@pyreon/compiler 0.13.1 → 0.15.0

package/src/test-audit.ts

@@ -0,0 +1,435 @@
+/**
+ * Test-environment audit for the `audit_test_environment` MCP tool (T2.5.7).
+ *
+ * Scans `*.test.ts` / `*.test.tsx` files under the `packages` tree
+ * for **mock-vnode patterns** — tests that construct `{ type, props,
+ * children }` object literals (or a custom `vnode(...)` helper) in
+ * place of going through the real `h()` from `@pyreon/core`. This
+ * class of pattern silently drops rocketstyle / compiler / attrs
+ * work from the pipeline, letting bugs through that production
+ * would hit immediately (see PR #197 silent metadata drop).
+ *
+ * The scanner does NOT run the tests or parse TypeScript — a fast
+ * regex pass is intentional. Accuracy trades for speed: the false-
+ * positive rate is low because the `{ type: ..., props: ...,
+ * children: ... }` shape is unusual outside of vnode construction.
+ *
+ * Output classification:
+ *   HIGH   — mock patterns present, no real `h()` calls and no `h`
+ *            import from `@pyreon/core`. Most at risk: the file has
+ *            no pathway to exercise the real pipeline.
+ *   MEDIUM — mock patterns present, some real `h()` usage — but the
+ *            mock count is still notable, so a parallel real-`h()`
+ *            test may be missing for specific scenarios.
+ *   LOW    — either no mocks, or mock count is dwarfed by real usage.
+ *
+ * Companion to the `validate` and `get_anti_patterns` tools: those
+ * tell an agent what to write; this one tells an agent which existing
+ * tests need strengthening.
+ */
+import { readdirSync, readFileSync, statSync } from 'node:fs'
+import { dirname, join, relative, resolve } from 'node:path'
+
+export type AuditRisk = 'high' | 'medium' | 'low'
+
+export interface TestAuditEntry {
+  /** Absolute path to the test file */
+  path: string
+  /** Path relative to the repo root for readable reporting */
+  relPath: string
+  /** Count of object-literal `{ type: ..., props: ..., children: ... }` patterns */
+  mockVNodeLiteralCount: number
+  /** Count of `vnode` / `mockVNode` / `createVNode` helper DEFINITIONS */
+  mockHelperCount: number
+  /**
+   * Count of CALLS to a known mock-helper name. Captures pervasiveness:
+   * a file with one helper definition and 50 call-sites has the same
+   * `mockHelperCount` (1) as one with zero calls, but very different
+   * exposure. This metric surfaces that.
+   */
+  mockHelperCallCount: number
+  /** Count of lines that look like real `h(...)` calls (`h(Tag, props)` / `h(Component, ...)` shape) */
+  realHCallCount: number
+  /** True if the file imports `h` from `@pyreon/core` */
+  importsH: boolean
+  /** Risk classification */
+  risk: AuditRisk
+}
+
+export interface TestAuditResult {
+  /** Repo root discovered by walking up for `packages/` */
+  root: string | null
+  /** Every test file scanned, sorted by risk (high → low) then path */
+  entries: TestAuditEntry[]
+  /** Total files scanned */
+  totalScanned: number
+}
+
+// ═══════════════════════════════════════════════════════════════════════════════
+// Discovery
+// ═══════════════════════════════════════════════════════════════════════════════
+
+function findMonorepoRoot(startDir: string): string | null {
+  let dir = resolve(startDir)
+  for (let i = 0; i < 30; i++) {
+    try {
+      if (statSync(join(dir, 'packages')).isDirectory()) return dir
+    } catch {
+      // fall through to parent walk
+    }
+    const parent = dirname(dir)
+    if (parent === dir) return null
+    dir = parent
+  }
+  return null
+}
+
+function walkTestFiles(dir: string, out: string[], depth = 0): void {
+  if (depth > 10) return
+  let entries: string[]
+  try {
+    entries = readdirSync(dir)
+  } catch {
+    return
+  }
+  for (const name of entries) {
+    if (name.startsWith('.')) continue
+    if (name === 'node_modules' || name === 'lib' || name === 'dist') continue
+    const full = join(dir, name)
+    let isDir = false
+    try {
+      isDir = statSync(full).isDirectory()
+    } catch {
+      continue
+    }
+    if (isDir) {
+      walkTestFiles(full, out, depth + 1)
+      continue
+    }
+    if (/\.test\.(ts|tsx)$/.test(name)) {
+      out.push(full)
+    }
+  }
+}
+
+// ═══════════════════════════════════════════════════════════════════════════════
+// Pattern detection
+// ═══════════════════════════════════════════════════════════════════════════════
+
+/**
+ * Matches an object literal carrying `type`, `props`, AND `children`
+ * keys — the canonical mock-vnode shape. The `s` flag spans newlines
+ * because vnode literals often wrap across multiple lines.
