@vzn/vx 0.0.16 → 0.0.21

package/src/cache/inputs.ts

@@ -6,22 +6,37 @@
 // `cache.inputs.env` is the cache-tracking axis for env vars; it's
 // independent of `exec.env`, which controls what reaches the child.
 //
-// File enumeration defers to git — same as Turbo and Nx. We ask git for
-// the file set via `git ls-files --cached --others --exclude-standard`,
-// which gives us:
-//   - all tracked files,
-//   - plus untracked-but-not-ignored files,
-//   - with nested .gitignore + .git/info/exclude + global excludes
-//     correctly applied (because git already does the cascade).
-// The user's `inputs.files` globs are then matched as a *filter* on
-// top of that file set. vx requires git to be installed and the
-// workspace to be a git work tree; non-git environments are not
-// supported.
+// File enumeration defers to git — same as Turbo and Nx. git-inputs.ts
+// asks for the tracked set with `git ls-files -s -v` (index OIDs and the
+// skip-worktree flag in the same spawn) and for the dirty and untracked
+// paths with `git status -uall`, with nested .gitignore, .git/info/exclude
+// and the global excludes applied because git already does the cascade.
+// The user's `inputs.files` globs are then matched as a *filter* on top
+// of that file set. vx requires git to be installed and the workspace to
+// be a git work tree; non-git environments are not supported.
 
 import path from 'node:path'
-import { rm } from 'node:fs/promises'
+import { lstatSync } from 'node:fs'
+import { realpath, rm, rmdir } from 'node:fs/promises'
 import type { CacheInputs } from '../config.js'
 import { UserError } from '../util/index.js'
+import { GitFilesCache, runGitLsFiles } from './git-inputs.js'
+import { normalizeGlob } from '../util/index.js'
+
+// The git side lives in git-inputs.ts; its whole public surface is
+// re-exported here so a reader that reaches the resolver for it (the
+// tests do, by deep import) keeps working.
+export {
+  GitFilesCache,
+  populateGitFilesCache,
+  runGitLsFiles,
+  startGitEnumeration,
+  applyGitEnumeration,
+  gitPathspecs,
+  parseCheckAttrOutput,
+  autocrlfConverts,
+  type GitEnumeration,
+} from './git-inputs.js'
 
 // vx-lock.json is committed (so git enumerates it) but it's vx's own
 // frozen-config metadata — never a task input. Excluded globally so a
@@ -74,7 +89,11 @@ export interface ResolveInputsArgs {
    * Run-scoped memo for `cache.inputs.runtime` command execution, keyed
    * by `projectDir + '\0' + command`. Shared across a run's tasks so a
    * project's command runs once even across build/test/lint and across
-   * the hash + sandbox-baseline resolveInputs calls.
+   * the hash + sandbox-baseline resolveInputs calls. The key is sound
+   * ONLY because the probe runs in vx's own ambient environment, never a
+   * task's `exec.env` (`runRuntimeCommand`): two tasks with different
+   * `define`s share one value. Threading the task env in would have to
+   * widen this key too, or task A's probe becomes task B's key component.
    */
   runtimeCache?: Map<string, Promise<string>>
   /**
@@ -82,8 +101,21 @@ export interface ResolveInputsArgs {
    * only — global dedup so a root-level probe spawns once per run.
    */
   workspaceRuntimeCache?: Map<string, Promise<string>>
+  /**
+   * Run-scoped memo for `cache.inputs.workspaceFiles`, keyed by the
+   * declaration and valid for one enumeration snapshot. A Turbo-mapped
+   * workspace gives every task the same `globalDependencies`, and
+   * resolving one literal against medusa's 24k-file enumeration per task
+   * was 930 ms of a 3.0 s warm no-op (159 scans, 2026-09-11).
+   */
+  workspaceFilesCache?: WorkspaceFilesCache
 }
 
+export type WorkspaceFilesCache = Map<
+  string,
+  { snapshot: readonly string[]; result: Promise<string[]> }
+>
+
 export async function resolveInputs(args: ResolveInputsArgs): Promise<ResolvedInputs> {
   const projectFiles = await resolveFiles({
     projectDir: args.projectDir,
@@ -101,26 +133,29 @@ export async function resolveInputs(args: ResolveInputsArgs): Promise<ResolvedIn
       workspaceFiles: wsDecl,
       ownWorkspaceOutputs: args.ownWorkspaceOutputs ?? [],
       ...(args.gitFilesCache !== undefined ? { gitFilesCache: args.gitFilesCache } : {}),
+      ...(args.workspaceFilesCache !== undefined ? { memo: args.workspaceFilesCache } : {}),
     })
     // Dedupe: when the project dir IS the workspace root (or a glob
     // overlaps), the same absolute path can arrive via both lists —
     // it must contribute to the key exactly once.
     if (wsFiles.length > 0) files = [...new Set([...projectFiles, ...wsFiles])].sort()
   }
-  const [runtimeValues, workspaceRuntimeValues] = await Promise.all([
-    resolveRuntimeValues(
-      args.inputs?.runtime ?? [],
-      args.projectDir,
-      args.runtimeCache,
-      `${args.projectDir}\0`,
-    ),
-    resolveRuntimeValues(
-      args.inputs?.workspaceRuntime ?? [],
-      args.workspaceRoot,
-      args.workspaceRuntimeCache,
-      '',
-    ),
-  ])
+  // Two awaited empty resolutions per task were ~3 ms of a 1000-task warm
+  // run (profiled 2026-09-09); a task declaring neither skips the fan-out.
+  const runtimeDecl = args.inputs?.runtime ?? []
+  const wsRuntimeDecl = args.inputs?.workspaceRuntime ?? []
+  const [runtimeValues, workspaceRuntimeValues] =
+    runtimeDecl.length === 0 && wsRuntimeDecl.length === 0
+      ? [[], []]
+      : await Promise.all([
+          resolveRuntimeValues(
+            runtimeDecl,
+            args.projectDir,
+            args.runtimeCache,
+            `${args.projectDir}\0`,
+          ),
+          resolveRuntimeValues(wsRuntimeDecl, args.workspaceRoot, args.workspaceRuntimeCache, ''),
+        ])
   return {
     files,
     envValues: resolveEnvValues(args.inputs?.env ?? [], args.envSource),
@@ -145,6 +180,7 @@ async function resolveWorkspaceFiles(args: {
   workspaceFiles: readonly string[]
   ownWorkspaceOutputs: readonly string[]
   gitFilesCache?: GitFilesCache
+  memo?: WorkspaceFilesCache
 }): Promise<string[]> {
   const positive: string[] = []
   const negative: string[] = []
@@ -154,10 +190,12 @@ async function resolveWorkspaceFiles(args: {
   }
   if (positive.length === 0) return []
 
