gstack/scripts/test-free-shards.ts

#!/usr/bin/env bun
/**
 * test-free-shards — enumerate, shard, curate, and run the free test suite.
 *
 * Four jobs:
 *   1. Enumeration. Walk `browse/test/`, `test/`, `make-pdf/test/` and return
 *      every `*.test.{ts,tsx,js,jsx,mjs,cjs}` that isn't a paid-eval test.
 *   2. Sharding. Duration-pack local and isolated CI runs. Legacy --shard
 *      selection retains stable hash assignment.
 *   3. Curation (Windows-safe filter). Scan each test's content for POSIX-only
 *      patterns (`/bin/bash`, `sh -c`, raw `/tmp/`, `chmod`, `xargs`). Files
 *      that match are excluded from the Windows-safe subset — they would fail
 *      on `windows-latest` no matter how the runner shards them.
 *   4. Execution. Spawn `bun test` children and refuse to trust their exit
 *      code alone: every byte of output is classified through
 *      scripts/test-strict-output.ts, so a child that exits 0 without bun's
 *      terminal summary (a mid-suite process.exit truncation), with `(fail)`
 *      result lines, or with fewer files run than planned is a FAILURE. An
 *      external wall-clock timeout SIGKILLs the child's process group and
 *      reports the shard as timed-out — distinct from failed.
 *
 * Execution strategy (decision ledger V3/D6 — evaluate the Bun built-in
 * first; probed 2026-08 on Bun 1.3.13):
 *   - Full-suite runs (`bun test` via package.json, `bun run test:free`) use
 *     N CONCURRENT SHARD PROCESSES, serial within each (the paid runner's
 *     model). A single `--parallel` invocation was probed and initially
 *     adopted, then abandoned: three distinct Bun 1.3.13 worker pathologies
 *     (segfault + crash-retry wedge, skipped-file hooks stalling a worker,
 *     spawn-heavy files hanging under load) each stalled the whole
 *     invocation, while process shards isolate any wedge to its own shard.
 *     Original --parallel probe results, kept for the record: it
 *     showed --parallel (a) prints the standard `Ran N tests across M files`
 *     terminal summary, (b) exits non-zero when any file fails, (c) runs each
 *     file in its own worker process (distinct pids, no shared globals), and
 *     (d) converts a mid-suite process.exit(0) — which silently truncates a
 *     serial run at exit 0 — into a per-file `(crashed: exited)` failure with
 *     a complete summary and exit 1. Strictly SAFER than the serial path and
 *     ~2x faster on a 6-file probe (0.22s -> 0.11s wall, 280% CPU); the win
 *     grows with suite size since the serial suite measured 454s.
 *   - Legacy runs (`--shards M --shard i`) keep the hash-partitioned
 *     one-child-per-shard path. Cross-runner partitioning must be
 *     deterministic and per-file stable, so bun's own `--shard=M/N`
 *     (round-robin over sorted paths — every assignment shifts when a file
 *     lands) is not used, and there are no static per-file weight lists.
 *     Shard indices are STABLE: assignFilesToShards never renumbers on
 *     occupancy, and an empty shard is a fast no-op success.
 *
 * Adapted from the McGluut/gstack fork's test-free-shards.ts (190 LOC). The
 * Windows-safe filter is upstream-original — codex flagged that sharding alone
 * doesn't fix POSIX-bound tests, so we curate the subset that actually runs
 * on the windows-latest CI job.
 *
 * Output contract (v1.66): the full child stream ALWAYS lands in a per-run
 * private log under .context/free-test-logs (path printed at start and in the
 * epilogue). The console is quiet by default — only the runner's own
 * [test:free] lines, `(fail)` result lines, bun error/crash markers
 * (`error:`, `panic:`, `crashed`, `Unhandled error`), and the terminal
 * `Ran N tests across M files` summary reach it; `--verbose` restores full
 * forwarding. After every run a stable epilogue names the failing tests
 * (attributed to files via bun's `path/to/file.test.ts:` chunk headers),
 * crashed+retried workers, and — on a wall-timeout kill — the wedge-suspect
 * files. The strict classifier consumes the FULL stream regardless of what
 * the console shows.
 *
 * Exit codes: 0 pass, 1 fail, 124 wall-clock timeout.
 *
 * Usage:
 *   bun run scripts/test-free-shards.ts                            # full suite, N concurrent shard processes
 *   bun run scripts/test-free-shards.ts --list                     # show all
 *   bun run scripts/test-free-shards.ts --windows-only --list      # show curated
 *   bun run scripts/test-free-shards.ts --windows-only             # run curated
 *   bun run scripts/test-free-shards.ts --shards 4 --shard 1       # one shard (CI matrix)
 *   bun run scripts/test-free-shards.ts --wall-timeout 600         # override the kill deadline
 *   bun run scripts/test-free-shards.ts --verbose                  # forward the full child stream
 *   bun run scripts/test-free-shards.ts --quick                    # explicit fast subset, not acceptance
 *   bun run scripts/test-free-shards.ts --ci-plan plan.json --shards 20
 *   bun run scripts/test-free-shards.ts --ci-run plan.json --shard 1 --result result.json
 *   bun run scripts/test-free-shards.ts --ci-verify plan.json --results results/
 */

import * as fs from 'fs';
import * as os from 'os';
import * as path from 'path';
import { spawn, spawnSync } from 'child_process';
import { StringDecoder } from 'node:string_decoder';
import { createHash, randomUUID } from 'node:crypto';
import { isPaidTestFile } from '../test/helpers/paid-test-set';
import {
  BunTestOutputClassifier,
  exactTestFileSelectors,
  installChildSignalForwarding,
  isTerminationRequested,
  killProcessGroup,
  BunFailureSummaryParser,
  parseBunFailureResult,
  normalizeRelativePath,
  strictTestExitCode,
  stripAnsiLine,
} from './test-strict-output';

export { normalizeRelativePath } from './test-strict-output';

const ROOT = path.resolve(import.meta.dir, '..');
// design/test was silently absent from BOTH the package.json test script and
// this list — design tests (including a teardown bomb) never ran in any CI
// or local free run. Keep the two lists in sync. This list is the single
// source of truth for free-suite roots: package.json's `test` script routes
// through this runner rather than passing its own directory globs.
export const TEST_ROOTS = [
  'browse/test',
  'test',
  'make-pdf/test',
  'design/test',
  // v1.65 orphan wire-in (decision D3a): these ran under NO script or CI —
  // written coverage that caught nothing. All were green on arrival.
  'ios-qa/daemon/test',
  'ios-qa/scripts',
  'browser-skills',
] as const;
const TEST_FILE_REGEX = /\.test\.(?:[cm]?[jt]s|tsx|jsx)$/;

// POSIX-only patterns that indicate a test will fail on windows-latest no
// matter how the runner shards. Codex's v1.18.0.0 review flagged the first
// three as concrete examples in the existing free suite (test/ship-version-sync.test.ts:72,
// test/helpers/providers/claude.ts:22, package.json:12). We scan the test's
// own content here so the filter stays automatic as new tests land. The
// "Windows-incompatible APIs" patterns at the bottom were added after the
// first windows-free-tests CI run surfaced concrete failure modes.
const WINDOWS_FRAGILE_PATTERNS: Array<{ pattern: RegExp; reason: string }> = [
  // Hardcoded POSIX shells / commands.
  { pattern: /['"`]\/bin\/(?:ba)?sh/, reason: 'hardcoded /bin/sh or /bin/bash' },
  { pattern: /spawnSync\(['"]sh['"],|spawn\(['"]sh['"],|exec\(['"]sh /, reason: 'spawn("sh", ...)' },
  { pattern: /['"]bash -c['"]|['"]sh -c['"]/, reason: 'bash -c / sh -c' },
  { pattern: /['"`]\/tmp\//, reason: 'raw /tmp/ path (use os.tmpdir())' },
  { pattern: /['"]chmod\b/, reason: 'chmod shell command' },
  { pattern: /['"]xargs\b/, reason: 'xargs pipeline' },
  { pattern: /\bwhich claude\b/, reason: 'which claude (use Bun.which)' },
  // Windows-incompatible APIs.
  { pattern: /\.mode\s*&\s*0o[0-7]+/, reason: 'POSIX file mode bitmask (mode & 0o600 etc — Windows fakes mode bits)' },
  { pattern: /\.endsWith\(['"]\//, reason: 'hardcoded forward-slash path assertion (Windows uses \\\\)' },
  { pattern: /['"]\.\/[a-zA-Z][^"']*['"]\)\s*\.\s*toBe\(true\)/, reason: 'forward-slash path comparison' },
  // Tests that spawn a bash shebang script in bin/ via spawnSync. Git Bash on
  // Windows can run `bash /path/to/script` but spawnSync(scriptPath, ...)
  // tries to execute the file directly via CreateProcess, which fails on the
  // shebang. The pattern matches `, 'bin'` as a path-join argument (closing
  // OR followed by another segment), which catches:
  //   - path.join(ROOT, 'bin', 'script-name')        — typical
  //   - join(import.meta.dir, '..', 'bin', 'name')   — destructured (diff-scope)
  //   - path.join(ROOT, 'bin')                       — bare BIN constant (brain-sync)
  { pattern: /,\s*['"]bin['"]\s*[,)]|['"]\.?\/?bin\/[a-z][\w-]+['"]/, reason: 'spawns bin/ shebang script (Windows CreateProcess does not parse shebangs)' },
  // Tests that launch a real Playwright browser. The windows-free-tests CI job
  // runs a curated subset that intentionally does NOT install Chromium —
  // browser bring-up on Windows is a separate concern (see PR #1238). Tests
  // matching `await foo.launch(` need Chromium and fail with "Executable
  // doesn't exist" on the runner.
  { pattern: /await\s+\w+\.launch\(/, reason: 'launches Playwright browser (Chromium not installed in windows-free CI)' },
  // Tests that spawn the browse server as a subprocess via `bun run server.ts`.
  // The Bun → server.ts → Playwright path is the same one that doesn't work
  // on Windows (PR #1238 windows-pty-bun-pty-fix). Tests typically set
  // BROWSE_HEADLESS_SKIP=1 to skip the browser launch but still need a working
  // server, which they don't get on Windows.
  { pattern: /BROWSE_HEADLESS_SKIP|spawn\(\[['"]bun['"],\s*['"]run['"]/, reason: 'spawns the browse server subprocess (Bun-driven path is Windows-broken)' },
];

// Explicit known-Windows-incompatible test files that don't fit a regex
// pattern. Listed here with the precise reason. Prefer adding a pattern above
// when possible; this list is for environment-/runtime-specific tests where
// the failure mode is structural rather than detectable via source-file scan.
export const KNOWN_WINDOWS_INCOMPATIBLE: Array<{ file: string; reason: string }> = [
  {
    file: 'test/qa-evidence-producer.test.ts',
    reason: 'executes the registered Linux native actor and its inotify observer; portable capture and Windows job behavior are covered by qa-evidence.test.ts',
  },
  {
    file: 'test/qa-functional-fixture.test.ts',
    reason: 'executes graceful POSIX signal cancellation; Bun on Windows uses TerminateProcess and cannot run the fixture SIGTERM cleanup handler',
  },
  {
    file: 'test/qa-functional-observer-atomic.test.ts',
    reason: 'exercises real Linux inotify inode and directory watches through libc.so.6; Windows has no equivalent kernel interface',
  },
  {
    file: 'test/docsync-report-interface.test.ts',
    reason: 'executes registered native documentation callbacks with their real Linux inotify write observer before the model boundary',
  },
  {
    file: 'test/setup-gbrain-fixture.test.ts',
    reason: 'the fixture invokes real POSIX detector/verifier helpers through executable shebang wrappers',
  },
  {
    file: 'test/hermetic-skills-seeding.test.ts',
    reason: 'seeds the POSIX PTY skill runtime, whose embedded shell paths require a POSIX temporary root',
  },
  {
    file: 'test/hermetic-wiring.test.ts',
    reason: 'its runtime contract check seeds the POSIX PTY skill runtime; the curated Windows lane does not run that harness',
  },
  {
    file: 'test/pty-workspace-trust.test.ts',
    reason: 'launches the POSIX PTY harness with a fake executable and bound skill runtime',
  },
  {
    file: 'test/host-config.test.ts',
    reason: 'asserts "claude" binary on PATH (only true when running inside Claude Code, not on bare CI runner)',
  },
  {
    file: 'browse/test/findport.test.ts',
    reason: 'asserts Bun.serve.stop() is fire-and-forget — Bun behavior differs on Windows for this polyfill',
  },
  // First full run of the expanded lane (v1.66, 13 → ~258 files) surfaced
  // seven POSIX-bound files the content patterns cannot see (their
  // POSIX-ness is what they TEST, or arrives via a variable). Receipts:
  // PR #2593 windows-free-tests run 31918591602.
  {
    file: 'test/codex-under-codex-detection.test.ts',
    reason: 'drives the rendered preflight bash under a hardcoded POSIX PATH (/usr/bin:/bin) — bash is unreachable through that PATH on Windows, so every case sees empty output (v1.67 windows lane run 95234224148)',
  },
  {
    file: 'test/regression-pr1169-build-app-sed.test.ts',
    reason: 'tests sed escape sequences in build-app.sh — sed/bash are the subject under test',
  },
  {
    file: 'test/setup-conductor-worktree.test.ts',
    reason: 'tests ln -snf symlink semantics in the setup script — POSIX ln is the subject under test',
  },
  {
    file: 'test/artifacts-init-migration.test.ts',
    reason: 'runs a bash migration script + jq against a scaffolded git state — POSIX toolchain paths break under cmd spawn',
  },
  {
    file: 'test/gstack-decision-semantic.test.ts',
    reason: 'installs a fake gbrain SHEBANG SHIM on PATH; Windows spawn cannot exec shebang scripts',
  },
  {
    file: 'test/question-log-hook.test.ts',
    reason: 'spawns the PostToolUse hook script (bash shebang) directly; Windows spawn cannot exec it',
  },
  {
    file: 'browse/test/browser-skills-e2e.test.ts',
    reason: 'asserts forward-slash tier paths (<repo>/browser-skills/) that resolve with backslashes on Windows',
  },
  {
    file: 'design/test/variants-retry-after.test.ts',
    reason: 'wall-clock retry-timing assertions — flaky on the slow windows-latest runner even with widened bounds',
  },
  // Round-2 census (PR #2593 run 31919227507) after the first seven:
  {
    file: 'test/skill-census.test.ts',
    reason: 'census walk throws at module load on Windows (skill-census.ts:63) — the skills-tree symlink layout needs Developer Mode that CI runners lack',
  },
  {
    file: 'browse/test/browser-manager-unit.test.ts',
    reason: 'wedges the shard to its wall deadline on windows-latest (in-flight at kill); needs a Windows repro to diagnose — macOS + Linux lanes cover the file',
  },
  // Round-3 census (PR #2593 run 31919871680): the round-2 wedge had been
  // TRUNCATING its shard, so these seven only surfaced once shard 2 completed.
  // All the same POSIX-environment classes: PID/cmdline identity probing,
  // bash scripts as the subject under test, env-scrubbed child spawns.
  {
    file: 'browse/test/server-embedder-terminal-port.test.ts',
    reason: 'identity-based terminal-agent kill probes PID/cmdline with POSIX semantics; teardown asserts fail on windows-latest',
  },
  {
    file: 'design/test/daemon-discovery.test.ts',
    reason: 'verifyIdentity matches a spawned daemon via /proc-style cmdline probing — POSIX identity semantics',
  },
  {
    file: 'test/context-save-hardening.test.ts',
    reason: 'bash context-save/migration scripts (HOME-unset semantics, random-suffix path) are the subject under test',
  },
  {
    file: 'test/eval-list-cli.test.ts',
    reason: 'spawns the eval:list CLI via bun with a constructed env — bun resolution fails under Windows spawn',
  },
  {
    file: 'test/memory-cache-injection.test.ts',
    reason: 'exercises hook/deny-enforcement shell scripts — POSIX toolchain is the subject under test',
  },
  {
    file: 'test/migrations-v1.65.0.0.test.ts',
    reason: 'bash migration script (bunx re-fetch, .done markers) is the subject under test',
  },
  {
    file: 'test/question-preference-hook.test.ts',
    reason: 'spawns the PreToolUse preference hook (shebang script) directly; Windows spawn cannot exec it',
  },
  // Round-4 census (PR #2593 run 31920052810): unhandled errors with no
  // (fail) lines — attributed statically (the lane had no log artifact yet).
  {
    file: 'browse/test/browser-skill-commands.test.ts',
    reason: 'spawnSkill spawns bun with a constructed env — bun resolution fails under Windows spawn (unhandled, no (fail) line)',
  },
  {
    file: 'browse/test/security-audit-r2.test.ts',
    reason: 'symlink-attack fixtures (evil-link) need Developer Mode CI runners lack; expect(toThrow) fires unhandled on Windows',
  },
  // CSO comprehensive execution is qualified only for Linux containers behind
  // the POSIX watchdog and Unix-domain registry broker. Keep the portable
  // static/parser contracts in the Windows lane while leaving these exact
  // containment suites to the Linux and macOS gates.
  {
    file: 'test/cso-preparation-adversarial.test.ts',
    reason: 'exercises POSIX prepared-tree and archive-cache containment for qualified Linux Docker execution, which Windows does not admit',
  },
  {
    file: 'test/cso-preparation-container.test.ts',
    reason: 'asserts POSIX permission and symlink semantics for inert exports consumed by qualified Linux Docker execution',
  },
  {
    file: 'test/cso-preparation-executor.test.ts',
    reason: 'executes the Linux Docker acquisition path and its Unix-domain registry broker; comprehensive execution is unavailable on Windows',
  },
  {
    file: 'test/cso-verification-cleanup.test.ts',
    reason: 'spawns the POSIX detached watchdog used by contained repair verification, which Windows intentionally leaves unavailable',
  },
  {
    file: 'test/cso-witness.test.ts',
    reason: 'tests the contained repair witness with POSIX private-directory and compiled-helper assumptions; comprehensive execution is unavailable on Windows',
  },
  {
    file: 'test/cso-scanner-cli.test.ts',
    reason: 'drives the prebuilt POSIX CSO launcher with /usr/bin/git and a POSIX-only PATH; native Windows launcher behavior is covered by the dedicated cso-windows-launcher gate',
  },
];

