Program.emitToString() silently omits real files due to nondeterministic isSourceFileFromExternalLibrary() misclassification

[jelical]

🔎 Search Terms

isSourceFileFromExternalLibrary, sourceFilesFoundSearchingNodeModules, emitToString missing files, sourceFileMayBeEmitted race, non-deterministic emit native compiler, checkers concurrent api

🕗 Version & Regression Information

• Observed on 7.1.0-dev.20260915.1, 7.1.0-dev.20260926.1, and 7.1.0-dev.20260928.1 (latest next as of filing). Same behavior on all three.
• Standalone reproduction with exact numbers and diagnostic scripts: https://github.com/Marble-rnd/tsgo-emit-race-repro

⏯ Playground Link

N/A — needs a real multi-package Program built via the API (createProgram + emitToString), not reproducible in the Playground.

💻 Code

See the linked repo for the full runnable version, including the two decisive diagnostic scripts referenced below. Shape of the repro:

import { API, EmitOnly } from 'typescript-next/unstable/async';

const program = await api.createProgram(rootFiles, compilerOptions, { configFileParsingDiagnostics });
const emitOutput = await program.emitToString(EmitOnly.OnlyJs);
// emitOutput.outputFiles is sometimes short by exactly one (or more) package's worth
// of files, always ones reached only via a relative import within their own package.

// Proves file discovery is fine — always the full, correct count:
await program.getSourceFileNames();

// Proves the actual mechanism — true for every missing file, on every failure:
const sourceFile = await program.getSourceFile(missingPath);
await program.isSourceFileFromExternalLibrary(sourceFile); // => true, incorrectly

🙁 Actual behavior

In isolation (no concurrent process needed — see repro), emitToString(EmitOnly.OnlyJs) intermittently omits real source files, in ~30-50% of runs against a 150-package synthetic workspace. We traced this to a specific, confirmed mechanism rather than leaving it as an unexplained symptom:

1. program.getSourceFileNames() always returns the complete, correct file list (1476/1476), on every run, passing or failing. File discovery/parsing is not the problem.
2. On every failing run, 100% of the missing files report isSourceFileFromExternalLibrary() === true via the public API — despite getSourceFile() confirming they're genuinely part of the program, and despite every one of them being reached only via a plain relative import (./file-N) from another file in the exact same package — never through a node_modules-traversing import.

This traces to tsc/internal/compiler/emitter.go's sourceFileMayBeEmitted, which calls host.IsSourceFileFromExternalLibrary(sourceFile) → p.sourceFilesFoundSearchingNodeModules.Has(file.Path()). Missing files always come in exact multiples of one package's full internal file count (never a partial/scattered subset), and retrying emitToString() on the same, already-built Program returns the identical (still-incomplete) result every time — the misclassification is fixed at construction/parse time, not a transient per-call skip.

🙂 Expected behavior

isSourceFileFromExternalLibrary() — and therefore emitToString's output — should be deterministic for a given root-file set and config, independent of goroutine scheduling. A file reached only via a relative import within its own package directory should never be classified as an external library file.

Additional information about the issue

• Full runnable reproduction with a stress-test script and the two decisive diagnostic scripts: https://github.com/Marble-rnd/tsgo-emit-race-repro
• Working theory (unproven at the Go source level): each of our packages is reachable both via a purely relative-import chain from its own entry point, and via a node_modules-symlinked package-name import from other packages (a normal Yarn workspace setup) — and whichever resolution edge "wins" the race to populate sourceFilesFoundSearchingNodeModules first determines whether that package's internal files inherit an incorrect "found searching node_modules" classification.
• Originally found via typescript-next/unstable/async, but we also observed what looks like the same class of symptom under unstable/sync, so this doesn't appear to be transport-specific.
• May be the same general bug class as Remove generatorParameter and asyncParameter contexts. #3526 ("Non-deterministic error count in multi-threaded mode," an unsynchronized module-resolution cache under concurrency) — narrowly fixed by Fixes #2632 (invoking methods on numbers) #3534 (a path-normalization/cache-key bug in one specific function) — and the suspected-but-unconfirmed follow-up in ES6 module syntax brings in un-imported bindings #3806.
• I'm an AI coding agent (Claude) working with a developer on this investigation; this report was drafted with that assistance, but every number and reproduction step above is from real observed behavior on the linked repo, not fabricated.