-  const excludeGlobs = [...ALWAYS_IGNORE, ...args.ownWorkspaceOutputs, ...negative].map(
-    (p) => new Bun.Glob(p),
-  )
-  const positiveGlobs = positive.map((p) => new Bun.Glob(p))
+  const excludeGlobs = [
+    ...ALWAYS_IGNORE,
+    ...asTrees(args.ownWorkspaceOutputs),
+    ...asTrees(negative),
+  ].map(globFor)
+  const positiveGlobs = asTrees(positive).map(globFor)
   // Workspace-wide partition, keyed by the workspace root. Populated
   // up-front by `populateGitFilesCache(..., workspaceWide: true)` when
   // any loaded task declares workspaceFiles; a missing/invalidated
@@ -167,23 +205,70 @@ async function resolveWorkspaceFiles(args: {
     gitFiles = runGitLsFiles(args.workspaceRoot).files
     args.gitFilesCache?.set(args.workspaceRoot, gitFiles)
   }
+  // The memo is valid for the snapshot it was computed over: a task that
+  // wrote workspace outputs mid-run replaces the partition, and the next
+  // caller sees a different array and scans again.
+  const memoKey =
+    args.memo === undefined
+      ? undefined
+      : JSON.stringify([positive, negative, args.ownWorkspaceOutputs])
+  if (memoKey !== undefined) {
+    const hit = args.memo!.get(memoKey)
+    if (hit !== undefined && hit.snapshot === gitFiles) return hit.result
+  }
+  const result = resolveWorkspaceFilesOver(args, gitFiles, positive, positiveGlobs, excludeGlobs)
+  if (memoKey !== undefined) args.memo!.set(memoKey, { snapshot: gitFiles, result })
+  return result
+}
+
+async function resolveWorkspaceFilesOver(
+  args: { workspaceRoot: string; gitFilesCache?: GitFilesCache },
+  gitFiles: readonly string[],
+  positive: readonly string[],
+  positiveGlobs: readonly Bun.Glob[],
+  excludeGlobs: readonly Bun.Glob[],
+): Promise<string[]> {
+  // Second call site of the literal-input guard. `resolveWorkspaceFiles`
+  // carries its own copy of the filter-over-git-set design, so the same
+  // silently-folds-nothing hazard exists here — and a fix applied only to the
+  // project half would pass that half's tests while leaving this one live.
+  const unmatchedLiterals = new Set(
+    positive.map(normalizeGlob).filter(isLiteralPath).map(stripTrailingSlash),
+  )
   const candidates: string[] = []
   for (const rel of gitFiles) {
+    if (unmatchedLiterals.size > 0) settleLiterals(unmatchedLiterals, rel)
     if (!positiveGlobs.some((g) => g.match(rel))) continue
     if (excludeGlobs.some((g) => g.match(rel))) continue
     candidates.push(path.resolve(args.workspaceRoot, rel))
   }
+  if (unmatchedLiterals.size > 0) {
+    await assertNoInvisibleLiteralInputs(unmatchedLiterals, args.workspaceRoot, 'workspaceFiles')
+  }
   // Same OID-trust shortcut as project files: a clean-per-status
   // tracked file necessarily exists on disk.
   const oids = args.gitFilesCache?.oidsFor(args.workspaceRoot)
-  const exists = await Promise.all(
-    candidates.map((abs) => oids?.has(abs) === true || Bun.file(abs).exists()),
-  )
-  const matches: string[] = []
-  for (let i = 0; i < candidates.length; i++) {
-    if (exists[i]) matches.push(candidates[i]!)
+  return candidates.filter((abs) => oids?.has(abs) === true || isInputOnDisk(abs)).sort()
+}
+
+/**
+ * What an enumerated path must be to count as an input: a regular file, or
+ * a symlink (to anything, or to nothing — its target STRING is what folds,
+ * as in git). A directory is not one: git lists a gitlink (a submodule) at
+ * its path and the glob `**\/*` matches it, and the hasher has nothing to
+ * read there. `Bun.file(p).exists()` answered false for every symlink to a
+ * directory and every dangling link too, which silently dropped a tracked
+ * link from the key — retargeting it was a stale hit. lstat, synchronously:
+ * the paths that reach this probe are the ones without a trusted index OID
+ * (untracked and dirty files), a handful on a warm run.
+ */
+function isInputOnDisk(abs: string): boolean {
+  try {
+    const st = lstatSync(abs)
+    return st.isFile() || st.isSymbolicLink()
+  } catch {
+    return false
   }
-  return matches.sort()
 }
 
