[dmccoystephenson]

Summary

A Stage A documentation sweep: README.md was checked against the gates, the workflows, the scenario docstrings and the tags. No code changed, and every addition describes current behaviour.

• Consumers gate: an assertion list was added, since the tests and save-compatibility gates had one and this gate did not. It records that control-<Name> always passes, that candidate-<Name> passes for a pre-existing consumer, and that control and harness are recorded only on failure. The harness's pytest and pytest-cov installs are also stated.
• All-skipped consumers: it is now stated that a run in which every consumer is skipped still passes. This is today's behaviour, and Gates can pass having checked nothing: every consumer skipped, or saves with no files #21 is open on whether to change it.
• result.json: the keys each gate adds were undocumented, although callers read them. A table was added under the shared shape.
• Fixture manifest: what expected and primaryFile do was added.
• "not in a tag yet" row: PR Read a scenario's view past stderr; match a pin URL only as a whole name #23 also made a non-object view fail, and the row did not say so. v5 takes any JSON on the last line as the view (confirmed with git show v5:gates/save_compat.py).

No tracking issue: this is a gap found during triage.

Checked with no drift found

• Input tables and defaults, against both trigger blocks of tests.yml, consumers.yml and save-compat.yml; the two blocks agree in all three.
• The tests and save-compatibility assertion lists, against the result.check(...) calls, in order.
• The time-limits table, against run(), the suite timeouts, the manifest download, pip_install() and timeout-minutes.
• The tag table, against git ls-tree and git grep on v1 to v5.
• External names: the baseline tags in the dispatch examples, and the Cave, Night Ferry, ophidian and plugin-fixtures repositories, all exist. The tak/fishe-2026-07-04 tag exists.
• Module docstrings in gates/*.py, workflow header comments, and the docstrings of tak_slots.py, tak_game_saves.py, cave_saves.py and fishe_saves.py. The remaining scenarios' docstrings were not read in this sweep, and the claims in the README's prose about them were not traced into their code.

Found but not changed

• workflow_call: 'Checkout gates' appears to check out the caller, not this repository #24 (filed): under workflow_call, the "Checkout gates" step appears to check out the caller rather than this repository. That is a code question, so it was filed rather than documented. The README's workflow_call sentence stays until the owner decides.
• CLAUDE.md is stale: the repository layout omits tests/, and the CI comment describes tests/ as covering declared migrations only. Editing agent-loaded config needs separate authorization, so it was left alone.
• installed_commit reads a stderr line as the commit, failing candidate-install on a correct install #19 (comment added): two more call sites read git rev-parse HEAD through run() and share the root cause.

Test plan

• git diff --stat origin/main: README.md only, 20 insertions and 2 deletions.
• Each added claim was read against its source, cited in the self-review.
• Check workflow green on the PR head.

Gate run: UNVERIFIED, not applicable. python3.12 needed an approval the dispatch could not receive, so no gate was run. This PR changes no gate, scenario or workflow, so none of the files a gate run would validate are touched.

Issues not taken this cycle

• installed_commit reads a stderr line as the commit, failing candidate-install on a correct install #19 was deferred: the fix is in gates/common.py, which is on the do-not-auto-merge list, and it needs a real gate run that this dispatch cannot make.
• Gates can pass having checked nothing: every consumer skipped, or saves with no files #21 was deferred: the issue asks the owner to choose between a new input and an unconditional failure.
• workflow_call: 'Checkout gates' appears to check out the caller, not this repository #24 was filed this cycle and needs an owner decision, as well as a real call from a second repository to verify any fix.

This PR description was drafted during a Gardener session (https://github.com/Stephenson-Software/gardener).

---

drafted by Claude on behalf of Daniel Stephenson

🤖 Generated with Claude Code

[dmccoystephenson]

Self-review rubric, scored against commit 0b2ab4d.

• CI + local gate run: PASS for CI; gate run UNVERIFIED-not-applicable. The Check workflow (compile, which also runs the unit suite on 3.12) passed on the head. python3.12 needed an approval this dispatch could not receive, so no gate ran, and git diff --name-only origin/main...HEAD lists only README.md, so nothing a gate run would validate was touched.
• Gate-run evidence: N/A — no script, scenario or workflow is touched.
• Scope: PASS — one file, 20 insertions and 2 deletions, all from the sweep.
• Tests-new: N/A — no code.
• Tests-fix: N/A — no bug fix.
• Tests-mutation: N/A — no characterization tests.
• Sibling structure: PASS — the assertions paragraph takes the form of the two sibling sections, with a per-consumer qualifier because this gate repeats per consumer, as the tests gate's line does per interpreter.
• Sibling renames: N/A — nothing was renamed.
• Docs: PASS — Phase 7 table pass results:
  • Gate tables against inputs: of all three workflows (both blocks): no drift.
  • Assertion lists against result.check(...): consumers list was missing, added; the other two had no drift.
  • result.json shape against Result.doc: shared keys had no drift; per-gate keys were missing, added from the result.doc writes in each gate.
  • Fixture manifest against fixture_phase(): expected and primaryFile behaviour added. README.md:147 was read against gates/save_compat.py:325-335.
  • Module docstrings and workflow header comments: no drift.
  • CLAUDE.md: drift found and not corrected (see below).
• Visuals: N/A — nothing a README image depicts changed.
• Issue resolution: N/A — no Closes reference.
• Input parity: N/A — no workflow touched; the blocks were compared anyway and agree.
• Read-only: PASS — git grep -n "secrets\." .github/workflows is empty.
• Stdlib only: N/A — no code.
• No tag moved: PASS — git ls-remote --tags origin still shows v1 to v5, tak/fishe-2026-07-04 and roam/longplay-0.12.0; none was pushed.

Claims traced to source:

• control-<Name> always passes: gates/consumers_gate.py:151, called with True.
• candidate-<Name> passes when pre-existing: gates/consumers_gate.py:166-168.
• control recorded only on failure: gates/consumers_gate.py:119,128,145, all called with False.
• pytest, pytest-cov and extras: gates/consumers_gate.py:141.
• All-skipped run passes: gates/consumers_gate.py:133-137, and the reproduction in Gates can pass having checked nothing: every consumer skipped, or saves with no files #21.
• Non-object view in v5: v5:gates/save_compat.py:204 returns any parsed JSON.

Judgment calls for a reviewer:

• README.md:103 — the all-skipped sentence documents behaviour that Gates can pass having checked nothing: every consumer skipped, or saves with no files #21 may change. Whichever way Gates can pass having checked nothing: every consumer skipped, or saves with no files #21 is decided, this sentence must change with it.
• The per-gate keys table is a new surface to keep in sync, since the Phase 7 table's result.json row names only Result.doc.
• The sweep did not read the docstrings of gardener_state.py, usd_recorder_recordings.py, trace_client_install_id.py, roam_saves.py, ophidian_saves.py or menu_bot.py, so the README's prose about them is unverified by this pass.
• CLAUDE.md omits tests/ from the repository layout and describes it as covering declared migrations only. It was left unedited because editing agent-loaded config needs separate authorization.

Phases 5 to 7: no human reviewer can be awaited in a headless dispatch, and no review comment changed anything, so the table pass above stands. Merge is judged next: README.md matches no do-not-auto-merge path, and no assertion's name or meaning changes.

This comment was drafted during a Gardener session (https://github.com/Stephenson-Software/gardener).

---

drafted by Claude on behalf of Daniel Stephenson