// Force-include overrides: files a WINDOWS_FRAGILE_PATTERNS regex excludes for
// a reason that does not actually apply to them. Each entry documents WHY the
// pattern hit is a false positive — the point of these files is Windows
// coverage, so auto-excluding them defeats the regression tests they carry.
const KNOWN_WINDOWS_SAFE: Array<{ file: string; reason: string }> = [
  {
    file: 'test/qa-evidence.test.ts',
    reason: 'invokes the production helper through Bun argv and exercises native Windows job cleanup, private file captures and backpressured receipt output',
  },
  {
    file: 'test/qa-evidence-selection.test.ts',
    reason: 'bin/ strings are literal dependency and Windows-selection assertions; no native actor or shebang command is launched',
  },
  {
    file: 'test/qa-deadline.test.ts',
    reason: 'launches the guard through Bun argv; mode assertions and POSIX signal cases are platform-gated, while Windows job cleanup must execute natively',
  },
  {
    file: 'test/qa-deadline-selection.test.ts',
    reason: 'bin/ strings are dependency-selection inputs; this suite never launches a shebang executable',
  },
  {
    file: 'test/shared-libs-source-reads.test.ts',
    reason: 'bin/ literal is a mocked launch assertion; actual worktree fingerprinting explicitly invokes Bash on Windows',
  },
  {
    file: 'test/claude-code-windows-job.test.ts',
    reason: 'invokes Bun directly; verifies Windows job containment at the standalone CLI boundary',
  },
  {
    file: 'test/claude-code-runner.test.ts',
    // The bin/ path is launched through process.execPath (Bun), never as a
    // shebang executable. Keep taskkill tree supervision in the Windows lane.
    reason: 'invokes the runner via Bun argv; fake CLI and timeout descendant assertions cover native Windows taskkill',
  },
  {
    file: 'test/setup-gbrain-remote-caller.test.ts',
    // bin is an expected PATH component; the adapter injects the SDK boundary
    // and never launches a shebang. Keep the native delimiter cases in CI.
    reason: 'replays the registered SDK callback with fixture-only bin paths; covers native Windows PATH composition',
  },
  {
    file: 'test/cso-windows-build-contract.test.ts',
    // bin is a temporary staging directory. The adapter injects spawnSync;
    // real PowerShell/native execution remains in cso-windows-launcher.
    reason: 'replays the native build callbacks with an injected subprocess; bin paths are staging fixtures, not shebang launches',
  },
  {
    file: 'test/setup-windows-rerun-refresh.test.ts',
    // Trips the "spawns bin/ shebang script" pattern via path.join(..., 'bin',
    // 'tool.sh') fixture paths, but every spawn goes through test/helpers/bash-script.ts
    // (bash <tempfile>) — Git Bash executes it fine on windows-latest, with no argv-length ceiling. This file IS
    // the #2444 Windows regression coverage (IS_WINDOWS=1 copy-refresh path);
    // excluding it here would keep the bug class unexercised on the one
    // platform it bites.
    reason: 'bin/ hits are fixture path segments; spawns bash explicitly — the IS_WINDOWS=1 refresh path must run on windows-latest',
  },
  {
    file: 'test/uninstall-windows-copies.test.ts',
    // Trips the "spawns bin/ shebang script" pattern via the
    // path.join(ROOT, 'bin', 'gstack-uninstall') constant, but the script is
    // always spawned through spawnSync('bash', [UNINSTALL, ...]). This file
    // carries the #2563 Windows real-dir-copy uninstall coverage — the bug
    // ONLY reproduces on the copy install shape windows-latest exercises.
    // The symlink-shape describe block self-skips on win32.
    reason: 'bin/ hit is a bash-spawned script path; #2563 real-dir uninstall coverage must run on windows-latest',
  },
  {
    file: 'browse/test/file-permissions.test.ts',
    // Trips the POSIX-mode-bitmask pattern, but every `mode & 0o777` assertion
    // is platform-guarded: win32-only tests return early, POSIX-only tests
    // guard the bitmask behind `process.platform !== 'win32'`, and the
    // symlink-skip regression test both wraps symlinkSync in try/catch
    // (runners without Developer Mode can't create symlinks) and guards its
    // bitmask — on win32 it asserts behavior (warns, skips, doesn't throw,
    // target stays usable), never fake Windows mode bits (dirs stat 0o777
    // there, so a 0o755 expectation fails on runner semantics, not our code).
    // This file carries the win32-only icacls-by-SID regression tests, which
    // can ONLY execute on windows-latest — excluding it here means the
    // machine-account ACL lockout regression is never exercised on the one
    // platform it bricks.
    reason: 'every mode-bitmask assertion is guarded off win32 (behavior asserted instead); win32-only ACL regression tests must run on windows-latest',
  },
  {
    file: 'browse/test/terminal-agent-owner-watchdog.test.ts',
    // Trips the spawn(['bun','run',...]) pattern, whose reason is the
    // Playwright-bound browse server. This test spawns terminal-agent.ts,
    // which imports only fs/path/crypto + local helpers (no Playwright, no
    // PTY at module scope) and boots under Bun on Windows — the owner-PID
    // orphan leak it pins was reported on Windows (#2019).
    reason: 'spawns terminal-agent (no Playwright), not the browse server; owner-orphan leak is a Windows defect',
  },
];

export const DEFAULT_SHARD_COUNT = 20;
// Per-test timeout passed to `bun test --timeout`. 30s matches what
// package.json's `test` script used before it was repointed at this runner —
// the runner is now the single owner of that semantic.
export const FREE_TEST_TIMEOUT_MS = 30_000;
// External wall-clock deadline per spawned child (whole shard or the single
// full-suite --parallel invocation). A wedged child — a spinning main thread
// no in-process --timeout timer can interrupt — is SIGKILLed at the group
// level and reported 'timed-out', distinct from 'failed'.
// ~3.5x the observed full-suite wall (~100-160s). A wedged run should be
// killed-and-diagnosed (the epilogue prints the in-flight suspects) in
// minutes, not sat out — 15min of silence was pure diagnosis latency.
// Override per run with --wall-timeout <secs>.
export const DEFAULT_WALL_TIMEOUT_MS = 6 * 60_000;
/**
 * Full-suite shards scale their wall deadline with shard size:
 * max(DEFAULT_WALL_TIMEOUT_MS, files × PER_FILE_WALL_MS). The 6-min floor
 * keeps wedge diagnosis fast on a typical ~70-file local shard, while a
 * low-core machine (jobs=1 → the whole suite in one shard) or the Windows
 * lane (~130 files/shard) gets proportional headroom instead of a false
 * timed-out kill of a healthy run. Explicit --wall-timeout disables scaling.
 */
export const PER_FILE_WALL_MS = 5_000;
export function wallTimeoutForShard(fileCount: number, baseMs = DEFAULT_WALL_TIMEOUT_MS): number {
  return Math.max(baseMs, fileCount * PER_FILE_WALL_MS);
}

/**
 * Wall for a duration-packed shard. The count heuristic above assumes count
 * approximates cost; LPT packing breaks that BY DESIGN (a shard may hold six
 * slow Playwright files), so packed shards get max(base, predicted x 3) —
 * generous against seed drift, still bounded.
 */
export function wallTimeoutForPackedShard(predictedMs: number, baseMs = DEFAULT_WALL_TIMEOUT_MS, fileCount = 0): number {
  // Predictions transfer badly across machines: the committed duration seed
  // is recorded on fast CI, and a syscall-supervised sandbox replays those
  // files 2-4x slower (observed: a 253-file shard predicted ~242s wall-killed
  // at its 725s predicted-x3 wall while genuinely still progressing). The
  // packed wall may therefore be LOOSER than the count heuristic, never
  // tighter — it keeps the per-file floor the runner has always guaranteed.
  return Math.max(baseMs, Math.ceil(predictedMs * 3), fileCount * PER_FILE_WALL_MS);
}
/**
 * Full-suite parallelism: use all available CPUs, with a floor of one and a
 * per-platform cap (maxFullSuiteJobs). Shards stay serial internally; separate
 * shard processes can overlap subprocess and I/O waits without a fixed CPU
 * reserve. Prefer availableParallelism() to honor CPU affinity (Bun also
 * honors a container's cgroup CPU quota there), falling back to cpus() on
 * runtimes without it. macOS and Windows keep the cap of 6: beyond that,
 * playwright-heavy shards contended on browser launches in the original
 * M-series measurement. Linux caps at 16: on a 16-vCPU Ubicloud VM with the
 * CI lane's environment (2026-09-28), 16 shards ran the complete suite in
 * 137s versus 327s for 6. More shards are not a guaranteed speedup; compare
 * complete-suite runs before raising either cap.
 *
 * GSTACK_FREE_JOBS overrides the computed count (the free runner's analogue
 * of the paid runner's EVALS_JOBS). Exists for syscall-supervised sandboxes:
 * on Vercel sandboxes, PID 1 (sandbox-init) installs a seccomp filter whose
 * user-space supervisor saturates under ~6 concurrent bun+playwright shards
 * and starts returning EACCES from plain file syscalls (measured: 200/200
 * `git init` probes in fresh mktemp dirs fail with
 * "Cannot access work tree: Permission denied" while the suite runs, 0/200
 * when idle — access(dir, X_OK) = EACCES under strace). Fewer shards keep
 * the supervisor inside its budget. Not clamped by maxFullSuiteJobs so a
 * beefy box can also raise it deliberately.
 */
export const MAX_FULL_SUITE_JOBS = 6;
export const MAX_LINUX_FULL_SUITE_JOBS = 16;

export function maxFullSuiteJobs(platform: NodeJS.Platform = process.platform): number {
  return platform === 'linux' ? MAX_LINUX_FULL_SUITE_JOBS : MAX_FULL_SUITE_JOBS;
}

export function fullSuiteJobs(platform: NodeJS.Platform = process.platform): number {
  const raw = process.env.GSTACK_FREE_JOBS;
  if (raw !== undefined && raw !== '') {
    // Strict digits-only: parseInt would silently truncate "2abc" -> 2 and
    // "3.7" -> 3, defeating the loud-failure contract the error text claims.
    if (!/^\d+$/.test(raw.trim()) || Number.parseInt(raw, 10) <= 0) {
      throw new Error(`GSTACK_FREE_JOBS must be a positive integer, got: ${raw}`);
    }
    return Number.parseInt(raw, 10);
  }
  const availableCpus = os.availableParallelism?.() ?? os.cpus().length;
  return Math.max(1, Math.min(maxFullSuiteJobs(platform), availableCpus));
}