 function resolveEnvValues(
@@ -197,6 +282,11 @@ function resolveEnvValues(
  * Run one runtime-input command via `sh -c` (so pipelines / redirects
  * work — "shell is the API"). Returns trimmed stdout+stderr. A non-zero
  * exit is a hard UserError naming the command (fail-loud, like git).
+ *
+ * The probe inherits vx's AMBIENT environment on purpose — not the task's
+ * `exec.env.define` / `passThrough`, which describe the command's
+ * environment, not the machine's. That is what lets the run-scoped memo
+ * key on (projectDir, command) alone; see `runtimeCache`.
  */
 async function runRuntimeCommand(command: string, cwd: string): Promise<string> {
   let proc
@@ -262,10 +352,70 @@ export async function resolveOutputs(args: {
   nestedProjectDirs: string[]
 }): Promise<string[]> {
   if (args.outputs.length === 0) return []
-  const excludeGlobs = boundaryIgnorePatterns(args.projectDir, args.nestedProjectDirs).map(
-    (p) => new Bun.Glob(p),
-  )
-  return [...(await scanUnion(args.outputs, excludeGlobs, args.projectDir))].sort()
+  const excludeGlobs = boundaryIgnorePatterns(args.projectDir, args.nestedProjectDirs).map(globFor)
+  const scanned = [
+    ...(await scanUnion(asTrees(args.outputs), excludeGlobs, args.projectDir, 'outputs')),
+  ]
+  // Containment, enforced HERE and not only at the loader. `cleanOutputs`
+  // DELETES whatever this returns, and `Bun.Glob.scan` happily walks `..` out
+  // of its cwd — so the loader's `..`/absolute rejection alone was a single
+  // point of failure. The resolver that feeds the delete refuses to name a path
+  // outside the project, so any future caller reaching it by another route (a
+  // programmatic embedder, a config source that skips the loader) is contained
+  // by construction.
+  return (await containedIn(args.projectDir, scanned)).sort()
+}
+
+/**
+ * Keep only the paths that are REALLY inside `root` — lexically, and after
+ * resolving symlinks.
+ *
+ * Lexical alone was sufficient only while `Bun.Glob.scan` refused to descend
+ * into symlinked directories, which this repo pinned as a deliberate tripwire
+ * on a DEPENDENCY's behaviour. **Bun 1.4.0 tripped it**: with `dist ->
+ * ../victim` the scan now yields `dist/precious.txt`, a path that is lexically
+ * inside the project while the file it names is not — and the caller rm()s
+ * whatever this returns. Measured on 1.4.0 before this guard: a plain
+ * `outputs.files: ['dist/**']` deleted a file outside the project.
+ *
+ * DIRECTORIES are what matter, not files: `rm` on a symlinked FILE unlinks the
+ * link and never its target, so a link sitting inside a real output directory
+ * is harmless. Resolving per directory also keeps this cheap — one syscall per
+ * distinct output directory rather than per output file, concurrently, and a
+ * `dist/**` of ten thousand files in one directory costs exactly one.
+ *
+ * A path whose directory will not resolve (a broken link, or a race with the
+ * task that produced it) is REFUSED. When the caller deletes, unresolvable
+ * means leave it alone.
+ */
+async function containedIn(root: string, paths: readonly string[]): Promise<string[]> {
+  // One `path.dirname` per path, kept alongside — computing it again in the
+  // final filter measured as the DOMINANT added cost on a wide output tree
+  // (string work, not syscalls: 10k files in 20 dirs cost more than 5k files
+  // in 200, which is the wrong shape for a per-directory probe).
+  const lexical: string[] = []
+  const lexDirs: string[] = []
+  for (const p of paths) {
+    if (!isInside(root, p)) continue
+    lexical.push(p)
+    lexDirs.push(path.dirname(p))
+  }
+  if (lexical.length === 0) return []
+  const realRoot = await realpath(root).catch(() => root)
+  const uniqueDirs = [...new Set(lexDirs)]
+  const resolved = await Promise.all(uniqueDirs.map((d) => realpath(d).catch(() => null)))
+  const contained = new Set<string>()
+  for (const [i, dir] of uniqueDirs.entries()) {
+    const real = resolved[i]
+    if (real !== null && real !== undefined && isInside(realRoot, real)) contained.add(dir)
+  }
+  return lexical.filter((_p, i) => contained.has(lexDirs[i]!))
+}
+
+/** Is `abs` the directory `dir` itself or something beneath it? */
+function isInside(dir: string, abs: string): boolean {
+  if (abs === dir) return true
+  return abs.startsWith(dir.endsWith(path.sep) ? dir : dir + path.sep)
 }
 
 /**
@@ -289,12 +439,43 @@ export async function cleanOutputs(args: {
   const files = await resolveOutputs(args)
   // `force: true` makes rm tolerate ENOENT (e.g. when two output
   // globs overlap and a sibling already deleted a path mid-iteration).
+  // A symlink is unlinked, never followed.
   await Promise.all(files.map((f) => rm(f, { force: true })))
+  await pruneEmptiedDirs(args.projectDir, files)
   // Project-relative posix paths of what was removed — the caller
   // feeds these to GitFilesCache.markOutputsChanged after a restore.
   return files.map((f) => path.relative(args.projectDir, f).split(path.sep).join('/'))
 }
 
+/**
+ * Remove the directories a clean emptied, bottom-up, never `root` itself.
+ * A restore writes files where the cached entry has them; a directory left
+ * standing where the entry holds a FILE of the same name (a task that once
+ * wrote `dist/out/…` and now writes `dist/out`) blocks the rename, and an
+ * empty directory is not an output anyone declared. A directory that still
+ * holds something — a stray the globs do not cover — stays, and the restore
+ * says so if it is in the way.
+ */
+async function pruneEmptiedDirs(root: string, removed: readonly string[]): Promise<void> {
+  const rootResolved = path.resolve(root)
+  const tried = new Set<string>()
+  // Deepest first, so a parent is attempted after every child had its turn.
+  const dirs = [...new Set(removed.map((f) => path.dirname(f)))].sort(
+    (a, b) => b.split(path.sep).length - a.split(path.sep).length,
+  )
+  for (let dir of dirs) {
+    while (dir !== rootResolved && dir.startsWith(rootResolved + path.sep) && !tried.has(dir)) {
+      tried.add(dir)
+      const gone = await rmdir(dir).then(
+        () => true,
+        (err: NodeJS.ErrnoException) => err.code === 'ENOENT',
+      )
+      if (!gone) break
+      dir = path.dirname(dir)
+    }
+  }
+}
+
 /**
  * Resolve declared `outputs.workspaceFiles` globs (workspace-root-
  * relative) to actual produced files. Live-FS glob like
@@ -307,7 +488,12 @@ export async function resolveWorkspaceOutputs(args: {
   outputs: string[]
 }): Promise<string[]> {
   if (args.outputs.length === 0) return []
-  return [...(await scanUnion(args.outputs, [], args.workspaceRoot))].sort()
+  const scanned = [...(await scanUnion(asTrees(args.outputs), [], args.workspaceRoot, 'outputs'))]
+  // Same containment as the project twin, anchored one level out. These globs
+  // deliberately ignore PROJECT boundaries — that is the escape hatch — but
+  // escaping the WORKSPACE was never part of it, and `cleanWorkspaceOutputs`
+  // deletes what this returns.
+  return (await containedIn(args.workspaceRoot, scanned)).sort()
 }
 