+ */
+const MOCK_VNODE_LITERAL_PATTERN =
+  /\{\s*type\s*:[^{}]*?(?:(?:\{[^{}]*\})[^{}]*?)*?props\s*:[^{}]*?(?:(?:\{[^{}]*\})[^{}]*?)*?children\s*:[^{}]*?(?:(?:\{[^{}]*\})[^{}]*?)*?\}/gs
+
+/**
+ * Matches a helper definition that produces a mock vnode. Recognises:
+ *   const vnode = (...) => ({ type, props, children })
+ *   const mockVNode = ({ type, props, children })
+ *   function createVNode(type, props, children)
+ *
+ * Does NOT match bindings that merely STORE a real VNode with a
+ * `vnode`-like name, which are common in component tests:
+ *   const vnode = defaultRender(...)         // real render result
+ *   const vnode = <span>cell content</span>  // real JSX expression
+ *   const vnode = h('div', null, 'x')        // real h() call
+ *
+ * Distinguisher: a mock helper definition either
+ *   (a) starts an arrow function / function — RHS begins with `(` or the
+ *       keyword `function`, OR
+ *   (b) is itself an inline object literal — RHS begins with `{`.
+ * `const vnode = <anything else>` is a binding, not a definition.
+ */
+const MOCK_HELPER_PATTERN =
+  /(?:(?:const|let)\s+(?:mockV[Nn]ode|vnode|createV[Nn]ode|V[Nn]odeMock|makeV[Nn]ode)\s*=\s*(?:\(|\{|function\b|async\s))|(?:function\s+(?:mockV[Nn]ode|vnode|createV[Nn]ode|V[Nn]odeMock|makeV[Nn]ode)\s*\()/g
+
+/**
+ * Matches CALLS to a known mock-helper name:
+ *   vnode('div', props, children)
+ *   mockVNode(Component, props)
+ *   createVNode(...)
+ *
+ * Non-word boundary before the name avoids hits inside other
+ * identifiers (`hasVNode`, `myVnodeImpl`). The helper-def pattern
+ * above ALSO matches definitions' own `<name>(` arg list, so the
+ * caller should subtract definition count from call count to get
+ * usage-only density — but for risk classification, the combined
+ * signal (any mock-helper activity) is what we want.
+ */
+const MOCK_HELPER_CALL_PATTERN =
+  /(?:^|[^a-zA-Z0-9_])(?:mockV[Nn]ode|vnode|createV[Nn]ode|V[Nn]odeMock|makeV[Nn]ode)\s*\(/g
+
+/**
+ * Matches calls to `h(…)` where the first arg is an uppercase
+ * identifier (component) or a lowercase string tag — the two real
+ * shapes. Avoids matching:
+ *   hasSomething(...)   — h followed by [a-z]
+ *   ch()                — single h as substring of another name
+ *   hash()              — same
+ * The `(?:^|\W)` boundary plus `[A-Z'"\s]` arg requirement handles both.
+ */
+const REAL_H_CALL_PATTERN = /(?:^|\W)h\s*\(\s*["'A-Z]/g
+
+const IMPORT_H_PATTERN = /import\s*(?:type\s*)?\{[^}]*\bh\b[^}]*\}\s*from\s*['"]@pyreon\/core['"]/
+
+/**
+ * Predicate: does the `{type, props, children}` literal at this
+ * position appear as an argument to a type-guard-like call
+ * (`isDocNode(...)`, `hasVNode(...)`, `assertVNode(...)`, etc.)?
