Sidecar entry counts preserved (not merely changed).
• Notebooks/markdown regenerated via generate.py from the repo root with the short project key (al = 308 scripts, ag = 132) — from the workspace root it prints "0 scripts" and silently regenerates nothing.
• features/ subpackage names are unchanged throughout.

Key Files

• */scripts/multi/ — the package being moved (both workspaces + both _test repos + profiling)
• */smoke_tests.txt, */config/build/no_run.yaml, */config/build/profile_release.yaml, */.script_sizes.json — path-keyed sidecars that fail open
• */workspace_index.json, autolens_workspace/llms-full.txt — generated indexes, regenerate
• PyAutoLens/docs/general/model_cookbook.md, PyAutoGalaxy/docs/general/model_cookbook.md, PyAutoGalaxy/docs/overview/overview_2_new_user_guide.md — GitHub blob + pinned Colab URLs
• autolens_jax_joss/benchmarks/*.py, autolens_jax_joss/results/** — paired_example pointers
• autolens_assistant/skills/al_multi_dataset.md, autolens_assistant/wiki/core/** — assistant pointers

Phasing note

pyauto-brain feature scored this too-large (28) and proposed the generic design → core_api → workspace_examples → docs split. That is the repo-count difficulty heuristic and is wrong here — there is no design step and no API to change. Overridden in favour of the per-repo, merge-dependency-ordered split above; the override is recorded in the prompt.

Original Prompt

Click to expand starting prompt

Rename the multi example package to multi_dataset

Type: refactor
Target: workspaces
Repos:

• autolens_workspace
• autogalaxy_workspace
• autolens_workspace_test
• autogalaxy_workspace_test
• autolens_profiling
• autolens_jax_joss
• autolens_assistant
• PyAutoLens
• PyAutoGalaxy
  Difficulty: large
  Autonomy: supervised
  Priority: normal
  Status: formalised

Original request

Rename the multi package as multi_dataset, which is a huge endeavor with docs,
URLs and everything so be thorough

Why

multi is ambiguous — it collides conceptually with the sibling multi_galaxy
package, with multi_gaussian_expansion feature folders, and with the general
adjective "multi-" used throughout the prose ("multi-wavelength", "multi-plane",
"multi-band"). multi_dataset names what the package actually is: the examples
for fitting multiple datasets simultaneously via a factor graph.

Scope (confirmed with the user)

Full rename, including dataset and output paths:

Old	New
scripts/multi/	scripts/multi_dataset/
notebooks/multi/	notebooks/multi_dataset/
markdown/multi/	markdown/multi_dataset/
dataset/multi/	dataset/multi_dataset/
dataset_type = "multi"	dataset_type = "multi_dataset"
output/multi/... (path_prefix)	output/multi_dataset/...

All nine repos above are in scope. This is a breaking change for existing users'
on-disk output/multi/ directories — that is accepted.

Surveyed surface (~330 path-shaped references)

Directories to git mv (scripts / notebooks / markdown / dataset):

• autolens_workspace: scripts, notebooks, markdown, dataset (incl.
  dataset/multi/rxj1131 real data)
• autogalaxy_workspace: scripts, notebooks, markdown
• autolens_workspace_test: scripts/multi (jax_likelihood, visualization,
  images), dataset/multi/lens_sersic
• autogalaxy_workspace_test: scripts/multi (jax_grad, jax_likelihood)
• autolens_profiling: scripts/multi/likelihood_runtime,
  dataset/multi/imaging, results/runtime/multi

Path-keyed sidecars that fail OPEN on a git mv (must be swept explicitly —
see [[feedback_extension_filtered_grep_misses_dotfiles]] and
[[feedback_no_run_suffix_entries_break_on_moves]]):