 /**
@@ -324,166 +510,87 @@ export async function cleanWorkspaceOutputs(args: {
 }): Promise<string[]> {
   const files = await resolveWorkspaceOutputs(args)
   await Promise.all(files.map((f) => rm(f, { force: true })))
+  await pruneEmptiedDirs(args.workspaceRoot, files)
   return files.map((f) => path.relative(args.workspaceRoot, f).split(path.sep).join('/'))
 }
 
-/**
- * Per-run memo of each project's `git ls-files` output, plus the
- * staleness bookkeeping that lets the warm path avoid re-spawning git.
- *
- * After a cache-hit restore we know EXACTLY which paths changed on
- * disk: the declared outputs `cleanOutputs` wiped plus the artifact's
- * output files. `markOutputsChanged` records them; `snapshotFor`
- * hands back the existing snapshot when a resolving task's input
- * globs can't match any changed path — provably identical to what a
- * re-spawn would return, since glob matching ignores gitignore status
- * entirely when the path doesn't match. When globs DO overlap,
- * returning undefined forces the caller down the re-spawn path so
- * gitignore semantics stay byte-identical.
- *
- * The cache-miss save path still uses plain `delete` — an executed
- * task may write files outside its declared outputs, and only git can
- * see those.
- */
-export class GitFilesCache extends Map<string, readonly string[]> {
-  private changed = new Map<string, string[]>()
-  /**
-   * Per-project trusted index OIDs: absolute path → git blob OID, for
-   * tracked regular files whose working-tree state matched the index
-   * at populate time (per one `git status --porcelain` snapshot).
-   * These feed `Cache.key` via `CacheKeyInput.fileHashes` so a clean
-   * tree derives input hashes with zero reads / stats / SQLite.
-   *
-   * Dropped wholesale on `set` / `delete`: a mid-run re-enumeration
-   * proves the project's tree changed, and index OIDs can't be
-   * re-trusted without a fresh status — the per-file fallback
-   * (`Cache.hashFile`) computes the identical blob OID from disk, so
-   * dropping is a pure perf concession, never a correctness one.
-   */
-  private oids = new Map<string, Map<string, string>>()
-  /**
-   * Set when `populateGitFilesCache` stored a WORKSPACE-WIDE partition
-   * (any loaded task declares `inputs.workspaceFiles`). Null when the
-   * feature is unused — every workspace-partition hook below is then a
-   * no-op, so unused-feature behavior stays byte-identical.
-   */
-  private wsRoot: string | null = null
-  /**
-   * Whether the worktree had any uncommitted changes at populate time,
-   * derived from the SAME `git status --porcelain` spawn that prunes
-   * dirty paths from the trusted-OID set. Lets the Tier-3 invocation
-   * record report `dirty` without a SECOND status spawn. `null` until
-   * populate runs, or when the status spawn failed (non-repo).
-   */
-  private dirty: boolean | null = null
-
-  setWorkspaceRoot(root: string): void {
-    this.wsRoot = root
-  }
-
-  /** Aggregate worktree dirtiness from the populate-time status spawn. */
-  get worktreeDirty(): boolean | null {
-    return this.dirty
-  }
-
-  setWorktreeDirty(dirty: boolean | null): void {
-    this.dirty = dirty
-  }
-
-  markOutputsChanged(projectDir: string, relPaths: readonly string[]): void {
-    this.recordChanged(projectDir, relPaths)
-    // The workspace-wide partition sees the same files under
-    // root-relative names; forward so a downstream workspaceFiles
-    // task can't reuse a snapshot its globs could now contradict.
-    if (this.wsRoot !== null && this.wsRoot !== projectDir) {
-      this.recordChanged(
-        this.wsRoot,
-        relPaths.map((rel) =>
-          path.relative(this.wsRoot!, path.resolve(projectDir, rel)).split(path.sep).join('/'),
-        ),
-      )
-    }
-  }
-
-  /**
-   * Record root-anchored changed paths (cleaned/restored workspace
-   * outputs) against EVERY partition that can see them: the workspace
-   * partition under their root-relative names, and any project
-   * partition whose dir contains them (workspace outputs may land
-   * inside other projects' dirs — the no-boundary escape hatch).
-   */
-  markWorkspaceOutputsChanged(workspaceRoot: string, relPaths: readonly string[]): void {
-    if (relPaths.length === 0) return
-    for (const key of this.keys()) {
-      if (key === workspaceRoot) {
-        this.recordChanged(key, relPaths)
-        continue
-      }
-      const under: string[] = []
-      for (const rel of relPaths) {
-        const abs = path.resolve(workspaceRoot, rel)
-        if (abs.startsWith(key + path.sep)) {
-          under.push(path.relative(key, abs).split(path.sep).join('/'))
-        }
-      }
-      if (under.length > 0) this.recordChanged(key, under)
-    }
-  }
-
-  /**
-   * Drop the workspace-wide partition (if one exists). Called after a
-   * cache-miss save — an executed task may have written undeclared
-   * files anywhere in its project dir, and the workspace partition
-   * spans that subtree; only a fresh enumeration can see them. No-op
-   * when the feature is unused.
-   */
-  invalidateWorkspacePartition(): void {
-    if (this.wsRoot !== null) this.delete(this.wsRoot)
-  }
-
-  private recordChanged(partitionDir: string, relPaths: readonly string[]): void {
-    if (!this.has(partitionDir)) return
-    const cur = this.changed.get(partitionDir)
-    if (cur) cur.push(...relPaths)
-    else this.changed.set(partitionDir, [...relPaths])
-    // These paths just changed on disk; their index OIDs (if any) no
-    // longer describe the working-tree content.
-    const partitionOids = this.oids.get(partitionDir)
-    if (partitionOids) {
-      for (const rel of relPaths) partitionOids.delete(path.resolve(partitionDir, rel))
-    }
-  }
-
-  /** Trusted index OIDs for a project (abs path → oid), if any survive. */
-  oidsFor(projectDir: string): ReadonlyMap<string, string> | undefined {
-    return this.oids.get(projectDir)
-  }
+function isLiteralPath(glob: string): boolean {
+  return !/[*?[\]{}]/.test(glob)
+}
 