+ *
+ * Type guards take any object shape and return boolean — passing a
+ * `{type, props, children}` literal there is testing the guard's
+ * duck-typing, not building a mock vnode for a rendering pipeline.
+ * False-positive coverage for `utils-coverage.test.ts` and similar.
+ */
+function isLiteralInsideTypeGuardCall(source: string, literalStart: number): boolean {
+  // Scan back ~60 chars from the literal for `(\b(?:is|has|assert|validate|check)[A-Z]\w*\s*\()`.
+  // We're looking for a function-call opening paren that directly
+  // contains this literal (no closer `)` in between).
+  const window = source.slice(Math.max(0, literalStart - 60), literalStart)
+  // The nearest `(` before the literal — count unmatched parens.
+  let unmatched = 0
+  let openAt = -1
+  for (let i = window.length - 1; i >= 0; i--) {
+    const ch = window[i]
+    if (ch === ')') unmatched++
+    else if (ch === '(') {
+      if (unmatched === 0) {
+        openAt = i
+        break
+      }
+      unmatched--
+    }
+  }
+  if (openAt < 0) return false
+  // Is the token immediately before `openAt` an is*/has*/assert*/check*/validate* identifier?
+  const head = window.slice(0, openAt)
+  return /\b(?:is|has|assert|validate|check)[A-Z]\w*\s*$/.test(head)
+}
+
+/**
+ * Mask the inside of every backtick-delimited template-literal with
+ * spaces. Preserves length so positions/lines/columns stay aligned.
+ * Used to keep the literal scanner from counting `{type,props,children}`
+ * patterns that live inside test FIXTURE strings (the `cli/doctor.test.ts`
+ * case — those are fixtures for the audit tool itself, not actual code).
+ *
+ * Limitations: doesn't parse `${...}` interpolations precisely. If a
+ * fixture contains a balanced `${ ... }` with code we'd want scanned,
+ * the surrounding template string still masks it. In practice, mock-
+ * vnode literals are never interpolation expressions, so this is fine.
+ */
+function maskTemplateStrings(source: string): string {
+  return source.replace(/`(?:\\.|[^`\\])*`/g, (m) => `\`${' '.repeat(m.length - 2)}\``)
+}
+
+function countMatches(source: string, pattern: RegExp): number {
+  let count = 0
+  pattern.lastIndex = 0
+  while (pattern.exec(source) !== null) count++
+  pattern.lastIndex = 0
+  return count
+}
+
+/**
+ * Counts `{type, props, children}` literals, skipping those that
+ * appear inside a type-guard-looking call OR inside a template-literal
+ * (which is fixture text, not code). Dedicated because the existing
+ * `countMatches` helper has no context-aware skip.
+ */
+function countMockVNodeLiterals(source: string): number {
+  // First mask template-literal contents — fixtures inside backticks
+  // (e.g. `\`const v = { type, props, children }\`` written via
+  // writeFile in audit's own test) shouldn't count. The mask
+  // preserves positions, so the type-guard skip logic still works.
+  const masked = maskTemplateStrings(source)
+  const pattern = MOCK_VNODE_LITERAL_PATTERN
+  let count = 0
+  pattern.lastIndex = 0
+  let m: RegExpExecArray | null
+  while (true) {
+    m = pattern.exec(masked)
+    if (m === null) break
+    if (!isLiteralInsideTypeGuardCall(masked, m.index)) count++
+  }
+  pattern.lastIndex = 0
+  return count
+}
+
+function classifyRisk(entry: Omit<TestAuditEntry, 'risk'>): AuditRisk {
+  const mocks =
+    entry.mockVNodeLiteralCount + entry.mockHelperCount + entry.mockHelperCallCount
+  if (mocks === 0) return 'low'
+  if (!entry.importsH && entry.realHCallCount === 0) return 'high'
+  if (entry.realHCallCount >= mocks) return 'low'
+  return 'medium'
+}
+
+// ═══════════════════════════════════════════════════════════════════════════════
+// Public API
+// ═══════════════════════════════════════════════════════════════════════════════
+
+export function auditTestEnvironment(startDir: string): TestAuditResult {
+  // Caller-supplied `startDir` (no default) — `runtime-dom` transitively
+  // pulls this file via the `@pyreon/compiler` JSX runtime entry, and its
+  // tsconfig narrows `process` to `{ env: ... }` only. Calling
+  // `process.cwd()` here breaks that typecheck. MCP / CLI both have full
+  // node types; let them resolve cwd at the call site.
