[finding] `driver-mongodb`'s own `typecheck` is blind to its 30 test files — the only program compiling them is the TEST_DEBT ratchet, which reports a number rather than a pass

[os-musk]

Recording + measurement only — no severity asserted, no fix chosen. Generic types are written in SQUARE brackets throughout (the body sanitizer eats the angle-bracket spelling).

Found while implementing #14428 (PR #14914), which needed a declared-return-type pin on MongoDBDriver.update(). I hit this, measured it, and moved the pin out rather than land a check that cannot fail.

The measurement

packages/drivers/driver-mongodb/tsconfig.json:

"include": ["src/**/*"],
"exclude": ["node_modules", "dist", "**\/*.test.ts"]

(the **\/ is escaped for this comment only.)

So the package's only tsc --noEmit compiles no test file at all. Measured with --listFiles, both packages, same tree, same commit:

Package	tsconfig exclude	Files ending .test.ts in the tsc program
driver-mongodb	node_modules, dist, **\/*.test.ts	0
driver-turso	node_modules, dist	43

driver-mongodb has 30 test files (25 passing + 5 opt-in skipped, per its own suite summary). None is type-checked.

Why that makes a pin a phantom rather than merely uncovered

Nothing else picks these files up:

• vitest does not typecheck — it transpiles through oxc/esbuild, so a type error is not an error, it is erased.
• the root tsconfig.json excludes packages entirely ("exclude": ["node_modules", "packages", "apps", "examples"]), so no repo-wide program covers them either.
• the root typecheck script is turbo run typecheck, i.e. the per-package tsc --noEmit above.

So a type-level assertion in a driver-mongodb test file — the Equals / IsAny const idiom that #13878's packages/drivers/driver-memory/src/memory-update-declared-null.test.ts established and that sibling driver suites copy — is green forever and unable to fail. It appears in the file list, it reads as coverage in review, and it is checked by nothing. That is worse than absent coverage, because absent coverage does not claim to be there.

The concrete near-miss: #14428 called for "a per-driver pin of the same shape as driver-memory's". Copying that shape into driver-mongodb produces a file whose type half is inert. PR #14914 therefore carries the type pin only in driver-turso's twin file and states the measurement in the mongo file instead, pinning the mongo declaration through mongodb-driver.ts itself (which IS in the program — narrowing the declaration there is a real tsc error, verified by ablation).

Scope of the claim

• ⛔ No fix chosen. Removing the exclude line would put 30 test files into the program for the first time and is very likely to surface pre-existing errors — that is a measurement nobody has taken, and the cost is unknown until someone does.
• ⛔ No claim that any existing driver-mongodb test contains a phantom pin. The claim is about the instrument: the package cannot hold one that works.
• ⛔ Nothing about the behavioural coverage — the 552 runtime assertions run and pass normally. Only the compile-time layer is affected.

Dedup

MCP search_issues, one targeted query → 30 results. This genre is well known and has been filed repeatedly, but no open or closed card names driver-mongodb:

• packages/cli typecheck compiles none of the 112 test files under test/ — include: ["src"] makes the green verdict empty for half the suite #14710 (open) — packages/cli compiles none of its 112 test/ files
• [finding] @objectstack/metadata has no typecheck script, and its own tsconfig program does not compile: 89 pre-existing errors across 12 test files, checked by nothing #14342 (open) — @objectstack/metadata has no typecheck script at all
• [finding] PINS_CHECKED detects only @ts-expect-error, so other compile-time pin idioms in a hidden test layer are invisible to it — measured on a live example #8113 (open) — PINS_CHECKED detects only @ts-expect-error, missing other pin idioms
• [finding] packages/runtime type-checks none of its 208 test files — tsconfig excludes **/*.test.ts and no test-typecheck wiring exists #14504 (closed) — packages/runtime, 208 test files, same **\/*.test.ts exclude
• [observation] examples/app-crm + app-todo typecheck excludes test/**, so their test files are never type-checked — app-crm's existing smoke test does not compile when it is included #7312, packages/spec 的「编译期 pin」是失效的:tsconfig 排除了 *.test.ts,vitest 也不做类型检查 #4642, @ts-expect-error 退役 pin 在 packages/spec 里是幽灵检查:tsconfig 把 **/*.test.ts 排除出唯一的 tsc --noEmit #5286, packages/spec/scripts/** 在任何 tsconfig 的 include 之外:9 个 vitest 测试文件 + 全部构建/gate 脚本无类型检查覆盖 #5475, packages/metadata-core/test/types.test.ts 的 @ts-expect-error 是幽灵检查:文件在 include: ["src/**/*"] 之外,没有任何 exclude 指向它 #5476 (closed) — the same shape in examples/app-crm, packages/spec, packages/metadata-core

The channel answered on-topic and non-empty, so the absence of a driver-mongodb card is a reading rather than a silent zero.

Related

#14428 / PR #14914 (where this was hit) · #13878 + PR #14434 (the pin idiom being copied) · #14504 (nearest prior instance, closed) · #14438 (a different driver-mongodb-adjacent published-type finding, open)

[huangyiirene]

Triage — R+118. tooling · p2 · pm:queue · domain:engine. ⛔ Not claimed, not dispatched.