/**
 * Files that crash or wedge Bun's --parallel WORKERS but run fine in a plain
 * serial process. Full-suite mode now uses shard PROCESSES (no workers), so
 * this list is inert placement-wise — retained as the paper trail of why the
 * one-invocation --parallel strategy was abandoned, and as the exclusion list
 * should anyone re-attempt it on a newer Bun.
 */
export const WORKER_HOSTILE: Record<string, string> = {
  'browse/test/security-live-playwright.test.ts':
    'Bun 1.3.13 segfaults running this file in a --parallel worker ("panic: '
    + 'Segmentation fault ... a bug in Bun"), and the crashed-worker retry then '
    + 'wedges the whole invocation past the wall clock. Passes serially.',
};

/**
 * Exclusive host-state fixtures: run in ONE serial shard AFTER the parallel
 * shards. The public name is retained for callers of the original tree-write
 * classification. Entries need a concrete shared-state hazard that fixture
 * directories cannot isolate, such as host-wide procfs visibility.
 * Keys are pinned against the live file census by test-free-shards.test.ts —
 * a renamed file fails the suite instead of silently dropping serialization.
 */
export const TREE_MUTATING: Record<string, string> = {
  'test/bootstrap-retention.test.ts': 'Creates nondumpable same-UID processes visible to every host procfs census; must not overlap other native-retention fixtures.',
};

export function isFreeTestFile(relativePath: string): boolean {
  const normalized = normalizeRelativePath(relativePath);
  if (!TEST_FILE_REGEX.test(normalized)) return false;
  return !isPaidTestFile(normalized);
}

/**
 * Returns the first POSIX-only pattern hit in the file, or null if Windows-safe.
 */
export function detectWindowsFragility(absolutePath: string): { reason: string } | null {
  let content: string;
  try {
    content = fs.readFileSync(absolutePath, 'utf-8');
  } catch {
    return null;
  }
  for (const { pattern, reason } of WINDOWS_FRAGILE_PATTERNS) {
    if (pattern.test(content)) return { reason };
  }
  return null;
}

function walkTestFiles(dirPath: string): string[] {
  const entries = fs.readdirSync(dirPath, { withFileTypes: true });
  const files: string[] = [];
  for (const entry of entries) {
    const fullPath = path.join(dirPath, entry.name);
    if (entry.isDirectory()) {
      files.push(...walkTestFiles(fullPath));
      continue;
    }
    if (TEST_FILE_REGEX.test(entry.name)) {
      files.push(fullPath);
    }
  }
  return files;
}

export function collectFreeTestFiles(rootDir = ROOT): string[] {
  const discovered = new Set<string>();
  for (const testRoot of TEST_ROOTS) {
    const absoluteRoot = path.join(rootDir, testRoot);
    if (!fs.existsSync(absoluteRoot)) continue;
    for (const fullPath of walkTestFiles(absoluteRoot)) {
      const relativePath = normalizeRelativePath(path.relative(rootDir, fullPath));
      if (isFreeTestFile(relativePath)) {
        discovered.add(relativePath);
      }
    }
  }
  return [...discovered].sort();
}

export interface CurationResult {
  safe: string[];
  excluded: Array<{ file: string; reason: string }>;
}

export function curateWindowsSafe(files: string[], rootDir = ROOT): CurationResult {
  const safe: string[] = [];
  const excluded: Array<{ file: string; reason: string }> = [];
  const knownBad = new Map(KNOWN_WINDOWS_INCOMPATIBLE.map((e) => [e.file, e.reason]));
  const knownSafe = new Set(KNOWN_WINDOWS_SAFE.map((e) => e.file));
  for (const relativePath of files) {
    const knownReason = knownBad.get(relativePath);
    if (knownReason) {
      excluded.push({ file: relativePath, reason: knownReason });
      continue;
    }
    if (knownSafe.has(relativePath)) {
      safe.push(relativePath);
      continue;
    }
    const absolute = path.join(rootDir, relativePath);
    const fragility = detectWindowsFragility(absolute);
    if (fragility) {
      excluded.push({ file: relativePath, reason: fragility.reason });
    } else {
      safe.push(relativePath);
    }
  }
  return { safe, excluded };
}

export function stableHash(input: string): number {
  let hash = 0x811c9dc5;
  for (let index = 0; index < input.length; index += 1) {
    hash ^= input.charCodeAt(index);
    hash = Math.imul(hash, 0x01000193);
  }
  return hash >>> 0;
}

/**
 * Hash-partition files across EXACTLY shardCount shards. Empty shards are
 * preserved: a file's shard index is a pure function of its own path and the
 * shard count, never of which other files happen to exist. A CI matrix keys
 * runners off the index, so filtering empty shards (the old behavior) would
 * renumber every later shard whenever occupancy shifted — runner 3 silently
 * running shard 4's files. An empty shard is instead a fast no-op success at
 * run time.
 */
export function assignFilesToShards(files: string[], shardCount: number): string[][] {
  if (!Number.isInteger(shardCount) || shardCount <= 0) {
    throw new Error(`Shard count must be a positive integer. Received: ${shardCount}`);
  }

  const shards = Array.from({ length: shardCount }, () => [] as string[]);
  for (const file of files) {
    const shardIndex = stableHash(file) % shardCount;
    shards[shardIndex].push(file);
  }

  return shards.map(filesInShard => filesInShard.sort());
}

// ─── Duration-aware packing (local full suite and explicit CI plans) ───────
// Hash sharding balances file COUNTS (~1.15x spread) but not cost: the 15
// Playwright-launching files land 4/3/4/1/2/1 across 6 shards, giving a
// measured 28s–97s shard spread and ~40s of idle tail on every run. LPT
// packing over recorded per-file durations reclaims most of it. The `--shard`
// legacy path is deliberately untouched — its contract is stable indices
// via assignFilesToShards/stableHash (empty shards no-op; see above).
//
// One store, no overlay: durations come from the committed seed
// (scripts/free-test-durations.json), refreshed occasionally via
// `--record-durations` (each file timed in its own child — exact, and immune
// to bun's stream buffering, where silent passers print no header to
// timestamp). GSTACK_FREE_TEST_DURATIONS overrides the path for experiments.
// The seed is a HINT, not a contract: missing file → hash-shard fallback;
// unknown file → 75th-percentile pessimism (placed early by LPT, bounding
// tail risk). CI shares one plan rather than independently recomputing it.

export const FREE_TEST_DURATIONS_FILE = 'scripts/free-test-durations.json';

export function loadFreeTestDurations(rootDir = ROOT): Record<string, number> | null {
  const file = process.env.GSTACK_FREE_TEST_DURATIONS
    ?? path.join(rootDir, FREE_TEST_DURATIONS_FILE);
  let raw: string;
  try {
    raw = fs.readFileSync(file, 'utf-8');
  } catch {
    return null; // no seed — hash sharding, silently (fresh checkouts are normal)
  }
  try {
    const parsed = JSON.parse(raw) as { durations?: Record<string, unknown> };
    const entries = Object.entries(parsed.durations ?? {})
      .filter((entry): entry is [string, number] =>
        typeof entry[1] === 'number' && Number.isFinite(entry[1]) && entry[1] >= 0);
    if (entries.length === 0) return null;
    return Object.fromEntries(entries);
  } catch (error) {
    // A corrupt seed (bad merge) must cost a warning, never the suite.
    console.error(`[test:free] WARNING: corrupt durations seed ${file} (${(error as Error).message}) — falling back to hash sharding`);
    return null;
  }
}

export interface PackedShards {
  shards: string[][];
  /** Predicted total per shard, aligned with `shards` — feeds walls + logs. */
  predictedMs: number[];
}

/**
 * Longest-processing-time-first bin packing: files sorted by predicted
 * duration (desc, path-stable tiebreak) each go to the currently-lightest
 * shard. Deterministic for a given (files, shardCount, durations).
 */
export function packShardsByDuration(
  files: string[],
  shardCount: number,
  durations: Record<string, number>,
): PackedShards {
  if (!Number.isInteger(shardCount) || shardCount <= 0) {
    throw new Error(`Shard count must be a positive integer. Received: ${shardCount}`);
  }
  const known = files
    .map((f) => durations[normalizeRelativePath(f)])
    .filter((v): v is number => typeof v === 'number')
    .sort((a, b) => a - b);
  // Unknown files get the 75th percentile of known durations: pessimistic, so
  // LPT places them early and a surprise long-runner can't recreate the tail.
  const fallback = known.length > 0 ? known[Math.min(known.length - 1, Math.floor(known.length * 0.75))] : 1;
  const predicted = (f: string): number => durations[normalizeRelativePath(f)] ?? fallback;

  const ordered = [...files].sort((a, b) => predicted(b) - predicted(a) || (a < b ? -1 : 1));
  const shards = Array.from({ length: shardCount }, () => [] as string[]);
  const loads = new Array<number>(shardCount).fill(0);
  for (const file of ordered) {
    let lightest = 0;
    for (let i = 1; i < shardCount; i += 1) {
      if (loads[i] < loads[lightest]) lightest = i;
    }
    shards[lightest].push(file);
    loads[lightest] += predicted(file);
  }
  return { shards: shards.map((s) => s.sort()), predictedMs: loads };
}

/**
 * Files missing from the duration seed are packed at the 75th percentile, so
 * one slow new file can become the whole run's long pole without anyone
 * noticing. Name them (on stderr: --ci-plan's stdout is the CI matrix).
 */
export function unseededFreeFiles(files: string[], durations: Record<string, number>): string[] {
  return files.filter((f) => durations[normalizeRelativePath(f)] === undefined);
}

function warnUnseededFreeFiles(files: string[], durations: Record<string, number>): void {
  const unseeded = unseededFreeFiles(files, durations);
  if (unseeded.length === 0) return;
  const shown = unseeded.slice(0, 5).join(', ') + (unseeded.length > 5 ? `, +${unseeded.length - 5} more` : '');
  console.error(`[test:free] ${unseeded.length} file(s) have no recorded duration and are packed at the 75th-percentile estimate: ${shown}.`
    + ' Refresh scripts/free-test-durations.json with `bun run test:ubicloud --record-durations`.');
}

export interface FreeCiPlan {
  version: 1;
  revision: string;
  id: string;
  shards: Array<{ shard: number; files: string[]; predictedMs: number }>;
}

const planDigest = (plan: Omit<FreeCiPlan, 'id'>): string =>
  createHash('sha256').update(JSON.stringify(plan)).digest('hex');

/** One immutable plan is shared by isolated CI machines; never repack per job. */
export function createFreeCiPlan(files: string[], count: number, durations: Record<string, number>, revision: string): FreeCiPlan {
  const readers = files.filter(file => !(file in TREE_MUTATING));
  const exclusive = files.filter(file => file in TREE_MUTATING).sort();
  const packed = packShardsByDuration(readers, count, durations);
  const shards = packed.shards.map((files, index) => ({ shard: index + 1, files, predictedMs: packed.predictedMs[index] }));
  if (exclusive.length) shards.push({ shard: shards.length + 1, files: exclusive, predictedMs: exclusive.reduce((ms, file) => ms + (durations[file] ?? 0), 0) });
  const body = { version: 1 as const, revision, shards };
  return { ...body, id: planDigest(body) };
}

export function validateFreeCiPlan(plan: FreeCiPlan, files: string[], revision: string): void {
  const { id, version, shards } = plan;
  if (version !== 1 || plan.revision !== revision || !Array.isArray(shards) || !shards.length
    || id !== planDigest({ version, revision: plan.revision, shards })) throw new Error('CI plan identity or revision mismatch');
  if (shards.some((shard, index) => shard.shard !== index + 1 || !Array.isArray(shard.files)
    || !Number.isFinite(shard.predictedMs) || shard.predictedMs < 0)) throw new Error('Invalid CI shard plan');
  const planned = shards.flatMap(shard => shard.files).sort();
  if (new Set(planned).size !== planned.length || JSON.stringify(planned) !== JSON.stringify([...files].sort())) throw new Error('CI plan must cover every free file exactly once');
}

export interface FreeCiResult {
  planId: string;
  revision: string;
  outcome: FreeShardOutcome;
  retry: FreeShardOutcome | null;
}

/** Preserve the full-suite retry cap across independently running CI jobs. */
export function eligibleFreeRetryFiles(outcomes: FreeShardOutcome[]): string[] | null {
  if (!outcomes.every(outcome => hasScopedFailureAttribution(outcome) && (outcome.status === 'passed'
    || (outcome.status === 'failed' && outcome.failingFiles.length > 0 && outcome.unattributedFailures === 0)))) return null;
  const files = [...new Set(outcomes.flatMap(outcome => outcome.failingFiles))];
  return files.length > 0 && files.length <= 5 ? files : null;
}

function hasScopedFailureAttribution(outcome: FreeShardOutcome): boolean {
  return new Set(outcome.failingFiles).size === outcome.failingFiles.length
    && outcome.failingFiles.every(file => outcome.files.includes(file));
}

export function verifyFreeCiResults(plan: FreeCiPlan, results: FreeCiResult[]): void {
  if (results.length !== plan.shards.length) throw new Error('Missing or duplicate CI shard results');
  const seen = new Set<number>();
  for (const result of results) {
    const outcome = result.outcome;
    const shard = plan.shards[outcome.shard - 1];
    if (result.planId !== plan.id || result.revision !== plan.revision || !shard || seen.has(outcome.shard)
      || JSON.stringify(outcome.files) !== JSON.stringify(shard.files)) throw new Error('CI result identity, shard or file coverage mismatch');
    seen.add(outcome.shard);
    if (!hasCompleteCiSummary(outcome)) throw new Error('Missing or incomplete CI execution summary');
    if (outcome.status === 'passed') {
      if (outcome.exitCode !== 0 || outcome.failingFiles.length || outcome.unattributedFailures || result.retry) throw new Error('Inconsistent passing CI result');
    } else {
      const retryFiles = eligibleFreeRetryFiles([outcome]);
      const retry = result.retry;
      if (!retryFiles || !retry || retry.status !== 'passed' || retry.exitCode !== 0
        || retry.failingFiles.length || retry.unattributedFailures
        || !hasCompleteCiSummary(retry)
        || JSON.stringify([...retry.files].sort()) !== JSON.stringify(retryFiles.sort())) throw new Error('Failed or incomplete CI shard');
    }
  }
  if (results.some(result => result.retry) && !eligibleFreeRetryFiles(results.map(result => result.outcome))) {
    throw new Error('CI retries exceed the full-suite attribution or five-file limit');
  }
}