• smoke_tests.txt — all four workspaces have multi/... entries (autolens has
  a commented-out # multi/start_here.py line too)
• config/build/no_run.yaml — autolens_workspace_test (SLOW + BOOTSTRAP-TARGET
  entries, incl. a comment naming dataset/multi/lens_sersic),
  autogalaxy_workspace_test (SLOW + BOOTSTRAP-TARGET)
• .script_sizes.json — autolens_workspace + autogalaxy_workspace (path keys;
  do not run check_sizes.sh --update, which sweeps in unrelated repo-wide
  drift — see [[project_dspl_terminology_rename]])
• .navigator_check_ignore — currently no multi entries, but re-check
• autolens_workspace_developer/.gitignore — jax_profiling/dataset/multi/
• HowToLens/config/build/profile_smoke.yaml — comment naming
  multi/start_here
• config/build/profile_release.yaml — autolens_workspace has
  - pattern: "multi/start_here" (a pattern match, so it silently stops
  matching after the move rather than erroring)

Generated indexes keyed by path (regenerate, do not hand-edit):

• workspace_index.json — both workspaces (dozens of multi/... entries)
• llms-full.txt — autolens_workspace

Package-listing prose that names the folder:

• autolens_workspace/README.md, scripts/README.md, AGENTS.md
• autolens_workspace/start_here.ipynb (root) — */multi package pointer
• the moved packages' own README.md / features/README.md bodies
• guides/results/aggregator/ (both workspaces) — "The multi package of the
  workspace illustrates…"

Output path_prefix values (in-scope per the confirmed scope):

• path_prefix=Path("multi", "modeling"), Path("multi") / "features"
• path_prefix=Path("slam", "multi", "simultaneous") — note the nested
  form under output/slam/, which a naive leading-multi/ sweep misses
• dataset_main_path = Path("dataset", "multi", "imaging", dataset_name) and
  dataset_path = Path("dataset") / "multi" / "rxj1131"

Ambiguity trap — do NOT blind-sed

multi appears as an English word adjacent to a slash in prose that is not
a path. In scripts/point_source/features/multiple_sources/ and its notebooks,
"the multi/factor-graph API" means multi-dataset / factor-graph, not
multi/. Those must stay (or be reworded to "multi-dataset/factor-graph"),
not rewritten to multi_dataset/factor-graph. Every replacement needs eyes on
it; the sweep is grep-assisted, not sed-automated.

Conversely, prose that names the folder — "Unlike other multi simulators",
"The multi package extends…", "Checkout the autolens_workspace/*/multi
package" — does need rewriting, and a path-shaped regex will not catch it.

Docs & URLs:

• PyAutoLens/docs/general/model_cookbook.md — 2 GitHub blob URLs
• PyAutoGalaxy/docs/general/model_cookbook.md — 2 GitHub blob URLs
• PyAutoGalaxy/docs/overview/overview_2_new_user_guide.md — a GitHub blob URL
  and a version-pinned Colab URL (.../blob/2026.7.29.2/notebooks/multi/...)
• PyAutoLens/docs/overview/overview_3_features.md,
  PyAutoGalaxy/docs/overview/overview_3_features.md — */multi package prose
• In-notebook Binder / Colab / "download" links inside the moved notebooks
  themselves

Cross-repo consumers:

• autolens_jax_joss — README.md, benchmarks/*.py (paired_example=),
  results/*.json, results/quick/*.json, results/RESULTS.md
• autolens_assistant — skills/al_multi_dataset.md (already named
  multi_dataset!), wiki/core/api/datasets.md,
  wiki/core/api/analysis_objects.md, wiki/core/concepts/multi_wavelength.md
• Sibling workspace scripts referencing multi/ cross-links:
  interferometer/features/datacube, weak/features/strong_lensing,
  guides/modeling, imaging, misc/database/scrape,
  misc/searches, misc/simulators, gallery

Explicitly OUT of scope:

• PyAutoMind/complete/** and PyAutoMind/draft/** historical records — these
  are a log of what happened, not live references. Do not rewrite history.
• Generated .artifacts-* directories.
• Any word containing multi that is not this package: multi_galaxy,
  multi_gaussian_expansion, multipoles, multiplane, MultiStartProdigy,
  multiprocessing, and all prose uses of "multi-wavelength"/"multi-band".

