There is no CI — every "340 passed" is one person's local run #33

.github/workflows does not exist. Verified today: ls -d .github/workflows → No such file or directory.

So every green claim in every PR body in this repo is one local run on one machine, including all eight merged on 2026-10-07/08.

Why this is worth an issue rather than a shrug

The repo's whole argument is that a claim needs a check that can go red. It currently applies that to designs and not to itself.

Two things made it concrete this week:

• Copilot stopped reviewing. Measured 2026-10-08: zero reviews on README: say there is no flagship example, and open #24 for it #25 and potential_incoherence (33 members), check derived, and eval_battle(run_count=N) #26, and a deliberate probe PR in nobsmed-v2 also got none after 7 minutes, against a 1–5 minute historical norm. So for a stretch there was no second reader of any kind — not a human, not a bot, not CI.
• A local run is not a clean-checkout run. The main clone's .venv had no pytest and uv sync --group dev --all-extras did not repair it; the suite only ran in a fresh worktree. A green on a long-lived venv says little about a fresh clone.

Minimum that would help

uv sync
uv run pytest -q
uv run python3 -m graph_builder_spec.reference --check   # the generated tables

The third line matters as much as the second: both README reference tables are generated, and the only thing stopping a hand-edit is a check nobody is obliged to run.

⚠️ tests/test_browser.py and the island tests drive a real browser and may need playwright install or a marker exclusion. Start with everything not browser-marked rather than letting that block the whole file.

⚠️ And the honest counter-argument, recorded so it is not re-litigated from scratch: this is a one-person library with no users yet, and CI on a repo nobody depends on is ceremony. The thing that tips it is that two repos already track main — nobsmed-v2 and ai_computer_use both consume this unpinned, so a bad main reaches them on their next sync.