function hasCompleteCiSummary(outcome: FreeShardOutcome): boolean {
  const summary = outcome.summary;
  if (!summary || !Number.isInteger(summary.testsRan) || summary.testsRan! < 0
    || summary.filesRan !== outcome.files.length) return false;
  // Empty assigned shards deliberately do not launch Bun or invent a summary.
  return outcome.files.length === 0
    ? summary.testsRan === 0 && summary.sawTerminalSummary === false
    : summary.sawTerminalSummary === true;
}

export const QUICK_CORE = [
  'test/strict-output.test.ts', 'test/gen-skill-docs.test.ts',
  'test/skill-check-driver.test.ts',
  'test/skill-ceo-section-ordering.test.ts',
  'test/qa-functional-observer.test.ts', 'test/qa-checkpoint-evidence.test.ts',
  'test/test-free-shards-capture.test.ts',
];

export function selectQuickFreeFiles(files: string[], durations: Record<string, number>): string[] {
  return files.filter(file => isFreeTestFile(file)
    && (QUICK_CORE.includes(file) || (durations[file] !== undefined && durations[file] <= 2_000)));
}

export interface BuildShardArgsOptions {
  /**
   * Pass bun's --parallel (worker-per-file, implies --isolate). No production
   * caller today — full-suite mode uses N shard PROCESSES after the worker
   * pathologies documented in main(); retained for a future re-attempt on a
   * newer Bun (see WORKER_HOSTILE).
   */
  parallel?: boolean;
  rootDir?: string;
}

export function buildShardArgs(files: string[], options: BuildShardArgsOptions = {}): string[] {
  // Exact absolute selectors: bun treats positional test paths as substring
  // filters, so a relative `test/x.test.ts` would ALSO select
  // `browse/test/x.test.ts` — shard bleed that double-runs files.
  const selectors = exactTestFileSelectors(files, options.rootDir ?? ROOT);
  const args = ['test', ...selectors, `--timeout=${FREE_TEST_TIMEOUT_MS}`];
  if (options.parallel) args.push('--parallel');
  else args.push('--max-concurrency=1');
  return args;
}

type CliOptions = {
  dryRun: boolean;
  listOnly: boolean;
  recordDurations: boolean;
  windowsOnly: boolean;
  verbose: boolean;
  shardCount: number;
  shardIndex: number | null;
  wallTimeoutMs: number;
  /** True when --wall-timeout was passed explicitly; full-suite mode only auto-scales the default. */
  wallTimeoutExplicit: boolean;
  quick: boolean;
  ciPlan: string | null;
  ciRun: string | null;
  ciVerify: string | null;
  result: string | null;
  results: string | null;
};

function parseCliOptions(argv: string[]): CliOptions {
  let dryRun = false;
  let listOnly = false;
  let recordDurations = false;
  let windowsOnly = false;
  let verbose = false;
  let shardCount = DEFAULT_SHARD_COUNT;
  let shardIndex: number | null = null;
  let wallTimeoutMs = DEFAULT_WALL_TIMEOUT_MS;
  let wallTimeoutExplicit = false;
  let quick = false;
  const paths: Record<'ciPlan' | 'ciRun' | 'ciVerify' | 'result' | 'results', string | null> = {
    ciPlan: null, ciRun: null, ciVerify: null, result: null, results: null,
  };

  for (let index = 0; index < argv.length; index += 1) {
    const arg = argv[index];
    if (arg === '--dry-run') { dryRun = true; continue; }
    if (arg === '--list') { listOnly = true; continue; }
    if (arg === '--record-durations') { recordDurations = true; continue; }
    if (arg === '--windows-only') { windowsOnly = true; continue; }
    if (arg === '--verbose') { verbose = true; continue; }
    if (arg === '--quick') { quick = true; continue; }
    const pathKey = ({ '--ci-plan': 'ciPlan', '--ci-run': 'ciRun', '--ci-verify': 'ciVerify', '--result': 'result', '--results': 'results' } as const)[arg];
    if (pathKey) {
      const value = argv[++index];
      if (!value || value.startsWith('--')) throw new Error(`Missing path for ${arg}`);
      paths[pathKey] = value;
      continue;
    }
    if (arg === '--shards') {
      const value = argv[index + 1];
      if (!value) throw new Error('Missing value for --shards');
      shardCount = Number.parseInt(value, 10);
      index += 1;
      continue;
    }
    if (arg === '--shard') {
      const value = argv[index + 1];
      if (!value) throw new Error('Missing value for --shard');
      shardIndex = Number.parseInt(value, 10);
      index += 1;
      continue;
    }
    if (arg === '--wall-timeout') {
      const value = Number.parseInt(argv[index + 1] ?? '', 10);
      if (!Number.isInteger(value) || value <= 0) throw new Error('--wall-timeout needs a positive integer (seconds)');
      wallTimeoutMs = value * 1000;
      wallTimeoutExplicit = true;
      index += 1;
      continue;
    }
    throw new Error(`Unknown argument: ${arg}`);
  }

  const ciModes = [paths.ciPlan, paths.ciRun, paths.ciVerify].filter(Boolean).length;
  if (ciModes > 1 || (ciModes && (quick || listOnly || dryRun || recordDurations || windowsOnly))) throw new Error('CI modes cannot be combined with other selection modes');
  if (paths.ciRun && (shardIndex === null || !paths.result)) throw new Error('--ci-run requires --shard and --result');
  if (paths.ciVerify && !paths.results) throw new Error('--ci-verify requires --results');
  if (quick && (recordDurations || windowsOnly || shardIndex !== null)) throw new Error('--quick cannot change recording, Windows or shard selection');
  return { dryRun, listOnly, recordDurations, windowsOnly, verbose, shardCount, shardIndex, wallTimeoutMs, wallTimeoutExplicit, quick, ...paths };
}

function formatShardSummary(shards: string[][]): string[] {
  return shards.map((files, index) => {
    const preview = files.slice(0, 3).join(', ');
    const suffix = files.length > 3 ? ', ...' : '';
    return `Shard ${index + 1}/${shards.length}: ${files.length} files${preview ? ` -> ${preview}${suffix}` : ''}`;
  });
}

// ---------------------------------------------------------------------------
// Output contract: console filtering + per-file failure attribution.
//
// Bun groups each file's output under a `path/to/file.test.ts:` header line
// (cwd-relative, sometimes ../-prefixed through a symlinked cwd). The
// reporter tracks the current header while consuming the stream, attributes
// `(fail)` lines and crash markers to files, and decides which lines reach
// the console in the default quiet mode. All matching happens on
// ANSI-stripped lines — colored `(fail)` lines defeated a prior grep.
// ---------------------------------------------------------------------------

const TEST_PATH_SOURCE = String.raw`\.test\.(?:[cm]?[jt]s|tsx|jsx)`;
/** A file chunk header: the path bun printed, terminated by a bare colon. */
const FILE_HEADER_RE = new RegExp(`^(\\S.*${TEST_PATH_SOURCE}):$`);
/** bun --parallel retries a crashed worker once: `<icon> crashed running <path>, retrying`. */
const CRASH_RETRY_RE = new RegExp(`crashed running (\\S*${TEST_PATH_SOURCE}), retrying`);
/** The give-up marker after the retry also crashes: `✗ <path> (crashed: exited)`. */
const CRASH_FINAL_RE = new RegExp(`(\\S*${TEST_PATH_SOURCE}) \\(crashed: [^)]+\\)`);
const TERMINAL_SUMMARY_CAPTURE_RE = /^Ran (\d+) tests? across (\d+) files?\. \[/;
/** Substrings that must reach the console even in the default quiet mode. */
const CONSOLE_ALWAYS_MARKERS = ['error:', 'panic:', 'Unhandled error', 'crashed'] as const;

export type StreamOrigin = 'stdout' | 'stderr';

export interface FreeRunFailure {
  /** Planned relative path when attributable, else the raw header path, else null. */
  file: string | null;
  testName: string;
}

export interface FreeRunReport {
  testsRan: number | null;
  filesRan: number | null;
  sawTerminalSummary: boolean;
  /** Deduped `(fail)` lines in arrival order, attributed to the current file header. */
  failures: FreeRunFailure[];
  failedTests: number;
  unreportedFailures: number;
  /** Files that crashed a worker (bun retries once; a second crash is final). Deduped. */
  crashedFiles: string[];
  /**
   * "# Unhandled error between tests" markers, attributed to the chunk they
   * appeared in. These fail the shard via the strict classifier but produce
   * NO (fail) lines — without surfacing them here, the epilogue reads
   * "FAIL — 0 failing test(s)" and the culprit is undiscoverable from CI
   * output (first Windows lane run: a module-load throw in skill-census).
   */
  unhandledErrors: Array<{ file: string | null }>;
  /**
   * Wedge-suspect heuristic for a wall-timeout kill: files whose header was
   * seen but whose chunk never ENDED (chunk end = the next file's header, or
   * a final crash marker) before the terminal summary — i.e. "started but
   * never produced a result chunk end". Result lines deliberately do NOT end
   * a chunk: a file that printed a fail and then wedged stays listed. Known
   * limits of the approximation:
   *   - Serial (--shard CI path): bun streams live but prints a file's header
   *     lazily, on its first output line — a wedged file that printed ANY
   *     line is listed; a fully silent wedge is not.
   *   - Parallel (full-suite path): bun buffers a file's whole chunk until it
   *     COMPLETES, so a wedged file usually never prints a header (see
   *     filesWithNoOutput), and the LAST flushed chunk before the kill has no
   *     closing header, so one completed noisy file can be over-listed.
   */
  inFlight: string[];
  /** Planned files never observed in the stream (silent passers + never-flushed wedges). */
  filesWithNoOutput: number;
}

interface FileProgress {
  headerSeen: boolean;
  /** The file's chunk ended: a later file's header arrived, or it crashed out. */
  ended: boolean;
}

/**
 * Incrementally consumes the child's stdout/stderr (chunk boundaries need not
 * align to lines), attributing results to files and forwarding only
 * always-visible lines to `forward` (omit `forward` for verbose/quiet modes —
 * attribution still runs so the epilogue works in every mode).
 */
export class FreeRunReporter {
  private readonly decoders: Record<StreamOrigin, StringDecoder> = {
    stdout: new StringDecoder('utf8'),
    stderr: new StringDecoder('utf8'),
  };
  private readonly pending: Record<StreamOrigin, string> = { stdout: '', stderr: '' };
  private readonly plannedSet: Set<string>;
  private readonly canonicalCache = new Map<string, string>();
  private readonly progress = new Map<string, FileProgress>();
  private readonly failureKeys = new Set<string>();
  private readonly failures: FreeRunFailure[] = [];
  private namedFailureCount = 0;
  private reportedFailedTests = 0;
  private readonly failureSummary = new BunFailureSummaryParser();
  private readonly crashed = new Set<string>();
  private currentFile: string | null = null;
  private inRecap = false;
  private readonly unhandled: Array<{ file: string | null }> = [];
  private testsRan: number | null = null;
  private filesRan: number | null = null;
  private sawSummary = false;

  constructor(
    private readonly plannedFiles: string[],
    private readonly forward?: (text: string, origin: StreamOrigin) => void,
  ) {
    this.plannedSet = new Set(plannedFiles.map(normalizeRelativePath));
  }

  write(chunk: Uint8Array | string, origin: StreamOrigin): void {
    this.pending[origin] += typeof chunk === 'string'
      ? chunk
      : this.decoders[origin].write(Buffer.from(chunk));
    let newline = this.pending[origin].indexOf('\n');
    while (newline !== -1) {
      this.handleLine(this.pending[origin].slice(0, newline), origin);
      this.pending[origin] = this.pending[origin].slice(newline + 1);
      newline = this.pending[origin].indexOf('\n');
    }
  }

  /** Flush partial trailing lines (a stream killed mid-line still classifies). */
  end(): void {
    for (const origin of ['stdout', 'stderr'] as const) {
      this.pending[origin] += this.decoders[origin].end();
      if (this.pending[origin].length > 0) this.handleLine(this.pending[origin], origin);
      this.pending[origin] = '';
    }
  }

  report(): FreeRunReport {
    const inFlight = this.sawSummary
      ? []
      : [...this.progress.entries()]
          .filter(([, p]) => p.headerSeen && !p.ended)
          .map(([file]) => file)
          .sort();
    return {
      testsRan: this.testsRan,
      filesRan: this.filesRan,
      sawTerminalSummary: this.sawSummary,
      failures: [...this.failures],
      failedTests: Math.max(this.reportedFailedTests, this.failures.length),
      unreportedFailures: Math.max(0, this.reportedFailedTests - this.namedFailureCount),
      crashedFiles: [...this.crashed].sort(),
      unhandledErrors: [...this.unhandled],
      inFlight,
      filesWithNoOutput: this.plannedFiles.filter((f) => !this.progress.has(normalizeRelativePath(f))).length,
    };
  }

  private handleLine(rawLine: string, origin: StreamOrigin): void {
    // GitHub Actions: bun wraps each file's section in ::group::<header>.
    // Without stripping, the real header fails FILE_HEADER_RE, failures get
    // attributed to the PREVIOUS file, and the terminal recap's re-printed
    // (fail) lines land under a second phantom file (observed on the first
    // Linux run: 5 real failures reported as 10 across 2 files).
    const line = stripAnsiLine(rawLine).replace(/^::group::/, '');
    let visible = false;
    const failedCount = this.failureSummary.consume(line, origin);
    if (failedCount !== null) {
      this.reportedFailedTests = Math.max(this.reportedFailedTests, failedCount);
      visible = failedCount > 0;
    }

