fix: unbreak main by baselining the M4 view models until their caller… #327

Workflow file for this run

	name: CI

	on:
	push:
	branches: [main]
	pull_request:

	concurrency:
	group: ci-${{ github.ref }}
	cancel-in-progress: true

	jobs:
	test:
	runs-on: ubuntu-latest

	services:
	postgres:
	image: postgres:16
	env:
	POSTGRES_USER: sync
	POSTGRES_PASSWORD: sync
	POSTGRES_DB: sync
	ports:
	- 5432:5432
	options: >-
	--health-cmd "pg_isready -U sync"
	--health-interval 2s
	--health-timeout 3s
	--health-retries 15

	env:
	# docker-compose.yml maps Postgres to 5433 to stay clear of a local server.
	# A service container has the port to itself, so the default DSN is overridden
	# rather than the mapping reproduced. The `serial` job below does the opposite,
	# and the comment there says why the choice is not free.
	SYNC_DSN: postgresql://sync:sync@localhost:5432/sync

	steps:
	- uses: actions/checkout@v4

	- name: Install uv
	uses: astral-sh/setup-uv@v5
	with:
	enable-cache: true

	# `.oasdiff-version` is the pin, and `scripts/bootstrap_tools.sh` reads that same file, so
	# a job here and a developer's checkout cannot drift apart. The last three lines assert
	# rather than print: a version that is only echoed is a version nobody checks, which is
	# how a pinned job and seven working copies came to run two different builds.
	- name: Install oasdiff
	run: \|
	version="$(grep -m1 -Ev '^[[:space:]]*(#\|$)' .oasdiff-version)"
	version="${version//[[:space:]]/}"
	mkdir -p tools
	curl -fsSL -o /tmp/oasdiff.tar.gz \
	"https://github.com/oasdiff/oasdiff/releases/download/v${version}/oasdiff_${version}_linux_amd64.tar.gz"
	tar -xzf /tmp/oasdiff.tar.gz -C tools oasdiff
	installed="$(tools/oasdiff --version)"
	if [ "$installed" != "oasdiff version ${version}" ]; then
	echo "pinned v${version} in .oasdiff-version, installed '${installed}'" >&2
	exit 1
	fi
	echo "$installed"

	- name: Sync dependencies
	run: uv sync --all-extras --dev

	# Runs before the suite: a failure here is a fact about the source, and needs no
	# database, no binary, and no test run to establish.
	- name: Encoding lint
	run: uv run python scripts/lint_encoding.py src scripts tests

	- name: Import boundary
	run: uv run lint-imports

	# Being tested and being reachable are different properties, and until this step
	# existed only the first was checked -- four finished components shipped with no
	# caller anywhere in src/. The baseline lists the ones already accepted and only
	# shrinks: an entry that no longer violates fails here until it is deleted.
	- name: Dead links
	run: uv run python scripts/lint_dead_links.py src --baseline scripts/dead_links_baseline.txt

	- name: Tests
	run: uv run pytest

	# Recorded, never gated. `2026-07-27-sync-benchmark-gates.md` forbids inventing a
	# threshold -- a gate at an invented number either fires constantly and gets disabled or
	# never fires and provides false assurance -- and a percentage that fails a build is a
	# percentage people write tests to satisfy rather than to test something.
	#
	# The not-gating is deliberate rather than defensive and it stays: a coverage number must
	# not turn a green suite red. What the `\|\| true` that used to carry it could not
	# distinguish is a number that moved from a run that never produced one, and the exit code
	# it discarded is the only place that distinction lives. No `--cov-fail-under` is passed
	# here, so coverage cannot produce an exit code of its own at all -- every non-zero value
	# is pytest's, and 2 through 5 all mean no number was measured. `\|\| true` therefore bought
	# nothing against the risk its own comment named and cost the ability to see a collapsed
	# run, so the two are separated below instead.
	#
	# It is worth saying what this did not hide, because
	# `2026-07-30-n0-is-broken.md` records that it did: `addopts` applies to this step too, so
	# it has always run `-n auto`, never the serial configuration those 186 errors need. And
	# `SYNC_DSN` above is pinned, which sends every database this job creates outside
	# `conftest.LEAKED_DATABASE_PATTERN` under either scheduler. Nothing in CI hid that
	# defect; nothing in CI could reach it. The `serial` job is what reaches it.
	#
	# The baseline this is measured against is
	# `docs/superpowers/specs/2026-07-29-sync-coverage-baseline.md`, which also records what
	# this number cannot see -- seven components have shipped here fully covered and reachable
	# from nothing, and every one of them looked healthy to a line count.
	- name: Coverage (recorded, not gated)
	run: \|
	status=0
	uv run pytest -q --cov=sync --cov-report=term-missing:skip-covered \|\| status=$?
	# 0 is a clean run and 1 is a failing test. Neither is this step's business: `Tests`
	# above is the gate on the suite, and a dotted `--cov` legitimately changes behaviour
	# in this repository (`2026-07-29-psycopg-error-identity.md`).
	if [ "$status" -gt 1 ]; then
	echo "the coverage run did not complete: pytest exited $status" >&2
	exit "$status"
	fi