+  const root = findMonorepoRoot(startDir)
+  if (!root) return { root: null, entries: [], totalScanned: 0 }
+
+  const files: string[] = []
+  walkTestFiles(join(root, 'packages'), files)
+
+  const entries: TestAuditEntry[] = []
+  for (const path of files) {
+    let source: string
+    try {
+      source = readFileSync(path, 'utf8')
+    } catch {
+      continue
+    }
+    // Skip the scanner's own test fixtures so `audit_test_environment`
+    // doesn't report itself.
+    if (path.includes('test-audit.test.ts') || path.includes('test-audit-fixture')) {
+      continue
+    }
+
+    // Mask template-literal contents once, then run every counter
+    // against the masked source. Patterns inside backticks are
+    // FIXTURE strings (the audit tool's own test fixtures, doctest
+    // examples, etc.) — they shouldn't count toward any metric.
+    // `countMockVNodeLiterals` already does its own masking and runs
+    // on `source` so it can do its own work; we pass `source` to
+    // keep that contract intact.
+    const masked = maskTemplateStrings(source)
+    const mockVNodeLiteralCount = countMockVNodeLiterals(source)
+    const mockHelperCount = countMatches(masked, MOCK_HELPER_PATTERN)
+    const mockHelperCallCount = countMatches(masked, MOCK_HELPER_CALL_PATTERN)
+    const realHCallCount = countMatches(masked, REAL_H_CALL_PATTERN)
+    const importsH = IMPORT_H_PATTERN.test(masked)
+
+    const base = {
+      path,
+      relPath: relative(root, path),
+      mockVNodeLiteralCount,
+      mockHelperCount,
+      mockHelperCallCount,
+      realHCallCount,
+      importsH,
+    }
+    entries.push({ ...base, risk: classifyRisk(base) })
+  }
+
+  const riskRank = { high: 0, medium: 1, low: 2 }
+  entries.sort((a, b) => {
+    const cmp = riskRank[a.risk] - riskRank[b.risk]
+    if (cmp !== 0) return cmp
+    return a.relPath.localeCompare(b.relPath)
+  })
+
+  return { root, entries, totalScanned: files.length }
+}
+
+// ═══════════════════════════════════════════════════════════════════════════════
+// Formatter
+// ═══════════════════════════════════════════════════════════════════════════════
+
+export interface AuditFormatOptions {
+  /** Only include entries at or above this risk level. Default 'medium'. */
+  minRisk?: AuditRisk | undefined
+  /** Maximum entries to show per risk group. Default 20. */
+  limit?: number | undefined
+}
+
+function riskAtOrAbove(risk: AuditRisk, min: AuditRisk): boolean {
+  const rank = { high: 0, medium: 1, low: 2 }
+  return rank[risk] <= rank[min]
+}
+
+export function formatTestAudit(
+  result: TestAuditResult,
+  { minRisk = 'medium', limit = 20 }: AuditFormatOptions = {},
+): string {
+  if (!result.root) {
+    return (
+      'No monorepo root found. This tool scans `packages/**/*.test.{ts,tsx}` ' +
+      'for mock-vnode patterns. Run the MCP from the Pyreon repo root to ' +
+      'get useful output.'