-  setOids(projectDir: string, oids: Map<string, string>): void {
-    this.oids.set(projectDir, oids)
-  }
+function stripTrailingSlash(p: string): string {
+  return p.replace(/\/+$/, '')
+}
 
-  /** Snapshot if still valid for these input globs; undefined → re-spawn. */
-  snapshotFor(projectDir: string, inputGlobs: readonly Bun.Glob[]): readonly string[] | undefined {
-    const snap = this.get(projectDir)
-    if (snap === undefined) return undefined
-    const pending = this.changed.get(projectDir)
-    if (pending !== undefined && pending.some((p) => inputGlobs.some((g) => g.match(p)))) {
-      return undefined
+/**
+ * A literal entry names a file OR a directory tree: `src/` and `src` both
+ * mean everything under `src`, as they do in Turbo and every `.gitignore`.
+ * A glob matcher sees only the literal path, so `['src/']` folded ZERO
+ * files — a key that never moves with its source, and the most common
+ * turbo.json shape (`"outputs": ["dist"]`) captured nothing. Every
+ * literal therefore compiles to itself plus its subtree; a literal that
+ * names a file still matches exactly that file, since `x/**` matches
+ * nothing under a file.
+ */
+export function asTrees(patterns: readonly string[]): string[] {
+  const out: string[] = []
+  for (const raw of patterns) {
+    const p = normalizeGlob(raw)
+    if (!isLiteralPath(p)) {
+      out.push(p)
+      continue
     }
-    return snap
+    const lit = stripTrailingSlash(p)
+    if (lit.length === 0) continue
+    out.push(lit, `${lit}/**`)
   }
+  return out
+}
 
-  override set(key: string, value: readonly string[]): this {
-    this.changed.delete(key)
-    this.oids.delete(key)
-    return super.set(key, value)
+/** A literal is answered by the path itself or by anything under it. */
+function settleLiterals(unmatched: Set<string>, rel: string): void {
+  for (const lit of unmatched) {
+    if (rel === lit || rel.startsWith(`${lit}/`)) unmatched.delete(lit)
   }
+}
 
-  override delete(key: string): boolean {
-    this.changed.delete(key)
-    this.oids.delete(key)
-    return super.delete(key)
+/**
+ * Refuse a literal `cache.inputs.files` entry that EXISTS ON DISK but is
+ * invisible to git — gitignored, or otherwise absent from
+ * `git ls-files --cached --others --exclude-standard`.
+ *
+ * Such an entry contributes nothing to the cache key, so the task stops
+ * tracking a file its own config names as an input. That is a stale hit, and a
+ * quiet one: the run replays cached stdout, so it even LOOKS like it executed.
+ *
+ * Turbo honours the declaration instead (an explicit `inputs` entry overrides
+ * gitignore). vx cannot afford to, and the reason is not cost: honouring it
+ * would make the cache key depend on a change that `git diff` and
+ * `git ls-files --others` CANNOT SEE, so editing that input would re-key the
+ * task while `--affected` selected nothing. `docs/cli.md` states the invariant
+ * those two surfaces owe each other as a principle — "input hashing sees it, so
+ * `--affected` must too" — and it was closed for lockfiles only recently.
+ * Honouring a gitignored input would reopen it from the other side.
+ *
+ * An entry that does NOT exist on disk stays silent: that is an ordinary stale
+ * declaration, and refusing it would break every config that lists an
+ * optional file.
+ */
+async function assertNoInvisibleLiteralInputs(
+  literals: ReadonlySet<string>,
+  base: string,
+  field: 'files' | 'workspaceFiles' = 'files',
+): Promise<void> {
+  for (const rel of literals) {
+    if (!(await Bun.file(path.resolve(base, rel)).exists())) continue
+    throw new UserError(
+      `cache.inputs.${field}: "${rel}" exists in ${base} but git does not report it, ` +
+        `so it contributes NOTHING to the cache key — input globs filter the files git ` +
+        `lists, and a filter cannot add one back. The task would keep reporting ` +
+        `up-to-date after that file changed. If it is generated, depend on the task ` +
+        `that produces it via cache.inputs.tasks; if it should be tracked, remove it ` +
+        `from .gitignore.`,
+    )
   }
 }
 
@@ -512,9 +619,12 @@ async function resolveFiles(args: ResolveFilesArgs): Promise<string[]> {
   if (positive.length === 0) return []
 
   const boundaryIgnores = boundaryIgnorePatterns(args.projectDir, args.nestedProjectDirs)
-  const excludeGlobs = [...ALWAYS_IGNORE, ...boundaryIgnores, ...args.ownOutputs, ...negative].map(
-    (p) => new Bun.Glob(p),
-  )
+  const excludeGlobs = [
+    ...ALWAYS_IGNORE,
+    ...boundaryIgnores,
+    ...asTrees(args.ownOutputs),
+    ...asTrees(negative),
+  ].map(globFor)
 