	# The first gate on a quality number rather than on correctness, and the only one
	# `2026-07-27-sync-benchmark-gates.md` allows today: a directional floor on a
	# deterministic axis. The floors are not here. They are in `scripts/gate_corpus.py`,
	# beside the figure each one guards and the argument for guarding it, because a floor in
	# a workflow file is a string a reader meets without the number it came from and lowering
	# one is a config tweak rather than a reviewable act.
	#
	# No `\|\| true`. The coverage step above is recorded and this is gated, and the difference
	# between them is the whole of what this step adds.
	- name: Fetch the frozen corpus
	run: uv run python scripts/fetch_corpus_repositories.py

	# A database of its own. `score_corpus.py` truncates the graph once per pair, so pointing
	# it at the one the suite used would delete what a later step might read.
	- name: Score the frozen corpus
	run: \|
	uv run python -c "
	import psycopg
	from psycopg import sql
	with psycopg.connect('postgresql://sync:sync@localhost:5432/postgres', autocommit=True) as c:
	c.execute(sql.SQL('DROP DATABASE IF EXISTS {} WITH (FORCE)').format(sql.Identifier('sync_benchmark')))
	c.execute(sql.SQL('CREATE DATABASE {}').format(sql.Identifier('sync_benchmark')))
	"
	uv run python scripts/score_corpus.py \
	--score-dsn postgresql://sync:sync@localhost:5432/sync_benchmark \
	--json corpus-score.json

	- name: Binding floors over the frozen corpus (gated)
	run: uv run python scripts/gate_corpus.py --score corpus-score.json

	# The scheduler `addopts` never selects, run in the one configuration where it can fail.
	#
	# `-n0` is contracted here rather than a debugging convenience: `pyproject.toml` names it for a
	# focused run, and `2026-07-29-sync-verification-regime.md` names it as the way to run the e2e
	# test at all, because `addopts` deselects that test and pins the scheduler in one string.
	# Nothing verified it, and it was broken on `main` for an unknown length of time -- 186 errors
	# from one cause, `docs/superpowers/reports/2026-07-30-n0-is-broken.md`. The whole suite rather
	# than the database-touching part of it: the defect class is one test breaking a later test in
	# the same process, so any rule that picks a subset picks victims while the breaker is whatever a
	# new test brings. `2026-07-30-ci-does-not-run-serial.md` carries the cost this buys with.
	#
	# A job of its own rather than a step in `test`, because it needs the opposite database
	# configuration and that is the whole of why it can fail. No `SYNC_DSN` here, deliberately.
	# `conftest.database_for` returns None for a serial run handed a pin, so no per-run database is
	# created and the sweep this job watches has nothing of the run's to take -- and under `-n auto`
	# a pin yields `sync_gw0`, which is outside `LEAKED_DATABASE_PATTERN` as well. Measured on the
	# broken tree: the recipe that reproduces the defect in 33 s unpinned is green under `test`'s
	# pin. A serial step inside `test` would have stayed green through all 186 errors, and
	# `tests/test_ci_runs_the_serial_scheduler.py` refuses a pin here for that reason.
	serial:
	runs-on: ubuntu-latest

	services:
	postgres:
	image: postgres:16
	env:
	POSTGRES_USER: sync
	POSTGRES_PASSWORD: sync
	POSTGRES_DB: sync
	ports:
	# `conftest.DEFAULT_DSN`, reproduced rather than overridden. `test` overrides the DSN
	# instead and is right to -- but that override is exactly what hides the defect this
	# job exists for, so here the mapping moves and the DSN does not.
	- 5433:5432
	options: >-
	--health-cmd "pg_isready -U sync"
	--health-interval 2s
	--health-timeout 3s
	--health-retries 15

	steps:
	- uses: actions/checkout@v4

	- name: Install uv
	uses: astral-sh/setup-uv@v5
	with:
	enable-cache: true

	# Duplicated from `test` rather than shared, and the version cannot drift with it: both read
	# `.oasdiff-version`, so what is copied is the mechanism and not the number.
	# `tests/test_oasdiff.py` calls the binary with no skip guard, so a job without it errors
	# rather than quietly running a smaller suite -- which is why this is here and not omitted.
	- name: Install oasdiff
	run: \|
	version="$(grep -m1 -Ev '^[[:space:]]*(#\|$)' .oasdiff-version)"
	version="${version//[[:space:]]/}"
	mkdir -p tools
	curl -fsSL -o /tmp/oasdiff.tar.gz \
	"https://github.com/oasdiff/oasdiff/releases/download/v${version}/oasdiff_${version}_linux_amd64.tar.gz"
	tar -xzf /tmp/oasdiff.tar.gz -C tools oasdiff
	installed="$(tools/oasdiff --version)"
	if [ "$installed" != "oasdiff version ${version}" ]; then
	echo "pinned v${version} in .oasdiff-version, installed '${installed}'" >&2
	exit 1
	fi
	echo "$installed"

	- name: Sync dependencies
	run: uv sync --all-extras --dev

	# Gated, and the only gate here. No lints and no corpus: those are facts about the source
	# and about a frozen input, neither of which a scheduler can change, and running them twice
	# buys a second copy of the same answer.
	- name: Tests, serial scheduler
	run: uv run pytest -q -n0

Navigation Menu

fix: unbreak main by baselining the M4 view models until their caller… #327

Workflow file

fix: unbreak main by baselining the M4 view models until their caller… #327

Uh oh!

Workflow file for this run