Phasing (Brain override — recorded)

pyauto-brain feature scored this too-large (28) and proposed the generic
design → core_api → workspace_examples → docs split. That is the repo-count
heuristic ([[feedback_brain_repo_count_difficulty_proxy]]) and is wrong here —
there is no design step and no API to change. Overridden in favour of a
per-repo split ordered by merge dependency:

• Phase 1 — user workspaces (autolens_workspace, autogalaxy_workspace).
  The bulk of the work; everything downstream points at these.
• Phase 2 — test / profiling / dev (autolens_workspace_test,
  autogalaxy_workspace_test, autolens_profiling,
  autolens_workspace_developer, HowToLens comment).
• Phase 3 — downstream reference consumers (PyAutoLens/docs,
  PyAutoGalaxy/docs, autolens_assistant, autolens_jax_joss).
  Gated on phase 1 merging — these are GitHub blob/main URLs and
  paired_example pointers that 404 / dangle until the workspace move lands.

Phase 2 is BLOCKED at time of writing: autolens_workspace_test,
autolens_profiling and autolens_workspace_developer are real worktrees held
by the in-flight point-source-chi-squared-variants phase 3
(feature/point-source-chi-squared-variants). autogalaxy_workspace_test is
free. Phase 2 starts once that task releases its claim.

Execution notes

• Library-first merge gate does not apply cleanly here: the library-side
  changes are docs-only (URLs). But the docs URLs point at
  autolens_workspace@main, so the workspace PRs must merge first or the
  doc links 404 in the interim.
• README.md bodies inside the moved packages describe "the multi folder" —
  rewrite the prose, not just the paths.
• The features/ subpackage names stay unchanged.
• Prove each repo with a post-move grep for the old path shape returning zero
  live hits, and by the .script_sizes.json / smoke_tests.txt entry counts
  being preserved (not merely changed) — see
  [[feedback_smoke_entry_path_and_count]].
• Regenerate notebooks/markdown via generate.py from the repo root with the
  short project key (al = 308 scripts, ag = 132) — see
  [[feedback_generate_py_cwd_and_project_key]].

[Jammy2211]

Phase 1 shipped — PRs open

• autolens_workspace: refactor: rename the multi example package to multi_dataset #414
• autogalaxy_workspace: refactor: rename the multi example package to multi_dataset autogalaxy_workspace#194

Both labelled pending-release. Heart was YELLOW (score 75, no RED reasons); the three reasons are pre-existing organism-level drift unrelated to this change and were explicitly acknowledged by a human before pushing.

Two traps that would have shipped silently

1. bound="multi" in scripts/guides/modeling/searches.py (both workspaces, ×2 each) is a Dynesty search kwarg. A blind s/multi/multi_dataset/ would have corrupted four search configurations with green CI.
2. config/build/profile_release.yaml keys on - pattern: "multi/start_here" — a pattern match, so after the git mv it would have silently stopped matching rather than erroring. Same failure mode for smoke_tests.txt, .script_sizes.json, .url_check_allowlist.txt and config/build/markdown_examples.yaml; each swept explicitly and proved by entry count (al: 28 entries / 401 keys; ag: 22 / 174).

Scope notes discovered during the work

• dataset/multi/ is gitignored, not tracked — so it is a script-path and release-note concern, not a git mv. Users' local copies are simply re-simulated.
• Version-pinned Colab link: bump_colab_urls rewrites only the version component at release, never the path. The path was renamed, so the link resolves again at the next release tag — leaving it would have made it permanently wrong after the next release instead of temporarily wrong before it.
• markdown/ was updated in place, not regenerated. generate_markdown.py executes every curated script for real (no test mode, at-release cadence); for a pure rename that is hours of sampling to reproduce byte-identical figures. One captured execution log line from a 2026-07-10 run still shows an output/multi/ path — that is a record of an actual run and was left rather than rewritten. markdown/ is stale against scripts/ independently of this work and wants a real regeneration at release.