   // Defer to git for the file set (Turbo / Nx parity). Nested .gitignore
   // files, .git/info/exclude, and global excludes all participate
@@ -524,7 +634,7 @@ async function resolveFiles(args: ResolveFilesArgs): Promise<string[]> {
   // per task (build + test + lint + …). Spawning git N times for the
   // same project per run is wasteful; we cache the result for the
   // duration of one orchestrator run.
-  const positiveGlobs = positive.map((p) => new Bun.Glob(p))
+  const positiveGlobs = asTrees(positive).map(globFor)
   let gitFiles = args.gitFilesCache?.snapshotFor(args.projectDir, positiveGlobs)
   if (gitFiles === undefined) {
     // Mid-run re-enumeration. The OIDs this spawn could yield are NOT
@@ -536,9 +646,24 @@ async function resolveFiles(args: ResolveFilesArgs): Promise<string[]> {
     // set() also clears the project's pending-changed bookkeeping.
     args.gitFilesCache?.set(args.projectDir, gitFiles)
   }
+  // A LITERAL entry — one with no glob metacharacter — names exactly one file,
+  // so "this matched nothing" is unambiguous. For a glob it is not: matching
+  // nothing is perfectly legitimate (`src/**/*.gen.ts` in a package with no
+  // generated code), which is why only literals are tracked here.
+  //
+  // The reason it has to be tracked at all: the user's globs FILTER the set git
+  // reports, and a filter can only remove. So naming a gitignored file by hand
+  // folds NOTHING — silently. The task then reports `up-to-date` while replaying
+  // an artifact built from an older version of a file the config explicitly
+  // claims as an input. See the refusal below for why this is not simply
+  // honoured instead.
+  const unmatchedLiterals = new Set(
+    positive.map(normalizeGlob).filter(isLiteralPath).map(stripTrailingSlash),
+  )
   // First pass: glob-filter to candidate absolute paths (no I/O).
   const candidates: string[] = []
   for (const rel of gitFiles) {
+    if (unmatchedLiterals.size > 0) settleLiterals(unmatchedLiterals, rel)
     let matched = false
     for (const g of positiveGlobs) {
       if (g.match(rel)) {
@@ -550,280 +675,62 @@ async function resolveFiles(args: ResolveFilesArgs): Promise<string[]> {
     if (excludeGlobs.some((g) => g.match(rel))) continue
     candidates.push(path.resolve(args.projectDir, rel))
   }
-  // Second pass: parallel existence check — but ONLY for paths
-  // without a trusted index OID. A clean-per-status tracked file
-  // necessarily exists on disk, so skipping its probe keeps the warm
-  // path free of per-file syscalls. Paths without an OID keep the
-  // probe: `git ls-files -s` can surface staged entries whose
-  // working-tree file is gone; the hasher would otherwise throw
-  // ENOENT.
-  const oids = args.gitFilesCache?.oidsFor(args.projectDir)
-  const exists = await Promise.all(
-    candidates.map((abs) => oids?.has(abs) === true || Bun.file(abs).exists()),
-  )
-  const matches: string[] = []
-  for (let i = 0; i < candidates.length; i++) {
-    if (exists[i]) matches.push(candidates[i]!)
-  }
-  return matches.sort()
-}
-
-interface GitLsResult {
-  /** cwd-relative paths — same visibility set as `--cached --others`. */
-  files: string[]
-  /**
-   * cwd-relative path → index blob OID, for tracked REGULAR files
-   * (mode 100644 / 100755) at stage 0 only. Symlinks are excluded
-   * (their OID hashes the link-target string, not the dereferenced
-   * content our fallback hasher would read); merge-conflict stages
-   * and gitlinks are excluded too. NOT yet filtered by working-tree
-   * dirtiness — callers intersect with `git status` before trusting.
-   */
-  oids: Map<string, string>
-}
-
-// `<mode> <oid> <stage>\t<path>` — the staged-entry form of
-// `ls-files -s`. `--others` paths print bare; with `-z`,
-// core.quotePath quoting is off, so a bare path containing a literal
-// tab still can't match this fixed-form prefix.
-const LS_FILES_STAGE_RE = /^([0-7]{6}) ([0-9a-f]{40,64}) ([0-3])\t/
-
-/**
- * Run `git ls-files -s --others --exclude-standard -z .` in `cwd`.
- * One spawn yields BOTH the file list (identical visibility to the
- * pre-v20 `--cached --others` form, verified empirically: same set
- * including staged-but-deleted files and per-stage conflict
- * duplicates) AND each tracked file's index OID — the heart of the
- * Turbo-parity "hashes come from git's index" fast path. Throws a
- * `UserError` when git is unavailable or `cwd` isn't a git work
- * tree; vx requires git. `-z` survives filenames with newlines /
- * spaces.
- */
-function runGitLsFiles(cwd: string): GitLsResult {
-  let proc
-  try {
-    proc = Bun.spawnSync({
-      cmd: ['git', 'ls-files', '-s', '--others', '--exclude-standard', '-z', '.'],
-      cwd,
-      stdout: 'pipe',
-      stderr: 'pipe',
-    })
-  } catch {
-    throw new UserError(
-      `vx requires git: failed to spawn 'git' (working dir: ${cwd}). Install git and re-run.`,
-    )
-  }
-  if (proc.exitCode !== 0) {
-    // Exit 128 = not a git work tree; other non-zero = git failure.
-    // Either way we can't enumerate inputs reliably.
-    const stderr = new TextDecoder().decode(proc.stderr).trim()
-    throw new UserError(
-      `vx requires git: ${cwd} is not inside a git work tree. ` +
-        `Run 'git init' in your workspace root.${stderr ? ` (git: ${stderr})` : ''}`,
-    )
-  }
-  return parseLsFilesOutput(new TextDecoder().decode(proc.stdout))
-}
-
-function parseLsFilesOutput(out: string): GitLsResult {
-  const files: string[] = []
-  const oids = new Map<string, string>()
-  if (out.length === 0) return { files, oids }
-  // NUL-separated; trailing NUL produces an empty segment we skip.
-  for (const record of out.split('\0')) {
-    if (record.length === 0) continue
-    const m = LS_FILES_STAGE_RE.exec(record)
-    if (m === null) {
-      files.push(record) // --others entry: bare path
-      continue
-    }
-    const filePath = record.slice(m[0].length)
-    files.push(filePath)
-    const mode = m[1]!
-    const stage = m[3]!
-    if ((mode === '100644' || mode === '100755') && stage === '0') {
-      oids.set(filePath, m[2]!)
-    }
-  }
-  return { files, oids }
-}
-
-function parseStatusOutput(out: string): Set<string> {
-  const tokens = out.split('\0')
-  const dirty = new Set<string>()
-  for (let i = 0; i < tokens.length; i++) {
-    const token = tokens[i]!
-    if (token.length < 4) continue
-    dirty.add(token.slice(3))
-    // X = R or C → the next token is the rename/copy source path.
-    const x = token[0]
-    if ((x === 'R' || x === 'C') && i + 1 < tokens.length && tokens[i + 1]!.length > 0) {
-      i++
-      dirty.add(tokens[i]!)
-    }
-  }
-  return dirty
-}
-
-/**
- * Run `git ls-files` ONCE at the workspace root, then partition the
- * result by project. Populates `cache` for every project in
