dev-gate

name	dev-gate
description	Run and interpret the Saitenka quality gates — the fast pre-push `poe all` and the slower pre-tag `poe pre-release`. Use before pushing or when asked to "run the gate", "run poe all", "check it passes / is it green", "lint/types/arch/tests before commit", "pre-release gate", or when a gate task fails and you need to read the failure. Explains both bundles, the advisory tier, and the free-threaded / 3.13-pinned-env traps that bite agents. NOT for authoring a test (use the write-test skill); the release flow itself is RELEASING.md; NOT for mutation/fuzz/crosshair adequacy (the `test-adequacy` skill).
metadata	{"project":"saitenka"}

dev-gate

Two gates, both defined in pyproject.toml [tool.poe.tasks] — that is the source of truth for what each runs; uv run poe --dry-run <gate> prints the exact chain. Run either from the repo root. CI (.github/workflows/ci.yml) mirrors poe all on PRs + pushes to main.

• poe all — the fast pre-push PR gate. Run before every push.
• poe pre-release — the slower superset run once before tagging; release.py gates on this, not all. Adds the supply-chain + installer checks and the heavier real-mpv / network smokes; needs mpv + network, so not for PR iteration. The human-triage advisory tools (hygiene / ps1 / perf-risk) are deliberately not in it — run those by hand.

Reading a failure (the non-obvious ones)

• types — a real type error; don't # type: ignore without a reason.
• arch (import-linter) — a forbidden import (cycle, PIL-in-core, GPL chokepoint): move the code, don't relax the contract.
• complexity — a function got more complex: simplify. Regenerate the baseline (poe complexity-baseline) only after a deliberate refactor, never to silence a regression.
• cov — floor 85%: add a behaviour test, not a coverage-painting one.
• test-ft — a C-ext re-enabled the GIL, or a genuine no-GIL race.
• licenses — a copyleft dep leaked in (only our own GPL deinflect is allowed): drop it.
• bench — a crash is API rot; the printed numbers are informational, not pass/fail.
• loop-tools-test — deterministic Grow/Sharpen tool tests; adapter smokes, live audits, and adequacy campaigns remain explicit.
• smoke-live — overlay↔mpv breakage the fakes can't catch.
• hygiene / perf-risk — advisory: triage via the repowise navigator, don't fix blind.

Env traps (these bite)

• Never uv run --python 3.13 against the default env — it recreates .venv as 3.13 and clobbers the free-threaded 3.14t build. Run the 3.13-pinned tools through their poe task (poe invariants-taint / poe fuzz / poe crosshair) — each pins its own 3.13 env, and [tool.poe.tasks] is the SSOT for how, so never hand-roll the interpreter.
• cosmic-ray re-enables the GIL via SQLAlchemy in its own harness — expected, not a regression (the test subprocess still runs free-threaded).
• Don't run complexipy <narrow-path> for a spot check — it silently overwrites the full-repo snapshot baseline scoped to just those paths. Use poe complexity.

Verify

The gate exits 0. If a single task is red, run it alone (e.g. uv run poe types) to iterate, then re-run the gate.