+    )
+  }
+
+  const relevant = result.entries.filter((e) => riskAtOrAbove(e.risk, minRisk))
+  const counts = {
+    high: result.entries.filter((e) => e.risk === 'high').length,
+    medium: result.entries.filter((e) => e.risk === 'medium').length,
+    low: result.entries.filter((e) => e.risk === 'low').length,
+  }
+  const withMocks = result.entries.filter(
+    (e) => e.mockVNodeLiteralCount + e.mockHelperCount > 0,
+  ).length
+
+  const parts: string[] = []
+  parts.push(`# Test environment audit — ${result.totalScanned} test files scanned`)
+  parts.push('')
+  parts.push(
+    `**Mock-vnode exposure**: ${withMocks} / ${result.totalScanned} files construct \`{ type, props, children }\` literals or a custom \`vnode()\` helper instead of going through the real \`h()\` from \`@pyreon/core\`. This is the bug class that caused PR #197's silent metadata drop — mock-only tests pass while the real pipeline (rocketstyle attrs, compiler transforms, props forwarding) stays unexercised.`,
+  )
+  parts.push('')
+  parts.push(`**Risk counts**: ${counts.high} high · ${counts.medium} medium · ${counts.low} low`)
+  parts.push('')
+
+  if (relevant.length === 0) {
+    parts.push(`No files at risk level "${minRisk}" or above. Every test file either avoids mocks entirely or pairs them with real-\`h()\` coverage.`)
+    return parts.join('\n')
+  }
+
+  const byRisk = new Map<AuditRisk, TestAuditEntry[]>()
+  for (const entry of relevant) {
+    if (!byRisk.has(entry.risk)) byRisk.set(entry.risk, [])
+    byRisk.get(entry.risk)!.push(entry)
+  }
+
+  for (const [risk, group] of byRisk) {
+    const shown = group.slice(0, limit)
+    parts.push(`## ${risk.toUpperCase()} — ${group.length} file${group.length === 1 ? '' : 's'}${shown.length < group.length ? ` (showing ${shown.length})` : ''}`)
+    parts.push('')
+    parts.push(describeRisk(risk))
+    parts.push('')
+    for (const entry of shown) {
+      const mocks =
+        entry.mockVNodeLiteralCount + entry.mockHelperCount + entry.mockHelperCallCount
+      const breakdown: string[] = []
+      if (entry.mockVNodeLiteralCount > 0) breakdown.push(`${entry.mockVNodeLiteralCount} literal${entry.mockVNodeLiteralCount === 1 ? '' : 's'}`)
+      if (entry.mockHelperCount > 0) breakdown.push(`${entry.mockHelperCount} helper${entry.mockHelperCount === 1 ? '' : 's'}`)
+      if (entry.mockHelperCallCount > 0) breakdown.push(`${entry.mockHelperCallCount} helper call${entry.mockHelperCallCount === 1 ? '' : 's'}`)
+      const hSide =
+        entry.realHCallCount > 0
+          ? `${entry.realHCallCount} real h() call${entry.realHCallCount === 1 ? '' : 's'}`
+          : entry.importsH
+            ? `imports h but 0 calls found`
+            : `no h import`
+      parts.push(`- ${entry.relPath} — ${mocks} mock signal${mocks === 1 ? '' : 's'} (${breakdown.join(' + ')}), ${hSide}`)
+    }
+    parts.push('')
+  }
+
+  parts.push('---')
+  parts.push('')
+  parts.push(
+    'Fix: for each HIGH file, add at least one test that imports `h` from `@pyreon/core` and renders the actual component through `h(RealComponent, props)`. The mock version can stay for speed — it is the LACK of a real-`h()` parallel that blocks bug surfacing.',
+  )
+  return parts.join('\n')
+}
+
+function describeRisk(risk: AuditRisk): string {
+  if (risk === 'high') {
+    return 'Mock patterns present, no real `h()` calls, and no `h` import from `@pyreon/core`. The file has no pathway to exercise the real pipeline — bugs like PR #197 would slip through.'
+  }
+  if (risk === 'medium') {
+    return 'Mock patterns present AND some real `h()` usage — but mocks outnumber real calls, so specific scenarios may be mock-only. Spot-check that each contract the tests assert on goes through at least one real-`h()` path.'
+  }
+  return 'Mocks dwarfed by real usage OR no mocks at all — low risk.'
