@gotcos/glasses-server 7.0.18 → 7.0.29

package/server/lib/web-key-signer-tls.ts

@@ -0,0 +1,790 @@
+// A publicly trusted certificate for the phone signer origin, https://<MagicDNS name>:3144 (QA
+// 2026-11-09 blocker B2; plan PLAN_web_turn_authority Phase 0 and Decision 1). Exists ONLY with
+// COS_WEB_KEY_AUTHORITY on and the 3144 listener planned (createSignerTlsForBoot); off, nothing here
+// runs: no Tailscale call, no file, no change to any listener.
+//
+// WHY. 3144 served the mkcert pair from server/certs/, whose SAN has no ts.net name, so iPhone Safari
+// had no secure context (no passkey, no WebCrypto key) and COS Control's signer-file audit failed its
+// hostname check. Tailscale's HTTPS Certificates feature issues a Let's Encrypt certificate for this
+// Mac's own tailnet name. That is the phone origin Miles chose.
+//
+// WHAT IT DOES.
+//   - Obtains the pair with `tailscale cert --cert-file - --key-file - <name>`. Both to STDOUT, not
+//     to file paths: the macOS GUI build's CLI runs in an app sandbox ("this binary's filesystem
+//     writes go to $HOME/Library/Containers/<app>/Data", read from the 1.102.2 binary), so a path
+//     under the data dir would not receive the files. The server parses the PEM blocks and writes the
+//     pair itself, atomically, into <data dir>/web-key-tls/ (dir 0700, key 0600, cert 0644).
+//   - Asks `tailscale status --json --peers=false` first (read only, bounded; SR-03). Empty CertDomains means HTTPS
+//     Certificates is off for the tailnet: no cert call is made, and the status says so
+//     (`tailnet_https_certs_off`). A name missing from CertDomains: `cert_domain_mismatch`.
+//   - Serves the pair on 3144 by SNI: a handshake naming exactly the certificate's host gets it; any
+//     other name, no name or an IP keeps the mkcert pair (the server's default context). Replacing
+//     the pair swaps the context the next handshake reads: no restart, no listener change.
+//   - Checks at boot and every SIGNER_TLS_CHECK_INTERVAL_MS (daily); renews when fewer than
+//     SIGNER_TLS_RENEW_BEFORE_MS (30 days) remain. A valid pair with more left calls nothing.
+//   - Each check first re-reads the stored pair: a longer-lived, valid pair for this name written there
+//     by hand or by another process is served from then on, no restart and no CLI call.
+//   - The prompt pickup (7.0.20), the fallback if this server's own fetch fails: a pair written into
+//     the dir by hand (`tailscale cert` in Terminal) or by another process. While nothing trusted
+//     serves, every 3144 handshake and every status read checks the two stored files with `stat` and
+//     re-reads them only when they changed (inode, size, mtime, ctime). So such a pair is served at the
+//     next handshake or status read, with no restart, no CLI call and no write, instead of at the next
+//     check (up to SIGNER_TLS_RETRY_MAX_MS away while failing). While a trusted pair serves, nothing
+//     extra runs. What to write: CONTRACT_stored_pair.md (phone-key build, 2026-11-10).
+//   - Every CLI call gets a minimal explicit env with SHLVL=1 (signerTlsCliEnv, 7.0.20), never this
+//     process's env: without SHLVL the macOS CLI under the LaunchAgent answers "The Tailscale GUI failed
+//     to start" (exit 0) and no certificate is fetched. That answer is reported as
+//     `tailscale_cli_gui_unavailable`.
+//   - Every CLI call has a timeout. Consecutive failures back off SIGNER_TLS_RETRY_BASE_MS doubling
+//     to SIGNER_TLS_RETRY_MAX_MS; no call is made inside the cooldown (the breaker). One check at a
+//     time. Never throws.
+//   - Logs reason codes, exit facts and the expiry date. Never stdout (it carries the private key).