- * `projectDirs` with project-relative path lists matching what a
- * per-project spawn would have produced. Throws `UserError` if the
- * workspace isn't a git work tree (vx requires git).
- *
- * Why bulk: each spawn costs ~5-10ms (fork+exec). On a 200-project
- * workspace that's 1-2s of pure overhead reclaimed.
- *
- * Files in nested-project subtrees stay in their parent's list — the
- * boundary-ignore globs in `resolveFiles` filter them out the same way
- * they did before. Cheaper to filter once-per-task than to subtract
- * here.
- *
- * v20: the same `ls-files -s` spawn also yields each tracked file's
- * index OID. A second spawn (`git status --porcelain`) prunes paths
- * whose working tree diverges from the index; what survives is
- * stored per project via `cache.setOids` and feeds `Cache.key`
- * directly — clean-tree input hashing costs zero reads/stats/SQLite.
- *
- * `workspaceWide`: set when any loaded task declares
- * `cache.inputs.workspaceFiles`. Disables pathspec scoping (those
- * globs must see every file from the root) and additionally stores a
- * workspace-wide partition keyed by `workspaceRoot` (files + trusted
- * OIDs), which `resolveWorkspaceFiles` consumes. When false, the
- * enumeration behavior — pathspecs, spawn count, stored partitions —
- * is byte-identical to the pre-workspaceFiles code.
- */
-export async function populateGitFilesCache(
-  workspaceRoot: string,
-  projectDirs: readonly string[],
-  cache: GitFilesCache,
-  workspaceWide = false,
-): Promise<void> {
-  // The two spawns are independent — run them concurrently so the
-  // bulk-populate costs max(ls-files, status) wall time, not the sum
-  // (status alone is ~74 ms on a 1000-project tree; serial spawning
-  // was a measurable warm-path regression vs the pre-OID code).
-  const spawnGit = async (args: string[]) => {
-    try {
-      const proc = Bun.spawn({
-        cmd: ['git', ...args],
-        cwd: workspaceRoot,
-        stdin: 'ignore',
-        stdout: 'pipe',
-        stderr: 'pipe',
-      })
-      const [stdout, stderr, exitCode] = await Promise.all([
-        new Response(proc.stdout).text(),
-        new Response(proc.stderr).text(),
-        proc.exited,
-      ])
-      return { exitCode, stdout, stderr }
-    } catch {
-      return null
-    }
-  }
-  // Pathspec scoping: when the run only needs a handful of projects
-  // (scoped config loading), let git scan just those dirs — 75 ms →
-  // 11 ms on an 11k-file repo. Above 64 dirs (or when a project IS
-  // the root) the whole-tree scan wins on arg/exec overhead anyway.
-  const rels = projectDirs.map((d) => path.relative(workspaceRoot, d).split(path.sep).join('/'))
-  const scoped =
-    !workspaceWide &&
-    rels.length > 0 &&
-    rels.length <= 64 &&
-    rels.every((r) => r !== '' && r !== '.')
-  const pathspecs = scoped ? rels : ['.']
-  const [ls, status] = await Promise.all([
-    spawnGit(['ls-files', '-s', '--others', '--exclude-standard', '-z', '--', ...pathspecs]),
-    spawnGit(['status', '--porcelain', '-z', '--', ...pathspecs]),
-  ])
-  if (ls === null) {
-    throw new UserError(
-      `vx requires git: failed to spawn 'git' (working dir: ${workspaceRoot}). Install git and re-run.`,
-    )
-  }
-  if (ls.exitCode !== 0) {
-    const stderr = ls.stderr.trim()
-    throw new UserError(
-      `vx requires git: ${workspaceRoot} is not inside a git work tree. ` +
-        `Run 'git init' in your workspace root.${stderr ? ` (git: ${stderr})` : ''}`,
-    )
-  }
-  const { files: all, oids } = parseLsFilesOutput(ls.stdout)
-  const dirty = status !== null && status.exitCode === 0 ? parseStatusOutput(status.stdout) : null
-  // Aggregate dirtiness for the Tier-3 invocation record — derived from
-  // this same status spawn so `run()` needs no second `git status`.
-  // null when the status spawn failed (non-repo / git error).
-  cache.setWorktreeDirty(dirty === null ? null : dirty.size > 0)
-  const trusted = dirty === null ? new Map<string, string>() : oids
-  if (dirty !== null) {
-    for (const rel of dirty) trusted.delete(rel)
-  }
-  // Sort once, then each project's files are a contiguous range found
-  // by binary search on its `dir/` prefix — O((F+P) log F) instead of
-  // the O(P·F) per-project startsWith scan (54 ms at 1090 projects ×
-  // ~9k files; ~5 ms this way). '/' sorts below most filename chars,
-  // so the range [prefix, prefix+'\xff…') is contiguous in the sorted
-  // array; lowerBound on `prefix` and on `prefix + '￿'` bracket it.
-  const sorted = [...all].sort()
-  const lowerBound = (key: string): number => {
-    let lo = 0
-    let hi = sorted.length
-    while (lo < hi) {
-      const mid = (lo + hi) >>> 1
-      if (sorted[mid]! < key) lo = mid + 1
-      else hi = mid
-    }
-    return lo
-  }
-  for (const projectDir of projectDirs) {
-    const relPrefix = path.relative(workspaceRoot, projectDir).split(path.sep).join('/')
-    if (relPrefix === '' || relPrefix === '.') {
-      cache.set(projectDir, all)
-      const rootOids = new Map<string, string>()
-      for (const [rel, oid] of trusted) rootOids.set(path.join(workspaceRoot, rel), oid)
-      cache.setOids(projectDir, rootOids)
-      continue
-    }
-    const prefix = `${relPrefix}/`
-    const start = lowerBound(prefix)
-    const end = lowerBound(`${prefix}￿`)
-    const matches: string[] = []
-    const projOids = new Map<string, string>()
-    for (let i = start; i < end; i++) {
-      const rel = sorted[i]!
-      matches.push(rel.slice(prefix.length))
-      const oid = trusted.get(rel)
-      if (oid !== undefined) projOids.set(path.join(workspaceRoot, rel), oid)
-    }
-    // setOids AFTER set — set() drops the project's OID slot.
-    cache.set(projectDir, matches)
-    cache.setOids(projectDir, projOids)
-  }
-  if (workspaceWide) {
-    const rootOids = new Map<string, string>()
-    for (const [rel, oid] of trusted) rootOids.set(path.join(workspaceRoot, rel), oid)
-    cache.set(workspaceRoot, all)
-    cache.setOids(workspaceRoot, rootOids)
-    cache.setWorkspaceRoot(workspaceRoot)
+  if (unmatchedLiterals.size > 0) {
+    await assertNoInvisibleLiteralInputs(unmatchedLiterals, args.projectDir, 'files')
   }
+  // Second pass: existence check — but ONLY for paths without a
+  // trusted index OID. A clean-per-status tracked file necessarily
+  // exists on disk, so skipping its probe keeps the warm path free of
+  // per-file syscalls. Paths without an OID keep the probe:
+  // `git ls-files -s` can surface staged entries whose working-tree
+  // file is gone; the hasher would otherwise throw ENOENT.
+  const oids = args.gitFilesCache?.oidsFor(args.projectDir)
+  return candidates.filter((abs) => oids?.has(abs) === true || isInputOnDisk(abs)).sort()
 }
 