+}

package/src/tests/depth-stress.test.ts

@@ -0,0 +1,16 @@
+import { transformJSX } from '../jsx'
+
+const t = (code: string) => transformJSX(code, 'input.tsx').code
+
+describe('T0.2 — chain depth stress test', () => {
+  for (const depth of [4, 5, 10, 20, 50]) {
+    test(`depth ${depth} chain compiles without crashing`, () => {
+      const lines = ['const v0 = props.x']
+      for (let i = 1; i <= depth; i++) lines.push(`const v${i} = v${i - 1} + 1`)
+      const code = `function Comp(props) { ${lines.join('; ')}; return <div>{v${depth}}</div> }`
+      const result = t(code)
+      expect(result).toContain('props.x')
+      expect(result).toContain('_bind')
+    })
+  }
+})

package/src/tests/detector-tag-consistency.test.ts

@@ -0,0 +1,86 @@
+import { readFileSync } from 'node:fs'
+import { dirname, resolve } from 'node:path'
+import { fileURLToPath } from 'node:url'
+
+// Drift guard between the static `detectPyreonPatterns` detector codes
+// and the `[detector: <code>]` annotations on `.claude/rules/anti-patterns.md`.
+// Without this test, a new bullet can land without a detector tag, or
+// a detector code can be renamed without updating the doc. Either
+// direction is a silent inconsistency — consumers read the doc and
+// expect the detector to back it up.
+//
+// The test does one thing: every `[detector: CODE]` tag in the doc
+// must reference a PyreonDiagnosticCode (or the literal `N/A` for
+// bullets explicitly declared doc-only), and every
+// PyreonDiagnosticCode must appear at least once in the doc so the
+// tag-documentation loop is closed.
+
+const HERE = dirname(fileURLToPath(import.meta.url))
+const REPO_ROOT = resolve(HERE, '../../../../../')
+const ANTI_PATTERNS_PATH = resolve(REPO_ROOT, '.claude/rules/anti-patterns.md')
+
+// Kept in sync with the `PyreonDiagnosticCode` union in
+// `pyreon-intercept.ts`. When adding a new code, ALSO add a bullet
+// (with the `[detector: <code>]` tag) to `anti-patterns.md`.
+const KNOWN_CODES = [
+  'for-missing-by',
+  'for-with-key',
+  'props-destructured',
+  'process-dev-gate',
+  'empty-theme',
+  'raw-add-event-listener',
+  'raw-remove-event-listener',
+  'date-math-random-id',
+  'on-click-undefined',
+  'signal-write-as-call',
+  'static-return-null-conditional',
+  'as-unknown-as-vnodechild',
+] as const
+type KnownCode = (typeof KNOWN_CODES)[number]
+
+function readAntiPatterns(): string {
+  return readFileSync(ANTI_PATTERNS_PATH, 'utf8')
+}
+
+function extractDetectorTags(doc: string): string[] {
+  const re = /\[detector:\s*([a-z-/ ]+?)\]/gi
+  const found: string[] = []
+  for (const m of doc.matchAll(re)) {
+    // Some bullets document multiple codes on one pattern, e.g.
+    // `[detector: raw-add-event-listener / raw-remove-event-listener]`.
+    const raw = m[1]!
+    for (const code of raw.split('/')) {
+      const trimmed = code.trim()
+      if (trimmed) found.push(trimmed)
+    }
+  }
+  return found
+}
+
+describe('anti-patterns.md detector tags vs PyreonDiagnosticCode', () => {
+  const doc = readAntiPatterns()
+  const tags = extractDetectorTags(doc)
+
+  it('every [detector: CODE] tag references a known PyreonDiagnosticCode', () => {
+    const validCodes = new Set<string>(KNOWN_CODES)
+    const unknown = tags.filter((t) => !validCodes.has(t) && t !== 'N/A')
+    expect(unknown).toEqual([])
+  })
+
+  it('every PyreonDiagnosticCode appears at least once as a [detector:] tag', () => {
+    const tagSet = new Set(tags)
+    const missing: KnownCode[] = []
+    for (const code of KNOWN_CODES) {
+      if (!tagSet.has(code)) missing.push(code)
+    }
+    // If this fails, add a bullet for the new detector code to
+    // `.claude/rules/anti-patterns.md` with the `[detector: <code>]`
+    // suffix. The doc is the human-readable catalog; the detector is
+    // the static enforcement arm — they have to name each other.
+    expect(missing).toEqual([])
+  })
+
+  it('reports at least as many tags as detector codes (multi-code bullets allowed)', () => {
+    expect(tags.length).toBeGreaterThanOrEqual(KNOWN_CODES.length)
+  })
+})