feat(ci): typecheck lexicon tests, scripts and examples, against a shrinking baseline (#1388)

[lex00]

Part of #1388 — the gating half. Working the backlog down is the other half.

The gap

CI typechecks packages/core and packages/k8s-client, and check-lexicons builds each lexicon's tsconfig.build.json — which excludes **/*.test.ts, correctly, since a published lexicon should not ship its tests. Nothing covered what fell between:

$ tsc -p lexicons/forgejo/tsconfig.build.json --listFiles | grep -c '\.test\.ts'
0
$ tsc -p packages/core/tsconfig.json     --listFiles | grep -c '\.test\.ts'
266

A type-level assertion written in a lexicon test was evaluated by nothing. That is why #1344's ChantConfig augmentation proofs had to live in compiled source instead of the test where they belong — and the next person would reach for the test file first.

Why a ratchet rather than switching it on

I filed #1388 estimating this at 56 example errors. Measured properly it is ten times that:

Population	Error lines	Covered today
packages/** (src and tests)	0	yes
lexicons/*/src non-test	0	yes
lexicons/*/src tests	405	no
lexicons/*/scripts	7	no
examples	546	no
total	958 across 118 files

A check that cannot pass is a check nobody runs. So this follows KNOWN_FAILURES in check-lexicons.ts: everything failing today is listed, and the list can only get smaller.

• a file not in the baseline that fails → error, reported by name and line
• a file in the baseline that now passes → error, with instructions to delete its line

Verified both directions:

$ npx tsx scripts/typecheck.ts        # after adding a broken test
1 file(s) newly failing the repo-wide typecheck:
  lexicons/forgejo/src/__ratchet-probe.test.ts
    (1,7): error TS2322: Type 'string' is not assignable to type 'number'.

$ npx tsx scripts/typecheck.ts        # after baselining a file that passes
1 baselined file(s) now typecheck cleanly:
  lexicons/forgejo/src/config.ts
  Remove them from scripts/typecheck-baseline.json …

Note on moduleResolution

bundler in this config, node everywhere else. Deliberate: tests and examples run under vitest and tsx, which resolve like a bundler. Shipped code keeps its node-resolution build configs, so nothing about the published packages' resolution story changes. Under node the same run reports 763 lines, a large share of them resolution artifacts rather than real type errors — which would make the baseline mostly noise.

For whoever takes the backlog

Concentrated by lexicon: github 118 lines, gcp 105, temporal 81, aws 51, k8s 37. One lexicon at a time drops the baseline in reviewable chunks.

Verification

• npm run typecheck → 118 known failure(s), no regressions
• New CI step in the check job, after the lexicon completeness contract

🤖 Generated with Claude Code

https://claude.ai/code/session_01Jp6SP7ZiavhoEB1AFuwwoJ