 /**
  * Union of files matching any positive pattern in `cwd`, minus files
  * matching any exclude glob (tested by Bun.Glob.match on the relative
  * path). Bun.Glob takes a single pattern per instance, so we iterate.
+ *
+ * `outputs` mode also yields SYMLINKS (never followed, never descended):
+ * a task that emits `dist/out -> ../src/x` has produced an output, so the
+ * save must capture it (as its target's bytes, `planArtifact`) and the
+ * clean must remove it — a link the clean leaves standing blocks the
+ * directory a later entry restores there. Bun's `onlyFiles` walk drops
+ * every symlink, so this mode lists everything and sorts by `lstat`.
  */
 async function scanUnion(
   positive: readonly string[],
   excludeGlobs: readonly Bun.Glob[],
   cwd: string,
+  mode: 'files' | 'outputs' = 'files',
 ): Promise<Set<string>> {
   const matches = new Set<string>()
   for (const pattern of positive) {
-    const glob = new Bun.Glob(pattern)
-    for await (const rel of glob.scan({ cwd, onlyFiles: true, dot: true })) {
+    const glob = globFor(pattern)
+    if (mode === 'outputs') {
+      for (const rel of glob.scanSync({
+        cwd,
+        onlyFiles: false,
+        followSymlinks: false,
+        dot: true,
+      })) {
+        if (excludeGlobs.some((g) => g.match(rel))) continue
+        const abs = path.resolve(cwd, rel)
+        const st = lstatSync(abs, { throwIfNoEntry: false })
+        if (st !== undefined && (st.isFile() || st.isSymbolicLink())) matches.add(abs)
+      }
+      continue
+    }
+    // Sync scan. The async walk was chosen on 2026-09-02, when a warm HIT
+    // globbed its outputs and the thread pool overlapped a thousand of
+    // them; the directory short-circuit took the glob off the hit path, and
+    // what is left runs on the MISS path twice per task (clean, then
+    // resolve), where the walk is CPU-bound and each async chunk costs a
+    // main-thread round trip. Measured 2026-09-09 on a 1,000-task cold run
+    // (see STATUS).
+    for (const rel of glob.scanSync({ cwd, onlyFiles: true, dot: true })) {
       if (excludeGlobs.some((g) => g.match(rel))) continue
       matches.add(path.resolve(cwd, rel))
     }
@@ -831,6 +738,16 @@ async function scanUnion(
   return matches
 }
 
+// Compiled once per pattern string for the life of the process: a `Bun.Glob`
+// is immutable, and the same handful of declared globs are compiled again
+// for every task otherwise (three sites, up to thousands of times per run).
+const globMemo = new Map<string, Bun.Glob>()
+function globFor(pattern: string): Bun.Glob {
+  let g = globMemo.get(pattern)
+  if (g === undefined) globMemo.set(pattern, (g = new Bun.Glob(pattern)))
+  return g
+}
+
 function boundaryIgnorePatterns(projectDir: string, nestedDirs: string[]): string[] {
   return nestedDirs.map((d) => {
     const rel = path.relative(projectDir, d).split(path.sep).join('/')