    // Bun's terminal recap ("N tests failed:") re-prints every (fail) line
    // WITHOUT re-printing file headers. Attributing those to the stale
    // currentFile invented a phantom failing file on the first Linux run
    // (5 real failures reported as 10 across 2 files, one innocent).
    if (/^\d+ tests? failed:$/.test(line)) {
      this.inRecap = true;
      if (this.currentFile) this.progressFor(this.currentFile).ended = true;
      this.currentFile = null;
    }

    if (line === '# Unhandled error between tests') {
      this.unhandled.push({ file: this.currentFile });
    }

    const header = FILE_HEADER_RE.exec(line);
    if (header) {
      const file = this.canonicalize(header[1]);
      // A new header ends the previous file's chunk — that file is no longer
      // a wedge suspect. (Bun 1.3.x prints NO (pass) lines, so chunk
      // delimiters, not result lines, are the completion signal.)
      if (this.currentFile && this.currentFile !== file) this.progressFor(this.currentFile).ended = true;
      this.currentFile = file;
      this.progressFor(file).headerSeen = true;
    } else {
      const fail = parseBunFailureResult(line);
      const retry = fail ? null : CRASH_RETRY_RE.exec(line);
      const final = fail || retry ? null : CRASH_FINAL_RE.exec(line);
      if (fail) {
        visible = true;
        // In the recap, a (fail) line only records a failure the main run
        // somehow never attributed (belt and braces); known names dedupe.
        const recapDuplicate = this.inRecap
          && this.failures.some((f) => f.testName === fail);
        if (!recapDuplicate) this.namedFailureCount += 1;
        const key = `${this.currentFile ?? ''}\u0000${fail}`;
        if (!recapDuplicate && !this.failureKeys.has(key)) {
          this.failureKeys.add(key);
          this.failures.push({ file: this.currentFile, testName: fail });
        }
      } else if (retry) {
        // The file will run again — a crash+retry does not end its chunk.
        visible = true;
        this.crashed.add(this.canonicalize(retry[1]));
      } else if (final) {
        visible = true;
        const file = this.canonicalize(final[1]);
        this.crashed.add(file);
        this.progressFor(file).ended = true;
      } else {
        const summary = TERMINAL_SUMMARY_CAPTURE_RE.exec(line);
        if (summary) {
          visible = true;
          this.sawSummary = true;
          this.testsRan = Number.parseInt(summary[1], 10);
          this.filesRan = Number.parseInt(summary[2], 10);
        }
      }
    }

    if (!visible) visible = CONSOLE_ALWAYS_MARKERS.some((marker) => line.includes(marker));
    if (visible && this.forward) this.forward(`${rawLine.replace(/\r$/, '')}\n`, origin);
  }

  private progressFor(file: string): FileProgress {
    let entry = this.progress.get(file);
    if (!entry) {
      entry = { headerSeen: false, ended: false };
      this.progress.set(file, entry);
    }
    return entry;
  }

  /**
   * Map a printed path back to its planned relative path. Bun prints paths
   * relative to the child's (real)cwd, so a symlinked cwd (macOS /tmp) yields
   * `../..`-prefixed forms — strip the prefix and suffix-match.
   */
  private canonicalize(printedPath: string): string {
    const cached = this.canonicalCache.get(printedPath);
    if (cached) return cached;
    const stripped = normalizeRelativePath(printedPath).replace(/^(?:\.{1,2}\/)+/, '');
    let resolved = stripped;
    if (!this.plannedSet.has(stripped)) {
      const match = this.plannedFiles.find(
        (planned) => stripped.endsWith(`/${planned}`) || planned.endsWith(`/${stripped}`),
      );
      if (match) resolved = match;
    }
    this.canonicalCache.set(printedPath, resolved);
    return resolved;
  }
}

/**
 * The stable post-run epilogue. Success is one line; failure names every
 * failing test (deduped, attributed) and crashed worker; a wall-timeout kill
 * additionally prints the wedge-suspect list (see FreeRunReport.inFlight for
 * the heuristic and its limits).
 */
export function buildRunEpilogue(
  status: FreeShardStatus,
  report: FreeRunReport,
  elapsedMs: number,
  logPath: string,
): string[] {
  const seconds = Math.round(elapsedMs / 1000);
  if (status === 'passed') {
    return [
      `[test:free] PASS — ${report.testsRan ?? '?'} tests, ${report.filesRan ?? '?'} files, ${seconds}s. Full log: ${logPath}`,
    ];
  }
  const failingFiles = new Set(report.failures.map((f) => f.file ?? '(unattributed)'));
  const lines = [
    `[test:free] FAIL — ${report.failedTests} failing test(s) in ${failingFiles.size} ${report.unreportedFailures > 0 ? 'identified ' : ''}file(s), `
    + `${report.crashedFiles.length} crashed worker(s)${report.unhandledErrors.length > 0 ? `, ${report.unhandledErrors.length} unhandled error(s) between tests` : ''}. Full log: ${logPath}`,
  ];
  for (const failure of report.failures) {
    lines.push(`  ✗ ${failure.file ?? '(unattributed)'} — ${failure.testName}`);
  }
  if (report.unreportedFailures > 0) {
    lines.push(`  ⚠ ${report.unreportedFailures} failure(s) reported without named result lines`);
  }
  for (const file of report.crashedFiles) {
    lines.push(`  ⚠ crashed+retried: ${file}`);
  }
  for (const u of report.unhandledErrors) {
    lines.push(`  ⚠ unhandled error between tests (around ${u.file ?? 'unknown file'})`);
  }
  if (status === 'timed-out') {
    if (report.inFlight.length > 0) {
      lines.push(`  ⏱ in flight at kill: ${report.inFlight.join(', ')}`);
    } else {
      lines.push(
        '  ⏱ in flight at kill: unknown — no open file chunk was observed '
        + '(bun --parallel buffers a file\'s output until it completes, so a silent wedge never prints); '
        + `${report.filesWithNoOutput} planned file(s) produced no output before the kill.`,
      );
    }
  }
  return lines;
}

export type FreeShardStatus = 'passed' | 'failed' | 'timed-out';

// ─── Flake ledger (WS1 telemetry) ───────────────────────────────────────────
// Single-writer JSONL: ONLY this parent runner appends (never shards, never
// tests — no concurrent-append hazard by construction). CI points
// GSTACK_FLAKE_LEDGER at $RUNNER_TEMP and uploads it as an artifact every
// run, so repeat offenders become an enumerable series instead of console
// scrollback. Fail-open with a loud stderr warning: a broken ledger must
// never red the only required lane.

export interface FlakeLedgerEntry {
  ts: string;
  runner: 'free';
  kind: 'flaky-pass';
  file: string;
  /** Shard the original failure surfaced in, when attributable. */
  shard?: number;
  /** Code-state attribution (review finding): without branch/sha the series
   *  can't tie an entry to the state that produced it, and the WS16
   *  promotion evidence needs exactly that. */
  branch?: string;
  git_sha?: string;
}

export function flakeLedgerPath(env: NodeJS.ProcessEnv = process.env): string {
  if (env.GSTACK_FLAKE_LEDGER) return env.GSTACK_FLAKE_LEDGER;
  // Local default: per-PROJECT, not the machine-global tmpdir — sibling
  // Conductor worktrees of DIFFERENT repos must not interleave into one
  // series (review finding). CI always sets GSTACK_FLAKE_LEDGER explicitly.
  try {
    const slug = spawnSync('bash', ['-c', '~/.claude/skills/gstack/bin/gstack-slug 2>/dev/null'], { stdio: 'pipe', timeout: 3000 })
      .stdout?.toString().match(/^SLUG=(.+)$/m)?.[1];
    if (slug) {
      const dir = path.join(os.homedir(), '.gstack', 'projects', slug);
      fs.mkdirSync(dir, { recursive: true });
      return path.join(dir, 'flake-ledger.jsonl');
    }
  } catch { /* fall through */ }
  return path.join(os.tmpdir(), 'gstack-flake-ledger.jsonl');
}

export function appendFlakeLedger(
  entries: FlakeLedgerEntry[],
  ledgerPath: string,
  warn: (line: string) => void = (line) => console.error(line),
): boolean {
  if (entries.length === 0) return true;
  try {
    fs.mkdirSync(path.dirname(ledgerPath), { recursive: true });
    fs.appendFileSync(ledgerPath, entries.map((e) => JSON.stringify(e)).join('\n') + '\n');
    return true;
  } catch (error) {
    warn(`[test:free] WARNING: could not append flake ledger at ${ledgerPath} `
      + `(${error instanceof Error ? error.message : String(error)}) — flaky-pass telemetry lost for this run, verdict unaffected`);
    return false;
  }
}

export interface FreeShardOutcome {
  shard: number;
  files: string[];
  status: FreeShardStatus;
  exitCode: number | null;
  elapsedMs: number;
  groupPid: number | null;
  /** Required by CI receipts; optional for existing local caller fixtures. */
  summary?: Pick<FreeRunReport, 'testsRan' | 'filesRan' | 'sawTerminalSummary'>;
  /**
   * Repo-relative files with attributed test failures or crashes, deduped.
   * Feeds the opt-in flaky retry pass (GSTACK_FREE_RETRY_FLAKY) — empty on
   * pass, and empty when every failure was unattributed (retry would be
   * meaningless without knowing what to re-run).
   */
  failingFiles: string[];
  /**
   * Count of failure evidence the retry pass CANNOT re-run by file: fail
   * lines seen before any file-chunk header, unhandled errors between tests,
   * and a truncated run (no terminal summary). Nonzero vetoes the flaky
   * retry for the whole run — retrying only failingFiles would re-run a
   * subset and mask the rest as a FLAKY-PASS, re-opening the silent-truncation
   * hole the strict classifier exists to close.
   */
  unattributedFailures: number;
}

export interface ShardCommand {
  command: string;
  args: string[];
}

export interface RunFreeShardOptions {
  /** External wall-clock deadline; on expiry the child's process GROUP is SIGKILLed. */
  wallTimeoutMs?: number;
  rootDir?: string;
  env?: NodeJS.ProcessEnv;
  /** Pass bun's --parallel. No production caller today (see BuildShardArgsOptions.parallel). */
  parallel?: boolean;
  /** Override the spawned command. Tests inject fake pass/fail/slow commands. */
  commandFor?: (files: string[]) => ShardCommand;
  /** Suppress ALL child output from the console (tests). The classifier and the log file still see every byte. */
  quiet?: boolean;
  /** Forward the full child stream to the console (legacy firehose). Default: the quiet filtered console. */
  verbose?: boolean;
  /**
   * Console sink for child-stream output (tests inject to assert quiet vs
   * verbose behavior). Default: process.stdout / process.stderr by origin.
   * Runner-owned [test:free] lines go through `log`, not this sink.
   */
  consoleWrite?: (text: string) => void;
  /** Per-run full-stream log path (tests inject). Default: a private retained file under .context/free-test-logs. */
  logFilePath?: string;
  log?: (line: string) => void;
}

const EPILOGUE_WORD: Record<FreeShardStatus, string> = {
  passed: 'pass',
  failed: 'fail',
  'timed-out': 'timed-out',
};

function trackShardBrowser(stateDir: string, env: NodeJS.ProcessEnv) {
  class BrowserCleanupError extends Error {}
  class CaptureStopped extends Error {}
  type Identity = { pid: number; parent: number; start: string; daemon: number; root: boolean };
  type Capture = { abort: AbortController; deadline: number; probes: Set<Promise<unknown>>; records?: [string, string] };
  const identities = new Map<number, Identity>();
  const nativeStarts = new Map<string, string>();
  const interrupted = new Set<string>();
  const errors = new Set<string>();
  const stateFile = env.BROWSE_STATE_FILE!;
  let stopping = false;
  let ready = true;
  let closed = false;
  let forced = false;
  let cancellation = false;
  let deadline = Infinity;
  let forceAt = Infinity;
  let active: Capture | null = null;
  let pending: Promise<void> | null = null;
  let observed = false;
  let alive = true;

  const check = (capture: Capture) => {
    if (closed || capture.abort.signal.aborted || active !== capture) throw new CaptureStopped();
    if (performance.now() >= Math.min(capture.deadline, deadline)) throw new BrowserCleanupError('browser ownership deadline exceeded');
  };
  const probe = async (capture: Capture, command: string, args: string[], timeout: number) => {
    check(capture);
    const remaining = Math.min(timeout, capture.deadline - performance.now() - 100, deadline - performance.now() - 100);
    if (remaining <= 0) throw new BrowserCleanupError('browser ownership deadline exceeded');
    const task = new Promise<{ status: number | null; stdout: string }>((resolve, reject) => {
      const child = spawn(command, args, { detached: true, stdio: ['ignore', 'pipe', 'ignore'], windowsHide: true });
      let output = '';
      let failed = false;
      let done = false;
      let reaper: ReturnType<typeof setTimeout> | undefined;
      const finish = (status: number | null) => {
        if (done) return;
        done = true;
        clearTimeout(timer);
        clearTimeout(reaper);
        capture.abort.signal.removeEventListener('abort', stop);
        child.stdout?.destroy();
        child.unref();
        if (failed) reject(new BrowserCleanupError('browser identity probe did not complete'));
        else resolve({ status, stdout: output });
      };
      const stop = () => {
        if (done || failed) return;
        failed = true;
        killProcessGroup(child, 'SIGKILL');
        reaper = setTimeout(() => finish(null), 100);
      };
      const timer = setTimeout(stop, Math.max(1, remaining));
      capture.abort.signal.addEventListener('abort', stop, { once: true });
      child.once('error', () => { failed = true; finish(null); });
      child.once('close', finish);
      child.stdout?.on('data', chunk => {
        output += chunk.toString();
        if (output.length > 65536) stop();
      });
    });
    capture.probes.add(task);
    try {
      const result = await task;
      check(capture);
      return result;
    } finally { capture.probes.delete(task); }
  };

  const failure = (error: unknown) => {
    if (error instanceof CaptureStopped) return;
    const code = (error as NodeJS.ErrnoException)?.code;
    const file = (error as NodeJS.ErrnoException & { path?: string })?.path;
    const pid = typeof file === 'string' ? /^\/proc\/(\d+)\//.exec(file)?.[1] : undefined;
    if (pid && identities.has(Number(pid)) && ['EACCES', 'EPERM', 'ENOENT', 'ESRCH'].includes(code ?? '')) return;
    errors.add(error instanceof BrowserCleanupError ? error.message : 'browser ownership unavailable');
  };

