[Jammy2211]

Phase 2 of the cortex-checkin epic (#349). Stacked on #348 — retargets to main when it merges. The Cortex half is PyAutoLabs/PyAutoCortex#11, stacked on PyAutoCortex#10.

What this adds

pyauto-brain cortex checkin — the verb behind "where is my science?". It composes the primitives phase 1 decoupled (run_pull, score_phase, _apply_checked, render_pages, the payload builders) and reasons nothing extra of its own:

1. Sync. Every status: active row of projects.yaml, plus any project owning a phase in submitted | running — a dormant project with a job still out there is still out there. Each is pulled with its own <local_path>/<sync_cli> pull, the verb all seven implement. Output is streamed, not captured: a pull runs for minutes and the human is watching this one command. A non-zero exit is recorded against that project and the sweep continues.
2. Manifest. After a good pull, <pull root>/.cortex/pull.json gets project / pulled_at / cmd / rc / phases_live — merged into whatever the project's own CLI already wrote there. autolens_profiling's hpc/sync writes the checkpoints / runs tables that are the only window onto the RAL-only checkpoint.hdf5; clobbering them would blind the scorer's checkpoint leg.
3. Score + move. collect's scorer over every live phase, then the same rehearsed _apply_checked moves (running → pulled → awaiting-ruling).
4. Render. dashboard.md + dashboard.html.
5. Push. claude/checkin-<YYYY-MM-DD> cut from a fresh origin/main, explicit paths, no force, never main — and refused before the branch is cut if scripts/ledger_merge.py classify calls the diff code.
6. Summarise, by project, last — so a chat sees it above the fold. Per project: local_path / mirror / RAL root, what the pull did, phase counts, and every phase a human could act on today with its health verdict, the run dirs and logs the scorer touched, and the fenced copy-ready prompt its state already has.

The push rule

--push is allowed only when gh auth status succeeds and the Cortex checkout is clean on main — read before anything is written, since "clean" stops being true the moment the phases move. That is the whole cloud/laptop split and it is also the default: a laptop pushes without asking, a cloud session cannot and is told so. Stated in the output either way.

/cortex, installed rather than advertised

skills/COMMANDS.md has advertised /cortex since the conductor was born; ~/.claude/commands/cortex.md never existed. The skill body is now the check-in sequence (verbs demoted to an appendix), the stale plan / gates --grade rows retired in phase 1 are gone, and the committed .claude/{commands,skills} + .codex/skills discovery links exist alongside the other 29 commands. bin/install.sh needed no change — it already links skills/<verb>/<verb>.md; nobody had run it since the skill was added.

Tests

tests/test_cortex_conductor.py gains a check-in block over the existing board fixture: the dry-run listing (projects, exact pull commands, phases, nothing written), dry-run-is-the-default, per-project pull-failure isolation (a fake sync_cli exiting 1 for one project — the other still pulls, every phase is still scored, the failure lands on the right project and the exit code is 1), manifest write, merge-not-clobber, --skip-pull running no CLI, --project, the summary's by-project keying and its position last, the render leg leaving dashboard --check clean, the no-push path, the preflight refusing anything that is not a clean main, and the push refusing a code-classified diff. No test runs a real hpc/sync pull — every pull under test is a fake script, and no fixture names a real org or repo.

python3 -m pytest -q: 874 passed, 3 failed → tests/test_gh_surface.py was mine (the new skill body mentions gh auth status; it now points at GITHUB_ACCESS.md) and is fixed; the two tests/test_branch_sweep.py failures reproduce unchanged on the phase-1 base and are not this branch's. Targeted re-run of test_cortex_conductor.py + test_skill_install.py + test_gh_surface.py: 68 passed.

Verified live, without touching RAL

checkin --dry-run against the real Cortex: 3 active projects (euclid, inference_programme, subhalo_validation), the exact cd … && hpc/sync pull for each, 2 live phases (both subhalo_validation), and push would be: yes. It SSHes nowhere. --apply was exercised only against fixtures, deliberately: a real pull is minutes of RAL traffic and the human's call.

Noted, not fixed

bin/install.sh --check-project-discovery reports pre-existing drift unrelated to this branch (batch / start_bundle missing from the committed links, register_and_iterate / run_queue present and unwanted). Nothing checks it in CI. Left alone rather than swept into a cortex PR.

Closes #349 (with PyAutoCortex#11).

🤖 Generated with Claude Code

https://claude.ai/code/session_013HsZA1ufn3msgPiDFxEXa6