# What the checker supports, what it will not, and why Status on 2026-09-16 (`just verify-checker-soundness`): ``` TP err+flag : 2665 (of which via parse rejection: 390) MISS in scope : 50 (the backlog — this one can reach zero) OUT OF SCOPE : 19 (declared in scripts/checker_out_of_scope.txt) FP ok +flag : 0 (soundness bugs — TS7 accepts these) PFLEGAL : 0 (parser rejects TS7-legal files — parser bugs) TN ok +quiet : 1750 ``` This document is the STRATEGY half: what the gate reports and why it is two numbers, which tier each remaining family sits in, how often each feature appears in real code, and what is declared out of scope. `src/checker/UNSUPPORTED.md` is the other half — every remaining MISS file with tsc's own diagnostic, a capability probe of the common shape of each feature, and the declared abstentions with their legal neighbour. `docs/checker-priority.md` is the TS6-era document whose conclusion ("incremental recall wins are exhausted") was refuted by measurement and is kept only as history. ## 1. Why the gate reports two MISS numbers A single `MISS` figure summed work worth doing now with files nobody should ever fix, so it could answer neither "are we done?" nor "what next?". The oracle therefore splits the bucket: 1. `scripts/checker_out_of_scope.txt` holds the declared remainder, one path per line with a kind and a reason. Every entry was decided by OPENING the file, not by its directory or its name — the first pass keyed on both and came out twice too large. 2. Only the MISS branch consults that file. A listed file can still be a TP, an FP or a PFLEGAL, so being out of scope withholds a rule and never excuses a wrong answer. The FP budget is untouched at 0. 3. The oracle reports a **STALE** entry the moment a listed path stops being a MISS, which is what keeps the file from becoming a suppression list. 4. `--max-miss` gates the in-scope number. That closes a direction nothing used to watch: a rule that stops firing moves a file from TP to MISS and every other number absorbs it silently. The budget is lowered whenever a batch improves it and never raised — it caught two accidental TPs on its first real use (batch DO). The asymmetry is deliberate. A MISS is expected because we model a subset of TypeScript; an FP never is. ## 2. The 50 in-scope MISS files, by machinery From `node scripts/checker_miss_rank.mjs` over the oracle's `--miss-list` — every file run through the local compiler under its own `// @option:` header and grouped by what a rule would need. Grouped by MACHINERY and not by error code, because a code is not a difficulty class: TS2322 is the largest bucket at 16 files and is fourteen unrelated causes, and 27 codes have exactly one file each. | family | files | tier | the one-line reason | |---|---|---|---| | generic inference and generic assignability | 12 | 3 | type arguments inferred from arguments, then checked against the rest of the call; generic signatures compared with each other | | object literals, contextual typing, widening | 10 | 3 | union normalization on widening, spread of a union, `(x \|\| {}).a`, tagged-template substitutions, return-type inference from a body | | classes, `this`, mixins | 8 | 3 | mixin over an intersection or type-variable base, polymorphic `this` as a value type, `typeof this.x` in a type position | | type-level machinery | 7 | 3 | variadic tuples, template-literal placeholders, TS2536 key checks, a conditional left unresolved inside `&` | | name resolution and declarations | 5 | 3 / declared | three of the five are recorded rejections or abstentions with a probe; `import a = A` needs the alias target resolved | | narrowing and control flow | 4 | 3 | aliased guards (fail direction is an FP), `typeof x === "object"` to `object \| null`, narrowing to `never`, definite assignment | | lib and host shapes | 4 | 3 | `Intl` members by lib version, the `Promise` constructor's callback parameter, structural `Generator` comparison | Every row is Tier 3, and that is the honest reading of MISS 50: **there is no Tier 1 or Tier 2 left.** Tier 1 (conditional through a generic alias, the utility-type table, overload selection, computed `unique symbol` keys as a cluster) was taken in batches DI–DM and EB, or opened and found to be eleven unrelated codes rather than a feature. Tier 2 (the grammar and declaration rules, the implicit-any family, the strict-null bucket) was taken in batches DN through ET. The thirty files between MISS 80 and MISS 50 took six batches and roughly one rule per file, which is the rate this tier now has. ### Tier 3 — DEFER: real but expensive All 50. Every one is a genuine TypeScript behaviour and none is reachable without machinery we would have to build — real generic inference, contextual typing, spread-type computation, a flow graph, lib interface merging. Do not take these for the MISS count. Take an individual file only when a real bridge input or `mtsc` target demands it, and record which one did; the measured rate here is about one file per investigation. Six of the 50 carry a **recorded rejection or abstention with a probe**, so they should not be re-attacked as written (details in `UNSUPPORTED.md` §1 and §3): `inferTypesInvalidExtendsDeclaration` (the parser reduces the conditional before the checker sees it), `await_incorrectThisType` (structural assignability of a phantom type parameter makes the naive rule unsound), `stringLiteralTypeIsSubtypeOfString` (173 lib interfaces are declared empty), `symbolProperty3` (needs the `Symbol` constructor as a value type, for a spelling nobody writes), `localTypes4` (a block-local type declared twice is deliberately left unregistered) and `intersectionWithIndexSignatures` (the deciding arm is never reached, measured). ### Tier 4 — WON'T SUPPORT: 19 files, declared out of scope Checked in as `scripts/checker_out_of_scope.txt` under six kinds: - **removed-feature** (4): `import x = module("m")`, and three files whose only error is TS6053 on a `///` — program construction, which the checker has no notion of. Two of those were TPs by ACCIDENT until batch DO fixed the `+=` string-building false positive that had been flagging them. - **malformed** (3): the parser's error-recovery corpus (`x: break`, `x: public`, `[public a]`). Taking this cluster for its size is fitting the corpus. - **lexer-adversarial** (1): `/**// asdf /`, written to defeat regex-versus-comment disambiguation. - **resource-mgmt** (5): `using` / `await using`, measured at ZERO occurrences across 5,697 real files (§3). Only files whose DIAGNOSTIC needs the feature are here — `usingDeclarationsWithObjectLiterals2` is named for `using` and errors on `value: null`, and is a TP since batch EC. - **no-local-oracle** (5): TS7 errors and the local compiler 6.0.3 accepts, so there is no compiler to probe a legal neighbour against. A HARNESS limit, not a judgement about the language — revisit when the local compiler moves. - **lib-diagnostic** (1): every error is inside `lib.es5.d.ts`, reached through an `interface Object` augmentation under `@skipDefaultLibCheck: false`. Found because the probe used to attribute OTHER files' diagnostics to the probed file. ## 3. How often each feature appears in real code A corpus count is not real-world frequency. The second axis is measured against real input: 3,000 `.d.ts` files from `node_modules` (what the bridge consumes) and 2,697 `.ts` sources from the type-aware corpus (what `mtsc` consumes). Measured once for the first version of this document; the corpus has not changed since. | feature | real `.d.ts` | real `.ts` app | |---|---|---| | conditional type | **345** | 24 | | `keyof` | 194 | — | | `unique symbol` | 183 | — | | `infer` | 156 | — | | mapped type | 143 | 31 | | `this` return type | 115 | — | | index signature | 102 | — | | computed `[Symbol.x]` key | 72 | — | | template-literal type | 57 | 151 | | variadic tuple | 20 | — | | `satisfies` | — | 238 | | optional chaining | — | 144 | | `as const` | — | 108 | | **`using` / `await using`** | **0** | **0** | | **decorator** | **1** | **0** | Two rows decide two tiers on their own. `using` declarations occur in zero real files and are five of the out-of-scope entries. Decorators occur in one `.d.ts` and no application source — a property of THIS corpus, which contains no Angular, NestJS or TypeORM, so that row is a caveat and not a verdict; the decorator rules that did ship (batches CU, EG) were taken because they were the applied-in-some-places family, not for this row. Against that table, the capability probe in `UNSUPPORTED.md` §2 says which features are still BLIND at the shape real code writes: variadic tuples, a template-literal placeholder, a computed `unique symbol` key, an anonymous index-signature READ, and a strict-null member-chain receiver (the last one deliberate). Conditional types, the utility table, mapped types, `keyof`, overload selection and generic inference are all CAUGHT at the common shape now — the four rows the first version of this document called blind "at the shape real code actually writes". ## 4. What the conformance number cannot see Three things, each with its own gate: - **False positives on real code.** `FP 0` is a statement about 4,484 conformance files. Batch EO measured `mtsc` on `zod@4.4.3` and found 272 diagnostics with tsc accepting all of them; 85 were one missing shadowing test and 187 remain (`TODO.md`, batch EO). A real-package FP gate is filed and not yet built. - **False positives on legal neighbours.** The corpus scores a file as a TP if we flag it AT ALL, so a wrong rule firing beside a right one is invisible. Batches EH, EM and EP–ET fixed eleven such false positives (five, two and four), every one found by probing a hand-written legal spelling or by a unit test and none by the corpus. - **Cost.** A rule quadratic in a module-wide list is invisible to a corpus of few-dozen-line files; `just verify-checker-scaling` fits an exponent per axis and fails on growth. ## 5. Where the history is Every batch from DZ onward is recorded in `TODO.md` under its own heading with the before/after numbers, the rules shipped, the false positives fixed and the rejections with their measurement. The summarized lessons — the applied-in-some-places family, the absent-versus-`: any` blocker, the label-for-objective substitution, probing a recorded abstention's REASON — are in `CLAUDE.md`'s `src/checker` section.