  const inspectLinux = (pid: number): Omit<Identity, 'daemon' | 'root'> | null => {
    try {
      const raw = fs.readFileSync(`/proc/${pid}/stat`, 'utf8');
      const fields = raw.slice(raw.lastIndexOf(') ') + 2).trim().split(/\s+/);
      if (!/^\d+$/.test(fields[19] ?? '')) throw new BrowserCleanupError('process start identity unavailable');
      return fields[0] === 'Z' || fields[0] === 'X' ? null
        : { pid, parent: Number(fields[1]), start: fields[19] };
    } catch (error) {
      if ((error as NodeJS.ErrnoException).code === 'ENOENT' || (error as NodeJS.ErrnoException).code === 'ESRCH') return null;
      throw error;
    }
  };
  const inspect = async (capture: Capture, pid: number): Promise<Omit<Identity, 'daemon' | 'root'> | null> => {
    check(capture);
    if (!Number.isSafeInteger(pid) || pid <= 1) throw new BrowserCleanupError('invalid process identity');
    if (process.platform === 'linux') return inspectLinux(pid);
    const result = await probe(capture, 'ps', ['-p', String(pid), '-o', 'ppid=,stat=,lstart='], 500);
    if (result.status === 1 && !result.stdout.trim()) return null;
    if (result.status !== 0) throw new BrowserCleanupError('process identity unavailable');
    const fields = result.stdout.trim().split(/\s+/);
    return fields[1]?.startsWith('Z') ? null : { pid, parent: Number(fields[0]), start: fields.slice(2).join(' ') };
  };
  const required = [`BROWSE_STATE_FILE=${stateFile}`, `GSTACK_FREE_SHARD_ID=${env.GSTACK_FREE_SHARD_ID}`];
  const boundLinux = (pid: number) => {
    try {
      const values = fs.readFileSync(`/proc/${pid}/environ`, 'utf8').split('\0');
      return required.every(value => values.includes(value));
    } catch (error) {
      if ((error as NodeJS.ErrnoException).code === 'ENOENT' || (error as NodeJS.ErrnoException).code === 'ESRCH') return false;
      throw error;
    }
  };
  const bound = async (capture: Capture, pid: number): Promise<boolean> => {
    check(capture);
    if (process.platform === 'linux') return boundLinux(pid);
    const [command, environment] = await Promise.all([
      probe(capture, 'ps', ['-ww', '-p', String(pid), '-o', 'command='], 500),
      probe(capture, 'ps', ['eww', '-p', String(pid), '-o', 'command='], 500),
    ]);
    if (command.status !== 0 || environment.status !== 0) return false;
    const prefix = command.stdout.trim();
    if (!prefix || !environment.stdout.trim().startsWith(prefix + ' ')) return false;
    const values = ' ' + environment.stdout.trim().slice(prefix.length).trim() + ' ';
    return required.every(value => values.includes(' ' + value + ' '));
  };
  const live = async (capture: Capture, identity: Identity): Promise<boolean> => {
    const current = await inspect(capture, identity.pid);
    check(capture);
    if (!current) return false;
    if (!current.start || current.start !== identity.start) {
      errors.add('captured process identity was replaced');
      return false;
    }
    return true;
  };
  const record = (file: string): any => {
    if (!fs.existsSync(file)) return null;
    const info = fs.lstatSync(file);
    if (!info.isFile() || info.size > 65536) throw new BrowserCleanupError('unsafe browser state record');
    return JSON.parse(fs.readFileSync(file, 'utf8'));
  };
  const nativeStart = async (capture: Capture, identity: Identity): Promise<string> => {
    check(capture);
    const key = `${identity.pid}:${identity.start}`;
    const recorded = nativeStarts.get(key);
    if (recorded) return recorded;
    if (!await live(capture, identity)) throw new BrowserCleanupError('process exited before its native identity was captured');
    const result = await probe(capture, 'ps', ['-p', String(identity.pid), '-o', 'lstart='], 2000);
    const value = result.status === 0 ? result.stdout.trim().replace(/\s+/g, ' ') : '';
    if (!value || !await live(capture, identity)) throw new BrowserCleanupError('native process identity unavailable');
    check(capture);
    nativeStarts.set(key, value);
    return value;
  };
  const remember = (capture: Capture, identity: Identity) => {
    check(capture);
    const previous = identities.get(identity.pid);
    if (previous && previous.start !== identity.start) throw new BrowserCleanupError('captured process identity was replaced');
    identities.set(identity.pid, identity);
  };
  const descendants = async (capture: Capture, parent: Identity, visited: Set<number>): Promise<void> => {
    check(capture);
    if (visited.has(parent.pid)) return;
    visited.add(parent.pid);
    if (visited.size > 256) throw new BrowserCleanupError('owned browser process limit exceeded');
    if (!await live(capture, parent)) return;
    let children: number[];
    if (process.platform === 'linux') {
      try {
        children = fs.readFileSync(`/proc/${parent.pid}/task/${parent.pid}/children`, 'utf8').trim().split(/\s+/).filter(Boolean).map(Number);
      } catch (error) {
        if ((error as NodeJS.ErrnoException).code === 'ENOENT') return;
        throw error;
      }
    } else {
      const result = await probe(capture, 'pgrep', ['-P', String(parent.pid)], 500);
      if (result.status !== 0 && result.status !== 1) throw new BrowserCleanupError('owned child identities unavailable');
      children = result.stdout.trim().split(/\s+/).filter(Boolean).map(Number);
    }
    for (const pid of children) {
      const child = await inspect(capture, pid);
      if (!child || child.parent !== parent.pid || !await live(capture, parent)) continue;
      const identity = { ...child, daemon: parent.daemon, root: false };
      remember(capture, identity);
      await descendants(capture, identity, visited);
    }
  };
  const capture = async (operation: Capture) => {
    check(operation);
    ready = true;
    if (process.platform === 'win32') return;
    try {
      if (fs.realpathSync(stateDir) !== stateDir) throw new BrowserCleanupError('shard directory was replaced');
      const directory = path.dirname(stateFile);
      if (fs.existsSync(directory) && (!fs.lstatSync(directory).isDirectory()
        || fs.lstatSync(directory).isSymbolicLink())) throw new BrowserCleanupError('browser directory was replaced');
      const state = record(stateFile);
      if (state?.pid !== undefined) {
        const current = await inspect(operation, state.pid);
        if (current) {
          if (!await bound(operation, current.pid)) {
            if (!await inspect(operation, current.pid)) return;
            throw new BrowserCleanupError('daemon is not bound to this shard');
          }
          remember(operation, { ...current, daemon: current.pid, root: true });
        } else if (!identities.has(state.pid)) {
          throw new BrowserCleanupError('daemon exited before ownership was captured');
        }
      }
      for (const identity of identities.values()) if (identity.root) await descendants(operation, identity, new Set());
      const validateChild = async (pid: unknown, start: unknown, daemon: unknown) => {
        check(operation);
        if (!Number.isSafeInteger(pid) || (pid as number) <= 1) throw new BrowserCleanupError('invalid browser child identity');
        const identity = identities.get(pid as number);
        if (!identity || identity.daemon !== daemon || identity.root) throw new BrowserCleanupError('browser child ownership is unconfirmed');
        if (await live(operation, identity) && (typeof start !== 'string' || !start
          || await nativeStart(operation, identity) !== start.replace(/\s+/g, ' '))) throw new BrowserCleanupError('browser child identity was replaced');
      };
      const agent = record(path.join(directory, 'terminal-agent-pid'));
      const validations: Promise<unknown>[] = [];
      if (agent) {
        const daemon = identities.get(agent.ownerPid);
        if (!daemon?.root || (state?.pid !== undefined && agent.ownerPid !== state.pid)) throw new BrowserCleanupError('terminal owner is unconfirmed');
        validations.push(nativeStart(operation, daemon).then(start => {
          if (start !== agent.ownerStartTime?.replace(/\s+/g, ' ')) throw new BrowserCleanupError('terminal owner is unconfirmed');
        }));
        if (agent.pid === 0) ready = false;
        else validations.push(validateChild(agent.pid, agent.startTime, agent.ownerPid));
      }
      if (state?.chromiumPid !== undefined) validations.push(validateChild(state.chromiumPid, state.chromiumStartTime, state.pid));
      const results = await Promise.allSettled(validations);
      check(operation);
      for (const result of results) if (result.status === 'rejected') throw result.reason;
      operation.records = [JSON.stringify(state), JSON.stringify(agent)];
    } catch (error) {
      check(operation);
      ready = false;
      failure(error);
    }
  };
  const signalOwned = async (operation: Capture, force: boolean) => {
    for (const identity of identities.values()) {
      if (!force && (!identity.root || !ready || errors.size > 0)) continue;
      const key = `${identity.pid}:${identity.start}`;
      if (!force && interrupted.has(key)) continue;
      try {
        if (!await live(operation, identity)) continue;
        if (identity.root && !await bound(operation, identity.pid)) {
          if (!await live(operation, identity)) continue;
          throw new BrowserCleanupError('daemon environment changed before termination');
        }
        if (!await live(operation, identity)) continue;
        check(operation);
        if (!force && (!operation.records || JSON.stringify(record(stateFile)) !== operation.records[0]
          || JSON.stringify(record(path.join(path.dirname(stateFile), 'terminal-agent-pid'))) !== operation.records[1])) {
          ready = false;
          continue;
        }
        process.kill(identity.pid, force ? 'SIGKILL' : 'SIGINT');
        if (!force) interrupted.add(key);
      } catch (error) {
        check(operation);
        if ((error as NodeJS.ErrnoException).code !== 'ESRCH') failure(error);
      }
    }
  };
  const forceLinux = () => {
    if (process.platform !== 'linux') return;
    for (const identity of identities.values()) {
      try {
        const current = inspectLinux(identity.pid);
        if (!current) continue;
        if (current.start !== identity.start) throw new BrowserCleanupError('captured process identity was replaced');
        if (identity.root && !boundLinux(identity.pid)) {
          if (!inspectLinux(identity.pid)) continue;
          throw new BrowserCleanupError('daemon environment changed before termination');
        }
        if (inspectLinux(identity.pid)?.start === identity.start) process.kill(identity.pid, 'SIGKILL');
      } catch (error) {
        if ((error as NodeJS.ErrnoException).code !== 'ESRCH') failure(error);
      }
    }
  };
  const enqueue = () => {
    if (closed || pending || process.platform === 'win32') return;
    const operation: Capture = { abort: new AbortController(), deadline: Math.min(performance.now() + 10000, deadline, forced ? Infinity : forceAt), probes: new Set() };
    active = operation;
    pending = (async () => {
      try {
        if (!forced) await capture(operation);
        if (stopping) {
          await signalOwned(operation, forced);
          let stillAlive = false;
          for (const identity of identities.values()) if (await live(operation, identity)) stillAlive = true;
          check(operation);
          alive = stillAlive;
          observed = true;
        }
      } catch (error) {
        if (!closed && !operation.abort.signal.aborted) failure(error);
      } finally {
        operation.abort.abort();
        await Promise.allSettled([...operation.probes]);
        if (active === operation) active = null;
      }
    })().finally(() => { pending = null; });
  };
  const signal = (force: boolean) => {
    if (closed) return;
    if (!cancellation) {
      cancellation = true;
      stopping = true;
      deadline = Math.min(deadline, performance.now() + 5500);
      forceAt = Math.min(forceAt, performance.now() + 5000);
      active?.abort.abort();
    }
    if (force) {
      forced = true;
      active?.abort.abort();
      forceLinux();
    }
    if (pending) void pending.then(enqueue);
    else enqueue();
  };
  const timer = setInterval(enqueue, process.platform === 'darwin' ? 1000 : 250);
  timer.unref();
  return {
    signal,
    async settle(): Promise<string | null> {
      clearInterval(timer);
      if (process.platform === 'win32') { closed = true; return null; }
      stopping = true;
      deadline = Math.min(deadline, performance.now() + 10000);
      forceAt = Math.min(forceAt, performance.now() + 5000);
      active?.abort.abort();
      try {
        while (true) {
          if (performance.now() >= deadline) {
            errors.add('owned browser settlement deadline exceeded');
            break;
          }
          if (!forced && performance.now() >= forceAt) {
            forced = true;
            active?.abort.abort();
            forceLinux();
          }
          if (pending) await pending;
          enqueue();
          if (pending) await pending;
          if (observed && !alive) break;
          await new Promise(resolve => setTimeout(resolve, Math.min(50, Math.max(0, deadline - performance.now()))));
        }
      } catch {
        errors.add('owned browser settlement could not be verified');
      } finally {
        closed = true;
        clearInterval(timer);
        active?.abort.abort();
        if (pending) await pending;
      }
      return errors.size ? [...errors].join('; ') : null;
    },
  };
}

/** One line per shard, printed after the run: `[test:free] shard i/N: M files, XXs, pass|fail|timed-out`. */
function shardEpilogue(outcome: FreeShardOutcome, totalShards: number): string {
  return `[test:free] shard ${outcome.shard}/${totalShards}: ${outcome.files.length} files, `
    + `${Math.round(outcome.elapsedMs / 1000)}s, ${EPILOGUE_WORD[outcome.status]}`;
}

/**
 * Run one shard (or the whole suite, in --parallel full-suite mode) in its own
 * bun process and classify the result strictly.
 *
 * Verdict integrity: the child's exit code is never trusted alone. Output is
 * fed through BunTestOutputClassifier, and strictTestExitCode requires bun's
 * terminal summary to report EXACTLY the planned file count — a shard that
 * exits 0 without the summary (mid-suite process.exit truncation), with
 * `(fail)` result lines, or having run fewer files than planned is a FAILURE.
 * This is enforced for injected fake commands too (unlike the paid runner),
 * so tests can pin the summary-missing => failure backstop; fake passing
 * commands must print a synthetic `Ran N tests across M files. [Xms]` line.
 *
 * Per-shard temp isolation: each spawned child gets its own throwaway TMPDIR
 * (TEMP/TMP on Windows) so shards can't trip over each other's temp files.
 * Deliberately NOT GSTACK_HOME: injecting one shared scratch home for a whole
 * invocation made 6,900 tests share a MUTABLE state dir — config tests wrote
 * keys into it and relink/update-check tests then read them (measured: 12
 * cross-contamination failures on the first full run). Tests that need
 * GSTACK_HOME isolation mkdtemp their own per test — the repo convention —
 * and the hermetic-env machinery covers E2E children.
 */
export async function runFreeShard(
  files: string[],
  shardNumber: number,
  totalShards: number,
  options: RunFreeShardOptions = {},
): Promise<FreeShardOutcome> {
  const log = options.log ?? ((line: string) => console.log(line));
  const label = `[test:free] shard ${shardNumber}/${totalShards}`;

  // Empty shard = fast no-op SUCCESS. Indices are stable for the CI matrix,
  // so an unoccupied index must not fail or shift work to a different runner.
  if (files.length === 0) {
    const outcome: FreeShardOutcome = {
      shard: shardNumber, files: [], status: 'passed', exitCode: 0, elapsedMs: 0, groupPid: null, failingFiles: [], unattributedFailures: 0,
      summary: { testsRan: 0, filesRan: 0, sawTerminalSummary: false },
    };
    log(shardEpilogue(outcome, totalShards));
    return outcome;
  }

  const rootDir = options.rootDir ?? ROOT;
  const wallTimeoutMs = options.wallTimeoutMs ?? DEFAULT_WALL_TIMEOUT_MS;
  log(`${label} (${files.length} files${options.parallel ? ', bun --parallel' : ''})`);