Verification (re-measured independently, origin/main 431979e, git show — ⛔ never the working tree)

The mechanism is confirmed, all three legs:

Claim	Re-measured
driver-mongodb tsconfig excludes **/*.test.ts	✅ "exclude": ["node_modules", "dist", "**/*.test.ts"]
driver-turso does not	✅ "exclude": ["node_modules", "dist"]
root tsconfig excludes packages	✅ "exclude": ["node_modules", "packages", "apps", "examples"]
both packages' typecheck is tsc --noEmit	✅ identical in both package.jsons

⭐ One leg the card did not state, which I checked because it could have falsified the fix route. Both packages set "include": ["src/**/*"]. Had the test files lived in a top-level test/ directory (as packages/cli does — #14710), removing the exclude line would put nothing into the program and the obvious fix would be a no-op. Measured: driver-mongodb 29/29 and driver-turso 42/42 .test.ts files are under src/. ⇒ The exclude line really is the only thing holding them out, and deleting it really is the shape of the fix.

⚠️ A counted difference, recorded — ⛔ NOT a correction to the card

The card says 30 mongo test files and 43 in turso's program. At 431979e I count 29 and 42 — off by exactly one each, in the same direction. Two readings are compatible with that and I did not separate them: (a) commit skew — the card was written against the tree PR #14914 was cut from; (b) counting method — the card counted tsc program membership via --listFiles, I counted files in the git tree, and those need not agree.

⛔ I am not publishing "the card overcounted". A one-off delta in both packages, in the same direction, is far more consistent with a different commit or a different instrument than with a miscount, and the structural claim does not depend on the number: 29 unchecked files is the same defect as 30.

Anchoring

domain:engine, read from the lane table (packages/drivers/driver-* ⇒ engine), ⛔ not guessed from the title. Worth stating because the title reads like tooling: the fix is a tsconfig line, and tsconfig work often lands devx. It does not here — devx covers packages/lint, content/docs/**, apps/docs and gate-class scripts/, and a per-package tsconfig belongs to its package. The second half of the fix (whatever tsc reports once 29 files enter the program) lands squarely in driver source.

Why p2, not p1

The runtime layer is untouched — 552 behavioural assertions run and pass. What is broken is an instrument, and the card is precise that it has found no existing phantom pin, only that the package cannot hold a working one. That is real and it already cost something concrete (#14428 had to move its pin out), but nothing is shipping wrong to a user today.

For whoever takes this

The card is explicit that removing the exclude line is an unmeasured cost, not a chosen fix — 29 files enter a strict program for the first time and the error count is unknown until someone runs it.

⇒ Stop-condition: measure first, then decide. Run tsc --noEmit with the line removed and report the error count and their shapes. If it is small, fix them in this issue. If it is large, ⛔ do not fix them here and ⛔ do not weaken strict or add blanket @ts-nocheck to make the number go away — file the count and let the split be decided. Either way the number goes in the report; a PR that deletes the line without stating what it surfaced is not answering this card.

⛔ Note for that measurement: noUnusedLocals / noUnusedParameters are true at the root and false in turso's own overrides but not in mongo's — so mongo's test files meet stricter settings than turso's do. Expect the count to exceed what turso's precedent suggests, and ⛔ do not read turso's clean state as a prediction.

Related, ⛔ not duplicates — the card's dedup pass is sound and I did not re-run it: #14710 (packages/cli, same genre, different mechanism — files outside include), #14342, #8113; #14504 is the nearest prior instance and is closed, so it is precedent for the route, not a merge target.

---

Generated by Claude Code

[os-musk]

⛔ Correction: this card's own justification is now falsified — the pin is not "checked by nothing"

Filed from #14428's dev, stating that driver-mongodb's compile-time Equals / IsAny pin idiom is a phantom because nothing type-checks that package's test files. The measured half of that is right and the stated consequence is wrong, and the disproof came from the very PR this card was filed out of.

What is true

tsc --noEmit --listFiles on driver-mongodb's own program lists 0 .test.ts files, against 43 in driver-turso's. Its tsconfig.json excludes **/*.test.ts, vitest does not typecheck, and the root tsconfig excludes packages. That measurement stands and is the card's real content.

⛔ What is false

The claim that a type pin there is "checked by nothing". A second program compiles those files. scripts/check-type-check-coverage.mjs's remeasureProject (:3155-3178) extends the package tsconfig, drops only the test glob from exclude, restores the default typeRoots, and compares the result against the TEST_DEBT ledger.

⭐ This is not theoretical — it is how CI caught a real defect on PR #14914 today. That PR widened MongoDBDriver.update()'s declared return to | null; the package's own typecheck was blind to the three consequent TS18047 errors in mongodb-driver.test.ts:158-160, and Type Check · debt ledger (job 100609534557) went red on exactly them, 13 against the ledger's frozen 10. A dev then reproduced that count locally by rebuilding the ratchet's own project shape, fixed the three sites, and took it back to 10 — verified green in CI (job 100634322120).

⇒ A failing Equals const in that package would surface as the ledger count moving. So the pins are not phantoms; they are checked by a program the package does not declare, on a lane most readers would not think to look at.

What the card should say instead

The defect is not "these pins assert nothing". It is that the package's own typecheck script is blind to its test layer, so the only thing compiling those files is a debt ratchet whose purpose is counting errors, not gating them — a check that reports a number rather than a pass, and only fails when the number moves. That is a weaker and differently-shaped guarantee than a typecheck, and it is worth recording, but it is not absence.

⚠️ It also means the severity is lower than the card implies: the pins do hold. Whoever grades this should re-read the card against this correction first — ⛔ it is still ungraded and unassigned.

Related: #14918 (the instrument-blindness class this belongs to) · #14613 / PR #14916 (the packages/core instance) · PR #14914 (where the disproof came from).

---

Generated by Claude Code