[P2] objectui: every package type-checks its tests or declares why not — and guards must prove they can fail (measured: 355 errors across 19 packages)

[os-zhuang]

Generalizes objectui#3009 from one package to the repo, with the scope measured rather than estimated. Companion to #4115 (which prevents new forks; this one makes existing guards actually run).

Why

objectui#3009 found that packages/types' spec-derivation guards — the satisfies checks meant to stop a vocabulary re-fork — were compiled by nothing: package tsconfigs exclude test files (correctly, they are build configs), and no other tsc invocation reads them. Reverting a derived alias produced zero errors; after #3009, TS1360 on the exact line.

Tests are where agents encode their understanding of a contract. Unchecked tests let wrong understanding accumulate silently and then read as evidence: p2-spec-exports.test.ts type-imported eight names whose absence its own sibling guard asserts, green the whole time; p1-spec-alignment.test.ts passes six "should accept ⟨shape⟩" claims the type rejects. An agent citing a green suite as proof of a contract is the #2901 backwards-premise incident again.

Measured scope (2026-07-30, objectui main @ a045a32)

Method: per package whose tsconfig excludes test globs, a temp tsconfig lifting only the test exclusion (emit off), tsc, errors classified config-tier (fixable in the test tsconfig: paths/rootDir leakage TS6059, lib too low TS2550, missing types: ["node"], cross-package dist TS2307) vs code-tier (real).

package	test files	code-tier errors	dominant codes
core	54	80	TS2741×32, TS2322×17, TS2304×10
react	28	76	TS2339×49, TS2769×9
sdui-parser	3	50	TS2304×47 (likely env globals — verify)
app-shell	240	37	TS2339×8, TS2345×6, TS2739×6
plugin-list	12	20	TS2739×6, TS2740×6, TS2741×5
plugin-form / i18n / permissions / auth	25/16/3/8	15/13/13/10	mixed
plugin-dashboard, components, plugin-kanban, plugin-detail, plugin-calendar, plugin-gantt, plugin-map, layout, plugin-chatbot, plugin-grid	—	8…2 each	mixed
mobile, plugin-designer, plugin-markdown, plugin-report, plugin-timeline, plugin-tree, react-runtime, runner	17	0	already clean

Totals: 27 packages, 612 test files, 355 code-tier errors in 19 packages; 8 packages are free wins. (5 further packages — fields, cli, data-objectstack, plugin-charts, plugin-editor — already include tests in tsc --noEmit and are covered today.)

The two dominant codes are not random: TS2741/TS2739 ("missing required properties") is the input-vs-output confusion #3009 fixed in p2-spec-exports (authoring fixtures typed as parsed output whose .default() fields are required), and TS2339/TS2353 ("property does not exist") is the implementation-wider-than-type dialect problem (#4074's sharing { visibility, enabled } shape). The per-package cleanup will mostly be those two playbooks applied repeatedly.

The ask

1. Coverage ratchet. Extend the check-type-check-coverage.mjs pattern (already the repo's best governance: declared, reasoned, shrink-only, #2911): every package either type-checks its tests (a tsconfig.test.json chained from type-check, per objectui#3009's template) or carries a DEBT entry with the measured error count and an issue link. Start by converting the 8 clean packages (pure plumbing, zero fixes), then burn down the table above.

2. Discrimination proof for guards. A guard PR shows the guard failing against the pre-fix code (one pasted run in the PR description). This is what exposed the inert satisfies checks, an already-vacuous "prefer canonical over alias" test (objectui#2990), and the tautological "should accept" suite — three cases in one day of applying it. Cheap for agents to execute mechanically; catches the exact failure mode ("test passes for the wrong reason") that review reliably misses.

3. Config-tier template. The per-package tsconfig.test.json should standardize: noEmit, composite: false, paths: {} (kills the TS6059×100 source-leak noise), lib matching what tests actually use, types: ["node"] where tests touch fs. objectui#3009's packages/types/tsconfig.test.json is the reference, comment included.

Same meta-rule as #4115: this lands as CI, not prose.

Refs objectui#3009, #4115, #4074, objectui#2911, #2901.

[os-zhuang]

Asks 1 and 3 landed in objectui#3032 (62311b6); ask 2 was applied as the working method throughout that PR (each of its three guards ships a pasted run failing against pre-fix code).

12 free wins, not 8. Four more fell out of a better config template than the measurement harness used: plugin-calendar and plugin-kanban needed only the jest-dom matcher types, providers was already clean, and sdui-parser's 50 TS2304s were env globals — exactly as this issue suspected ("likely env globals — verify"), so they were config-tier, not the second-largest code-tier debt in the table.

Wiring sdui-parser up surfaced a real bug, which is the best argument for the ratchet itself: render.test.tsx imports @object-ui/core and @object-ui/react with neither declared in its package.json — phantom dependencies resolved only through the root tsconfig's paths alias and vitest, never as a dependency a consumer would get. Both are devDependencies now. That is also what makes paths: {} in the template load-bearing rather than cosmetic: with the alias in place, an undeclared workspace import still compiles.

TEST_DEBT is re-measured, not copied. Against main at branch time: 240 code-tier errors across 15 packages. That is a fresh measurement rather than a subtraction from the table above — main moved since a045a32, and sdui-parser's 50 reclassified as config-tier. A package with test files now either compiles them or carries an entry with its measured count.

The guard grew a third failure mode worth noting here, because it's this thread's own shape: it fails on a tsconfig.test.json that exists but that the package's type-check script never runs. That is objectui#3009 exactly — a file that looks like enforcement while nothing compiles it. It also rejects a package listed in both DEBT and TEST_DEBT, since that's one gap counted twice; layout was caught that way during the sweep.

Remaining: the 240. The two dominant codes are the two playbooks this issue predicted, so it should stay mechanical — core (80) and react (76) last, after the smaller packages have worn the grooves in.

---

Generated by Claude Code