  // Full-stream capture: EVERY child byte lands here, whatever the console
  // shows. Printed once at start so a wedged or noisy run is inspectable
  // without a re-run.
  const logPath = options.logFilePath ?? nextDefaultLogPath(rootDir);
  const logStream = fs.createWriteStream(logPath, { mode: 0o600 });
  let logWriteFailed = false;
  logStream.on('error', (err) => {
    if (logWriteFailed) return;
    logWriteFailed = true;
    console.error(`${label} could not write the full log at ${logPath}: ${err.message}`);
  });
  log(`[test:free] full log: ${logPath}`);

  const { command, args } = options.commandFor
    ? options.commandFor(files)
    : { command: process.execPath, args: buildShardArgs(files, { parallel: options.parallel, rootDir }) };

  const env = { ...(options.env ?? process.env) };
  const stateDir = fs.realpathSync(fs.mkdtempSync(path.join(os.tmpdir(), 'gstack-free-shard-')));
  const childTmp = path.join(stateDir, 'tmp');
  fs.mkdirSync(childTmp);
  env.TMPDIR = childTmp;
  env.TEMP = childTmp;
  env.TMP = childTmp;
  // CLI renders otherwise attach to the repo's shared .gstack/browse.json,
  // even with distinct Chromium profiles. Concurrent shards and surviving
  // daemons from prior runs can then replace or remove each other's state.
  // Override inherited state too; the shard owns this directory's cleanup.
  env.BROWSE_STATE_FILE = path.join(stateDir, '.gstack', 'browse.json');
  env.GSTACK_FREE_SHARD_ID = randomUUID();
  // Per-shard Chromium profile (same isolation idea as TMPDIR): nine test
  // files launch in-process persistent contexts or daemons that default to
  // the SHARED ~/.gstack/chromium-profile, and two concurrent shards on one
  // profile dir kill each other's browser — observed live on CI once
  // duration packing recomposed shards (handoff's launchPersistentContext
  // died "Target page, context or browser has been closed" while a sibling
  // shard's daemon logged "Chromium process crashed"). Hash sharding had
  // masked the collision by chance placement. Within a shard, files run
  // serially, so sharing the per-shard profile is safe; config tests that
  // assert resolution order save/restore this env around their assertions.
  env.CHROMIUM_PROFILE = path.join(stateDir, 'chromium-profile');

  const startedAt = Date.now();
  const child = spawn(command, args, {
    cwd: rootDir,
    env,
    stdio: ['ignore', 'pipe', 'pipe'],
    detached: process.platform !== 'win32',
    windowsHide: true,
  });
  const groupPid = child.pid ?? null;
  const browser = trackShardBrowser(stateDir, env);
  // Group-kill on parent SIGINT/SIGTERM too, not just on timeout.
  const forwarding = installChildSignalForwarding({
    kill: (signal?: NodeJS.Signals | number) => {
      killProcessGroup(child, (signal as NodeJS.Signals) ?? 'SIGTERM');
      browser.signal(signal === 'SIGKILL');
      return true;
    },
  });

  const classifier = new BunTestOutputClassifier();

  // Console policy: quiet => nothing; verbose => the raw firehose; default =>
  // only always-visible lines (fail results, crash markers, error/panic
  // markers, the terminal summary), selected by the reporter. The reporter
  // consumes the stream in EVERY mode so the epilogue can attribute failures.
  const emitToConsole = (text: string, origin: StreamOrigin): void => {
    if (options.quiet) return;
    if (options.consoleWrite) {
      options.consoleWrite(text);
      return;
    }
    (origin === 'stdout' ? process.stdout : process.stderr).write(text);
  };
  const reporter = new FreeRunReporter(files, options.verbose ? undefined : emitToConsole);

  const captureFailures = new Map<StreamOrigin, Error>();
  const consumeStream = (stream: typeof child.stdout, origin: StreamOrigin): Promise<void> =>
    new Promise((resolve) => {
      if (!stream) {
        captureFailures.set(origin, new Error('configured pipe is missing'));
        resolve();
        return;
      }
      let ended = stream.readableEnded;
      const incomplete = (error?: Error | null): void => {
        // A delayed error replaces the initial destroyed-stream diagnostic
        // with its original cause. Failures are data, never early rejections
        // while the caller is still waiting for the child's real exit.
        if (error) captureFailures.set(origin, error);
        else if (!captureFailures.has(origin)) captureFailures.set(origin, new Error('stream closed before end'));
        resolve();
      };
      // Even an already-destroyed pipe can emit error on the next tick.
      stream.on('error', incomplete);
      stream.once('end', () => { ended = true; resolve(); });
      stream.once('close', () => {
        if (!ended) incomplete(stream.errored);
        else resolve();
      });
      stream.on('data', (chunk: Buffer | string) => {
        classifier.write(chunk, origin); // strict verdict ALWAYS sees the full stream
        if (!logWriteFailed) logStream.write(chunk);
        reporter.write(chunk, origin);
        if (options.verbose) emitToConsole(typeof chunk === 'string' ? chunk : chunk.toString('utf8'), origin);
      });
      if (ended) resolve();
      else if (stream.destroyed) incomplete(stream.errored);
    });

  let timedOut = false;
  const killTimer = setTimeout(() => {
    timedOut = true;
    killProcessGroup(child, 'SIGKILL');
  }, wallTimeoutMs);

  let exitCode: number | null = null;
  let cleanupError: string | null = null;
  try {
    const streams = [consumeStream(child.stdout, 'stdout'), consumeStream(child.stderr, 'stderr')];
    exitCode = await new Promise<number | null>((resolve, reject) => {
      child.once('error', reject);
      child.once('close', (code) => resolve(code));
    });
    await Promise.all(streams);
  } finally {
    clearTimeout(killTimer);
    // Reap survivors of this shard even on the clean path.
    killProcessGroup(child, 'SIGKILL');
    cleanupError = await browser.settle();
    reporter.end();
    for (const [origin, error] of captureFailures) {
      const diagnostic = `${label} ${origin} capture incomplete: ${error.message} `
        + `(child exit ${exitCode ?? 'signal'}). Full log: ${logPath}`;
      console.error(diagnostic);
      if (!logWriteFailed) logStream.write(diagnostic + '\n');
    }
    await new Promise<void>((resolve) => logStream.end(() => resolve()));
    try {
      if (!cleanupError) fs.rmSync(stateDir, { recursive: true, force: true });
    } catch {
      // Best-effort cleanup of a throwaway temp dir — a locked file on
      // Windows must not turn a real verdict into an exception.
      if (process.platform !== 'win32') cleanupError = 'could not remove the owned shard directory';
    }
    forwarding.dispose();
  }

  const summary = classifier.end();
  const status: FreeShardStatus = timedOut
    ? 'timed-out'
    : !cleanupError && !logWriteFailed && captureFailures.size === 0 && strictTestExitCode(exitCode ?? 1, summary, files.length) === 0 ? 'passed' : 'failed';

  if (cleanupError) console.error(`${label} browser cleanup failed: ${cleanupError}; retained ${stateDir}`);

  if (status === 'timed-out') {
    console.error(
      `${label} exceeded the ${Math.round(wallTimeoutMs / 1000)}s wall-clock deadline — `
      + 'killed the process group. Reporting as TIMED-OUT (distinct from failed).',
    );
  } else if (status === 'failed' && !cleanupError && !logWriteFailed && captureFailures.size === 0 && (exitCode ?? 1) === 0) {
    const reason = summary.failedTests > 0 || summary.unhandledBetweenTests > 0
      ? `reported ${summary.failedTests} failing test(s) and ${summary.unhandledBetweenTests} unhandled error(s) between tests`
      : summary.terminalFileCounts.length === 0
        ? "never printed bun's terminal summary — the run was truncated (a process.exit fired mid-suite)"
        : `bun's summary reported ${summary.terminalFileCounts.join(', ')} file(s), expected ${files.length}`;
    console.error(`${label} exited 0 but ${reason}. Treating as FAILED.`);
  } else if (status === 'failed' && (exitCode ?? 1) !== 0) {
    console.error(`${label} failed with exit code ${exitCode ?? 'signal'}`);
  }

  const report = reporter.report();
  const failingFiles = status === 'passed' ? [] : [...new Set([
    ...report.failures.map((f) => f.file).filter((f): f is string => !!f),
    ...report.crashedFiles,
  ])];
  const unattributedFailures = status === 'passed' ? 0
    : report.failures.filter((f) => !f.file).length
      + report.unreportedFailures
      + report.unhandledErrors.length
      + captureFailures.size
      + (logWriteFailed ? 1 : 0)
      + (cleanupError ? 1 : 0)
      + (report.sawTerminalSummary ? 0 : 1);
  const outcome: FreeShardOutcome = {
    shard: shardNumber, files, status, exitCode, elapsedMs: Date.now() - startedAt, groupPid, failingFiles, unattributedFailures,
    summary: { testsRan: report.testsRan, filesRan: report.filesRan, sawTerminalSummary: report.sawTerminalSummary },
  };
  log(shardEpilogue(outcome, totalShards));
  for (const line of buildRunEpilogue(status, report, outcome.elapsedMs, logPath)) log(line);
  if (status !== 'passed') {
    const problem = cleanupError ? 'Owned-process cleanup is unconfirmed; inspect the retained state before another run.'
      : logWriteFailed ? 'The evidence log could not be retained; repair the log destination before another run.'
        : captureFailures.size || !report.sawTerminalSummary ? 'Evidence capture is incomplete; repair the stream or early exit before another run.'
          : status === 'timed-out' ? 'Execution exceeded its deadline; inspect the last completed step before changing code or rerunning.'
            : 'A test or module failed; the root cause is not established. Inspect the full log and repair the cause first.';
    log(`[test:free] Recovery: ${problem} See docs/TESTING_INTERNALS.md.`);
    const focused = failingFiles.filter(file => files.includes(file) && fs.existsSync(path.resolve(rootDir, file)));
    if (!unattributedFailures && focused.length) {
      log(`[test:free] After repair, focused check: bun test ${focused.map(file => `'${file.replaceAll("'", "'\\''")}'`).join(' ')}`);
    } else {
      log('[test:free] No complete narrower failure scope is available; do not treat a subset rerun as complete coverage.');
    }
  }
  return outcome;
}

let logPathSequence = 0;

/** Timestamped retained log file; pid+sequence defeat same-ms collisions. */
function nextDefaultLogPath(rootDir: string): string {
  let directory = fs.realpathSync(rootDir);
  for (const part of ['.context', 'free-test-logs']) {
    directory = path.join(directory, part);
    const existing = fs.lstatSync(directory, { throwIfNoEntry: false });
    if (existing && (!existing.isDirectory() || existing.isSymbolicLink())) throw new Error('Free-test log directory must not traverse links');
    if (!existing) fs.mkdirSync(directory, { mode: 0o700 });
  }
  fs.chmodSync(directory, 0o700);
  const stamp = new Date().toISOString().replace(/[:.]/g, '-');
  logPathSequence += 1;
  return path.join(directory, `gstack-free-test-${stamp}-${process.pid}-${logPathSequence}.log`);
}

function exitCodeFor(status: FreeShardStatus): number {
  if (status === 'passed') return 0;
  return status === 'timed-out' ? 124 : 1;
}

/**
 * `--record-durations`: time every file in its own child (exact per-file wall,
 * immune to bun's stream buffering) and write the committed seed atomically.
 * Uses the same isolated, strictly classified children as a normal run.
 * Run against an immutable checkout; never time while editing its inputs.
 */
async function recordFreeTestDurations(files: string[], jobs: number): Promise<number> {
  const durations: Record<string, number> = {};
  const failed: string[] = [];
  const expectedCount = files.length;
  let cursor = 0;
  console.log(`[test:free] recording per-file durations: ${files.length} files across ${jobs} workers`);
  const worker = async (): Promise<void> => {
    for (;;) {
      if (isTerminationRequested()) return;
      const index = cursor;
      cursor += 1;
      if (index >= files.length) return;
      const file = files[index];
      const outcome = await runFreeShard([file], index + 1, files.length, {
        wallTimeoutMs: wallTimeoutForShard(1),
        quiet: true,
      });
      durations[normalizeRelativePath(file)] = outcome.elapsedMs;
      if (outcome.status !== 'passed') failed.push(file);
    }
  };
  // As in full-suite mode, finish parallel work before exclusive host-state fixtures.
  const exclusive = files.filter(file => file in TREE_MUTATING);
  files = files.filter(file => !(file in TREE_MUTATING));
  await Promise.all(Array.from({ length: Math.max(1, jobs) }, () => worker()));
  files = exclusive;
  cursor = 0;
  await worker();
  if (Object.keys(durations).length !== expectedCount) {
    throw new Error('Duration recording was interrupted; the seed was not replaced.');
  }

  const target = process.env.GSTACK_FREE_TEST_DURATIONS ?? path.join(ROOT, FREE_TEST_DURATIONS_FILE);
  const payload = {
    version: 1,
    recordedAt: new Date().toISOString(),
    durations: Object.fromEntries(Object.entries(durations).sort(([a], [b]) => (a < b ? -1 : 1))),
  };
  // Atomic temp+rename (capture-context-budget's pattern): a killed recorder
  // must never leave a truncated seed for loadFreeTestDurations to warn on.
  const tmp = `${target}.tmp-${process.pid}`;
  fs.writeFileSync(tmp, `${JSON.stringify(payload, null, 2)}\n`);
  fs.renameSync(tmp, target);
  console.log(`[test:free] wrote ${Object.keys(durations).length} durations to ${path.relative(ROOT, target)}`);
  if (failed.length > 0) {
    // Failures still recorded (a red file's duration is still a real cost),
    // but surfaced loudly — recording from a broken tree deserves a look.
    console.error(`[test:free] WARNING: ${failed.length} file(s) failed while recording:`);
    for (const f of failed) console.error(`  ✗ ${f}`);
    return 1;
  }
  return 0;
}