Drive-by fixes (all pre-existing)

Four "can be viewed in the folder" docstrings had dataset_type and dataset_label swapped; two of those also named the wrong dataset_name. Two stale markdown sentences referenced advanced/multi/modeling, a path in neither the old nor new layout.

Verification

Zero live hits for the old path shape in either repo (excluding the four legitimate bound="multi" kwargs and the captured log line). All package scripts byte-compile; every rewritten subprocess path resolves; simulator.py runs green in both repos, writing to dataset/multi_dataset/imaging/{lens_sersic,simple}. Regeneration counts matched exactly: al 308 scripts, ag 132.

---

Phase 2 (autolens_workspace_test, autogalaxy_workspace_test, autolens_profiling, autolens_workspace_developer, HowToLens) remains blocked on point-source-chi-squared-variants releasing its worktree claims on three of those repos.

Phase 3 (PyAutoLens/PyAutoGalaxy doc URLs, autolens_assistant, autolens_jax_joss) is gated on the two PRs above merging — its links are blob/main URLs and paired_example pointers that dangle until then.

[Jammy2211]

Phase 3 shipped — PRs open

Downstream reference consumers re-pointed at multi_dataset:

• docs: point multi-dataset example links at the renamed multi_dataset package PyAutoLens#673 — docs blob/main links + */multi prose
• docs: point multi-dataset example links at the renamed multi_dataset package PyAutoGalaxy#542 — same, plus the version-pinned Colab link
• docs: point workspace references at the renamed multi_dataset package autolens_assistant#105 — skill, wiki and audit path citations
• docs: point paired_example references at the renamed multi_dataset package autolens_jax_joss#2 — paired_example pointers in benchmarks, results and README

⚠️ Merge order

autolens_workspace#414 and autogalaxy_workspace#194 must merge first. The four PRs above are blob/main URLs and paired_example pointers that 404 or dangle until the package actually moves on main. Each PR body states this.

Three more multi tokens that are NOT this package

Found and excluded during the phase 3 sweep:

1. bound: multi in PyAutoLens/autolens/config/non_linear.yaml, test_autolens/config/non_linear.yaml and test_autogalaxy/config/non_linear.yaml — a Dynesty sampler config value. That is now eight bound="multi" sites across the organism that a blind s/multi/multi_dataset/ would have corrupted.
2. harness.fetch_url(HIPS2FITS_URL, "multi/sdp81", ...) (and rxj1131) in autolens_jax_joss — that argument is relative_path, this repo's own local dataset cache key (dataset/<relative_path>/<file_name>), not a workspace path. Renaming it would invalidate cached FITS downloads for no correctness gain.
3. id: assistant-hard-group-multi and the filename hard_group_multi.md — benchmark identifiers. Renaming them would break comparison against historical benchmark results, so only the workspace_packages: entry inside was renamed.

On editing recorded artifacts

paired_example in the JOSS results JSON/markdown is a pointer to the corresponding workspace script, not a measurement — re-pointing keeps the record findable and no timing, iteration count or likelihood value was touched. By the same principle, the captured output/multi/ path inside a 2026-07-10 execution log in markdown/multi_dataset/modeling.md was left alone, since rewriting it would misrepresent what that run actually produced.

---

Remaining: phase 2 only — autolens_workspace_test, autogalaxy_workspace_test, autolens_profiling, autolens_workspace_developer, HowToLens. Still blocked on point-source-chi-squared-variants releasing its worktree claims on three of those repos. autogalaxy_workspace_test is unclaimed and could go early if worth splitting.

