[Jammy2211]

Closes #545 (phase 2 of making the Cortex the Science project manager; phase 1 was #544).

Summary

• pyauto-brain cortex birth <key> --category active|development --assistant A --summary S registers a project folder that the domain assistant's start-new-project already scaffolded. Scaffolding stays with the assistants.
• It derives the projects.yaml row from the folder and the other rows:
  • the remote, from the git origin;
  • ral_root, the rows' RAL base plus the category and key;
  • sync_verbs, from the hpc/sync dispatch case (help excluded);
  • the ledger file.
• It's a dry run by default, which also enforces the 15-word summary limit.
• --apply (needs the Cortex clean on main) does five things:
  1. cuts claude/birth-<key> from origin/main;
  2. writes through PyAutoCortex's cortex.py birth; the conductor writes no ledger itself;
  3. commits;
  4. pushes only if ledger_merge.py classify --base origin/main calls the commit a birth, so it lands by itself;
  5. returns to main.
• A failed write or commit abandons the branch and leaves main clean. Exit codes: 2 refused, 1 for any git, write or push failure, 0 landed or --no-push.
• The /cortex skill gains a "Start a new project" section: scaffold with the assistant, take data from dataset/ or simulators/, then cortex birth.

API Changes

None to the library API. The conductor gains a verb: cortex birth. It needs PyAutoLabs/PyAutoCortex's companion PR (cortex.py birth); until that merges, the three --apply tests skip.

Validation

• A dry run on euclid_detailed derives a row identical to its registered one, then refuses because the key and path are taken.
• End-to-end --apply against a throwaway clone with a local bare origin: written, classified as a birth, pushed, back on clean main, drift clean.
• Derived sync_verbs match four of the seven registered active Science rows exactly. For analytic_gaussian, ep_toy_gaussian and slope_hierarchy_scale, the registered rows list fewer verbs than their hpc/sync now dispatches. That's stale rows, not this code.

Review findings addressed

• A failed git commit, for example from a hook, used to leave the birth staged on main and exit 0 with a false "committed locally" message. It now abandons the branch, leaves main clean and exits 1. There's a test that uses a rejecting pre-commit hook.
• A failed push and a classifier refusal now exit 1, as documented. There's a test.
• A backslash in a value was silently treated as a YAML escape. It's now refused, and the read-back compares every field (Cortex PR).
• help is no longer counted as a sync verb.
• The dry run enforces the summary limit.

Open question for the human: a development/ birth's ral_root is <base>/development/<key>, mirroring the laptop as decision 61 and drift expect. But the four existing development projects still sit under RAL active/, and no development/ folder exists on RAL yet.

Heart RED override (development only)

Authorized by the live human on 2026-10-29: "I authorize development only heart override" (for PyAutoBrain#545, after the exact RED reasons were shown). Development ship only: commit, push, PR-open. Merge needs a separate human /prm with every required check green. No release.

Exact RED reasons (pyauto-heart readiness, 2026-10-29):

• PyAutoArray: 4 uncommitted source change(s)
• PyAutoGalaxy: 9 commit(s) behind origin
• PyAutoLens: 7 commit(s) behind origin
• release validation incomplete: no rehearsal for current source
• release push credential check stale (pyautolabs-bot, 49h old)

None of these involves PyAutoBrain or PyAutoCortex. This branch does not fix them.

Passed branch gates: Brain tests/test_cortex_conductor.py 80 passed (3 end-to-end tests skip cleanly against a Cortex without birth); Cortex 70 tests + cortex check: OK; no new lint beyond the file's existing idiom; smoke n/a (organ tooling); independent review FINDINGS, all addressed (below).

🤖 Generated with Claude Code