async function retryFailedFreeFiles(
  outcomes: FreeShardOutcome[], totalShards: number,
  options: Pick<CliOptions, 'wallTimeoutExplicit' | 'wallTimeoutMs' | 'verbose'>,
): Promise<{ exitCode: number; retry: FreeShardOutcome | null }> {
  let worst = Math.max(...outcomes.map(outcome => exitCodeFor(outcome.status)));
  let retry: FreeShardOutcome | null = null;
  const shardTimeout = (count: number) => options.wallTimeoutExplicit
    ? options.wallTimeoutMs : wallTimeoutForShard(count, options.wallTimeoutMs);
  // Opt-in flaky retry (GSTACK_FREE_RETRY_FLAKY=1): when every failure is an
  // attributed test failure (no timeouts, no unattributed carnage), re-run
  // just the failing files ONCE in a fresh serial shard. A clean retry
  // downgrades the run to a loud flaky-pass; a repeat failure stays a
  // failure. Default OFF: dev boxes should see flakes, not absorb them.
  // Exists for syscall-supervised sandboxes (see fullSuiteJobs) where a run
  // lands 0-1 spurious browser-timing failures under an otherwise-green
  // suite. Capped so a genuinely broken tree never masquerades as flaky.
  const RETRY_CAP = 5;
  if (
    worst !== 0
    && process.env.GSTACK_FREE_RETRY_FLAKY === '1'
    && !isTerminationRequested()
    && outcomes.every((o) => o.status !== 'timed-out')
  ) {
    const flakyFiles = [...new Set(outcomes.flatMap((o) => o.failingFiles))]
      .filter((f): f is string => typeof f === 'string' && f.length > 0);
    // "Fully attributed" is per-failure, not per-shard: a shard with one
    // attributed failure PLUS a headerless failure / unhandled error /
    // truncated run must veto the retry — re-running only failingFiles would
    // mask the unattributable evidence as a FLAKY-PASS.
    const allAttributed = outcomes.every((o) => hasScopedFailureAttribution(o) && (o.status === 'passed'
      || (o.failingFiles.length > 0 && o.unattributedFailures === 0)));
    if (allAttributed && flakyFiles.length > 0 && flakyFiles.length <= RETRY_CAP) {
      console.log(`[test:free] flaky-retry: re-running ${flakyFiles.length} failing file(s) once, serially: ${flakyFiles.join(', ')}`);
      const retryOutcome = await runFreeShard(flakyFiles, totalShards + 1, totalShards + 1, {
        wallTimeoutMs: shardTimeout(flakyFiles.length),
        verbose: options.verbose,
      });
      retry = retryOutcome;
      if (retryOutcome.status === 'passed') {
        console.log(`[test:free] FLAKY-PASS — ${flakyFiles.length} file(s) failed once and passed on serial retry: ${flakyFiles.join(', ')}`);
        console.log('[test:free] treat repeat offenders as real flakes worth fixing, not noise.');
        // Durable record (WS1): console lines vanish with the scrollback; the
        // ledger makes repeat offenders rankable across runs (eval:flake-rank).
        const ts = new Date().toISOString();
        // Two separate calls: `rev-parse --abbrev-ref HEAD HEAD` abbreviates
        // BOTH revs, printing the branch twice — git_sha recorded the branch
        // name (codex adversarial finding).
        const ledgerBranch = (spawnSync('git', ['rev-parse', '--abbrev-ref', 'HEAD'], { cwd: ROOT, encoding: 'utf8', timeout: 5000 }).stdout ?? '').trim();
        const ledgerSha = (spawnSync('git', ['rev-parse', 'HEAD'], { cwd: ROOT, encoding: 'utf8', timeout: 5000 }).stdout ?? '').trim();
        appendFlakeLedger(
          flakyFiles.map((file) => ({
            ts,
            runner: 'free' as const,
            kind: 'flaky-pass' as const,
            file,
            shard: outcomes.find((o) => o.failingFiles.includes(file))?.shard,
            ...(ledgerBranch ? { branch: ledgerBranch } : {}),
            ...(ledgerSha ? { git_sha: ledgerSha.slice(0, 12) } : {}),
          })),
          flakeLedgerPath(),
        );
        worst = 0;
      } else {
        console.error('[test:free] flaky-retry FAILED — the failures reproduce serially; not flaky.');
      }
    } else {
      console.log(`[test:free] flaky-retry skipped: ${allAttributed ? `${flakyFiles.length} failing file(s) exceeds cap ${RETRY_CAP}` : 'failures not fully attributed'}.`);
    }
  }
  return { exitCode: worst, retry };
}

async function main(): Promise<number> {
  const options = parseCliOptions(process.argv.slice(2));
  const allFiles = collectFreeTestFiles();
  if (allFiles.length === 0) {
    throw new Error('No free test files were discovered.');
  }

  if (options.ciPlan || options.ciRun || options.ciVerify) {
    const git = spawnSync('git', ['rev-parse', 'HEAD'], { cwd: ROOT, encoding: 'utf8', timeout: 5_000 });
    if (git.status !== 0 || !git.stdout.trim()) throw new Error('Cannot bind CI plan to the checkout revision');
    const revision = git.stdout.trim();
    const writeJson = (file: string, value: unknown) => {
      fs.mkdirSync(path.dirname(path.resolve(file)), { recursive: true });
      const temporary = `${file}.tmp-${process.pid}`;
      fs.writeFileSync(temporary, JSON.stringify(value, null, 2) + '\n');
      fs.renameSync(temporary, file);
    };
    if (options.ciPlan) {
      const durations = loadFreeTestDurations() ?? {};
      warnUnseededFreeFiles(allFiles, durations);
      const plan = createFreeCiPlan(allFiles, options.shardCount, durations, revision);
      validateFreeCiPlan(plan, allFiles, revision);
      writeJson(options.ciPlan, plan);
      console.log(JSON.stringify({ shard: plan.shards.map(shard => shard.shard) }));
      return 0;
    }
    const plan = JSON.parse(fs.readFileSync((options.ciRun ?? options.ciVerify)!, 'utf8')) as FreeCiPlan;
    validateFreeCiPlan(plan, allFiles, revision);
    if (options.ciVerify) {
      const results = fs.readdirSync(options.results!).filter(file => file.endsWith('.json'))
        .map(file => JSON.parse(fs.readFileSync(path.join(options.results!, file), 'utf8')) as FreeCiResult);
      verifyFreeCiResults(plan, results);
      console.log(`[test:free] CI PASS: ${allFiles.length} files across ${results.length} isolated shards; slowest ${Math.round(Math.max(...results.map(result => result.outcome.elapsedMs + (result.retry?.elapsedMs ?? 0))) / 1000)}s including retries`);
      return 0;
    }
    const shard = plan.shards[options.shardIndex! - 1];
    if (!shard || shard.shard !== options.shardIndex) throw new Error('CI shard index is outside the plan');
    const outcome = await runFreeShard(shard.files, shard.shard, plan.shards.length, {
      wallTimeoutMs: options.wallTimeoutExplicit ? options.wallTimeoutMs
        : wallTimeoutForPackedShard(shard.predictedMs, options.wallTimeoutMs, shard.files.length),
      verbose: options.verbose,
    });
    const retried = await retryFailedFreeFiles([outcome], plan.shards.length, options);
    writeJson(options.result!, { planId: plan.id, revision, outcome, retry: retried.retry } satisfies FreeCiResult);
    return retried.exitCode;
  }

  let files = allFiles;
  if (options.quick) {
    const missing = QUICK_CORE.filter(file => !allFiles.includes(file));
    if (missing.length) throw new Error(`Quick core files missing: ${missing.join(', ')}`);
    const durations = loadFreeTestDurations() ?? {};
    files = selectQuickFreeFiles(allFiles, durations);
    const unknown = allFiles.filter(file => durations[file] === undefined && !QUICK_CORE.includes(file)).length;
    console.log(`[test:free] QUICK SUBSET: ${files.length}/${allFiles.length} files; ${unknown} unclassified and ${allFiles.length - files.length - unknown} slow files excluded. Full CI remains required; this is not release acceptance.`);
  }
  let curationReport: CurationResult | null = null;
  if (options.windowsOnly) {
    curationReport = curateWindowsSafe(allFiles);
    files = curationReport.safe;
    console.log(`[test:free] curated ${files.length} Windows-safe tests (${curationReport.excluded.length} excluded)`);
    if (options.listOnly && curationReport.excluded.length > 0) {
      console.log('\nExcluded (POSIX-fragile):');
      for (const { file, reason } of curationReport.excluded) {
        console.log(`  - ${file}  [${reason}]`);
      }
    }
  }

  if (options.listOnly) {
    console.log(`\nDiscovered ${files.length} test files.`);
    for (const file of files) console.log(`  ${file}`);
    return 0;
  }

  if (options.recordDurations) {
    return recordFreeTestDurations(files, fullSuiteJobs());
  }

  if (options.dryRun) {
    const shards = assignFilesToShards(files, options.shardCount);
    const occupied = shards.filter((s) => s.length > 0).length;
    console.log(
      `\nWould run ${files.length} files across ${shards.length} shards (${occupied} occupied). `
      + 'Without --shard, the full suite runs as N concurrent shard processes '
      + '(plus an exclusive host-state shard) instead.',
    );
    for (const line of formatShardSummary(shards)) console.log(line);
    return 0;
  }

  if (options.shardIndex !== null) {
    // Bounds-check against the REQUESTED shard count, not post-assignment
    // occupancy — indices must be stable for a CI matrix, and an empty shard
    // is a valid fast no-op.
    if (!Number.isInteger(options.shardIndex) || options.shardIndex < 1 || options.shardIndex > options.shardCount) {
      throw new Error(`--shard must be between 1 and ${options.shardCount}. Received: ${options.shardIndex}`);
    }
    const shards = assignFilesToShards(files, options.shardCount);
    const outcome = await runFreeShard(shards[options.shardIndex - 1], options.shardIndex, options.shardCount, {
      wallTimeoutMs: options.wallTimeoutMs,
      verbose: options.verbose,
    });
    return exitCodeFor(outcome.status);
  }

  // Full-suite mode: N concurrent shard PROCESSES, serial within each — the
  // paid runner's proven model. One `bun test --parallel` invocation was
  // tried first (decision V3) and abandoned after three distinct
  // worker-runtime pathologies in a single day on Bun 1.3.13: a segfault
  // whose crashed-worker retry wedged the run (security-live-playwright), a
  // gated file's still-running file-level hooks stalling a worker
  // (compare-board), and spawn-heavy files hanging workers under load
  // (session-runner-timeout). Plain child processes have none of these:
  // proven spawn semantics, per-shard group-kill, per-shard logs, and a
  // wedge only ever costs its own shard. WORKER_HOSTILE files are moot in
  // process shards (no workers) and fold back into normal assignment.
  const jobs = fullSuiteJobs();
  // Phase split: exclusive host-state fixtures run AFTER the parallel shards,
  // so their shared process or filesystem state cannot interfere with readers.
  const exclusive = files.filter((f) => f in TREE_MUTATING);
  const readers = files.filter((f) => !(f in TREE_MUTATING));
  const durations = loadFreeTestDurations();
  if (durations) warnUnseededFreeFiles(files, durations);
  const packed = durations ? packShardsByDuration(readers, jobs, durations) : null;
  const shards = packed ? packed.shards : assignFilesToShards(readers, jobs);
  const totalShards = jobs + (exclusive.length > 0 ? 1 : 0);
  console.log(`[test:free] full suite: ${readers.length} files across ${jobs} shard processes`
    + (packed ? ' (duration-packed)' : '')
    + (exclusive.length > 0 ? `, then ${exclusive.length} exclusive host-state file(s) serially` : ''));
  if (packed) {
    // One line per shard so a packing regression is diagnosable from any log.
    packed.predictedMs.forEach((ms, i) => {
      console.log(`[test:free]   shard ${i + 1}: ${shards[i].length} files, predicted ~${Math.round(ms / 1000)}s`);
    });
  }
  const shardTimeout = (fileCount: number): number =>
    options.wallTimeoutExplicit ? options.wallTimeoutMs : wallTimeoutForShard(fileCount, options.wallTimeoutMs);
  const outcomes = await Promise.all(
    shards.map((shardFiles, index) => runFreeShard(shardFiles, index + 1, totalShards, {
      // Packed shards get duration-aware walls: LPT decouples file count from
      // cost BY DESIGN, so the 5s/file heuristic would undersize a shard
      // holding few expensive files.
      wallTimeoutMs: packed && !options.wallTimeoutExplicit
        ? wallTimeoutForPackedShard(packed.predictedMs[index], options.wallTimeoutMs, shardFiles.length)
        : shardTimeout(shardFiles.length),
      verbose: options.verbose,
    })),
  );
  let worst = Math.max(...outcomes.map((o) => exitCodeFor(o.status)));
  // Cancellation stops the run: don't launch the exclusive host-state shard
  // after a SIGINT/SIGTERM already killed the parallel phase.
  if (exclusive.length > 0 && !isTerminationRequested()) {
    const exclusiveOutcome = await runFreeShard(exclusive, totalShards, totalShards, {
      wallTimeoutMs: shardTimeout(exclusive.length),
      verbose: options.verbose,
    });
    worst = Math.max(worst, exitCodeFor(exclusiveOutcome.status));
    if (exclusiveOutcome.status !== 'passed') {
      // Fixture safety rests on each test restoring default state itself; a
      // SIGKILL at the wall deadline (or a mid-regeneration crash) defeats
      // that by construction. Say so, loudly, before someone commits
      // regenerated SKILL.md / .agents artifacts by accident.
      const dirty = spawnSyncGitStatusGenerated();
      if (dirty.length > 0) {
        console.error('[test:free] ⚠ exclusive host-state shard did not finish cleanly — generated artifacts are dirty:');
        for (const line of dirty.slice(0, 20)) console.error(`[test:free]   ${line}`);
        console.error('[test:free]   restore with: bun run gen:skill-docs (or git checkout -- <paths>)');
      }
    }
    outcomes.push(exclusiveOutcome);
  }

  return (await retryFailedFreeFiles(outcomes, totalShards, options)).exitCode;
}

/** Dirty generated artifacts (SKILL.md / host outputs) after a failed exclusive shard. */
function spawnSyncGitStatusGenerated(): string[] {
  const result = spawnSync('git', ['status', '--porcelain'], { cwd: ROOT, encoding: 'utf8' });
  if (result.status !== 0 || !result.stdout) return [];
  return result.stdout.split('\n').filter((line) =>
    /SKILL\.md$/.test(line) || line.includes('.agents/') || line.includes('.factory/'));
}

if (import.meta.main) {
  try {
    process.exitCode = await main();
  } catch (error) {
    console.error(`[test:free] ${error instanceof Error ? error.message : String(error)}`);
    process.exitCode = 1;
  }
}