[Jammy2211]

Phase 2a shipped + phase 3 CI triage

Phase 2a — PyAutoLabs/autogalaxy_workspace_test#101

Of the five phase-2 repos, autogalaxy_workspace_test was the only one not held by another task's worktree, so it went now. Renamed scripts/multi/ → scripts/multi_dataset/ and dataset/multi/jax_test → dataset/multi_dataset/jax_test; swept smoke_tests.txt (9 entries preserved) and config/build/no_run.yaml (3, including the BOOTSTRAP-TARGET line whose trailing comment names the dataset path it produces — a path-shaped sweep of the entry alone would have missed it).

Verified end-to-end across the producer/consumer boundary: the BOOTSTRAP-TARGET simulator writes dataset/multi_dataset/jax_test, and jax_grad/lp.py reads that renamed path back and passes its JAX gradient checks.

Drive-by: both jax_grad scripts printed jax_grad/multi/<name>.py on success — path components reversed vs their real location. Pre-existing; corrected.

HowToLens needs no change

Its only hit is inside a dated 2026-07-23 historical comment recording which override patterns were removed and why. Those patterns literally said multi/start_here; rewriting the note would misrepresent the record. Same principle applied to the captured execution log in markdown/ and to the JOSS timing measurements.

Phase 3 CI

• docs: point multi-dataset example links at the renamed multi_dataset package PyAutoLens#673 — all green (docs-build + unittest 3.12/3.13)
• docs: point multi-dataset example links at the renamed multi_dataset package PyAutoGalaxy#542 — all green
• docs: point workspace references at the renamed multi_dataset package autolens_assistant#105 — wiki-currency failed, now fixed

The assistant failure was not the cross-repo citation gate — --check-citations passed with 0 missing paths even before the workspace PR merges. It was --check-provenance: each wiki/core page carries a content_sha256 stamp that any body edit invalidates, and it errored on exactly the three pages this branch touched. Re-stamped with --write-provenance --page <each> (the documented targeted form — "only stamp what you re-validated"); these edits changed workspace paths only, no API claim, so each page's pinned commit still holds. --check-provenance now reports 0 errors, down from 3.

Phase 1 CI — all green

Both PRs, both matrix legs, including navigator / Catalogue staleness and the renamed smoke entries — which is the direct proof that the smoke_tests.txt sweep landed.

---

Remaining: phase 2b only — autolens_workspace_test, autolens_profiling, autolens_workspace_developer, still held by point-source-chi-squared-variants (and autolens_workspace_test additionally by potential-correction-validation).

[Jammy2211]

The breaking change was silently filed as Internal

Caught while checking that the output/multi/ break would actually reach users.

PyAutoHands/autohands/generate_release_notes.py::classify_pr decides a PR's release-note category from body structure, not from the title, the labels, or how loudly the body says BREAKING:

has_breaking = any(h in combined for h in
    ["### Removed", "### Renamed", "### Changed Signature", "### Changed Behaviour"])

where combined comes from a ## API Changes section matched by regex. With no such section a PR falls through to internal.

All three workspace PRs already carried a prominent ### ⚠️ Breaking for existing users block explaining that output/multi/ results are not migrated — and still classified as internal. Verified by running the real classify_pr against the live PR bodies:

PR	before	after
autolens_workspace#414	internal	breaking
autogalaxy_workspace#194	internal	breaking
autogalaxy_workspace_test#101	internal	breaking

Fixed by appending an ## API Changes section with ### Renamed and ### Changed Behaviour to each. The four docs-only phase-3 PRs correctly remain internal — only the PRs that actually move the package need it.

This is the same fail-open shape as the path-keyed sidecars this rename already had to sweep by hand: nothing errors, CI stays green, and the omission is invisible until a user is surprised after a release. A heading that reads as breaking is not the heading the parser matches.

(gh pr edit fails on these repos with the Projects-classic GraphQL error; the bodies were patched via gh api -X PATCH repos/<owner>/<repo>/pulls/<num>.)

Worth considering as a follow-up: a CI check that fails a PR whose body says "breaking" in prose but classifies as internal.