+//
+// NOT HERE: 3143 keeps its single mkcert pair (no SNI selection): Even Hub and every existing client
+// see exactly the certificate they saw before.
+
+import { execFile } from 'node:child_process'
+import { closeSync, constants as fsConstants, fstatSync, lstatSync, mkdirSync, chmodSync, openSync, readSync } from 'node:fs'
+import { join } from 'node:path'
+import { X509Certificate, createPrivateKey, createPublicKey } from 'node:crypto'
+import { createSecureContext, rootCertificates, type SecureContext } from 'node:tls'
+import type { ServerOptions as HttpsServerOptions } from 'node:https'
+import { durableAtomicWriteFileSync } from './atomic-fs.js'
+import { normalizeTailscaleDnsName } from './network-policy.js'
+import {
+  parseTailscaleStatus,
+  resolveTailscaleBin,
+  type RunResult,
+} from './tailscale-nudge.js'
+
+const DAY_MS = 24 * 60 * 60_000
+/** Renew when fewer than this remain before notAfter. */
+export const SIGNER_TLS_RENEW_BEFORE_MS = 30 * DAY_MS
+/** How often a healthy pair is re-checked (also the first check's cadence after a success). */
+export const SIGNER_TLS_CHECK_INTERVAL_MS = DAY_MS
+/** First retry after a failed check; doubles per consecutive failure. */
+export const SIGNER_TLS_RETRY_BASE_MS = 60_000
+/** The longest cooldown between failed checks. */
+export const SIGNER_TLS_RETRY_MAX_MS = 60 * 60_000
+/** SR-04: the least time between two re-checks started by status reads. */
+export const SIGNER_TLS_RECHECK_FLOOR_MS = 15_000
+/** SR-04: the failures a user fixes outside this server, which a status read may re-check at once. */
+const RECHECK_ON_READ: ReadonlySet<string> = new Set(['tailnet_https_certs_off', 'tailnet_name_unknown'])
+export const SIGNER_TLS_STATUS_TIMEOUT_MS = 5_000
+/** ACME issuance through tailscaled can take tens of seconds. */
+export const SIGNER_TLS_CERT_TIMEOUT_MS = 90_000
+export const SIGNER_TLS_DIR = 'web-key-tls'
+export const SIGNER_TLS_CERT_FILE = 'tailscale-cert.pem'
+export const SIGNER_TLS_KEY_FILE = 'tailscale-key.pem'
+export const SIGNER_TLS_KEY_MODE = 0o600
+export const SIGNER_TLS_CERT_MODE = 0o644
+/** SR-05: the breaker and the day's cert-call count, saved beside the pair (0600) so they survive a restart. */
+export const SIGNER_TLS_STATE_FILE = 'fetch-state.json'
+/** SR-05: the most `tailscale cert` calls per UTC day (Let's Encrypt allows 5 duplicate certificates a week). */
+export const SIGNER_TLS_CERT_CALLS_PER_DAY = 3
+/** SR-08: the most bytes read from either stored file (a real chain plus key is about 4 KiB). */
+export const SIGNER_TLS_STORED_MAX_BYTES = 64 * 1024
+/**
+ * SR-03: the signer's own status call, without peers. CertDomains, BackendState and Self are top-level and
+ * stay in the answer (checked 2026-11-11 on Tailscale 1.102 for macOS: 0 peers, CertDomains present), so a
+ * tailnet with hundreds of devices cannot push the answer past the 1 MiB output cap.
+ */
+export const SIGNER_TLS_STATUS_ARGS: readonly string[] = Object.freeze(['status', '--json', '--peers=false'])
+/** Node's error code when a child's output passes maxBuffer (SR-03): never "Tailscale is not installed". */
+const MAXBUFFER = 'ERR_CHILD_PROCESS_STDIO_MAXBUFFER'
+/** Both outputs to stdout ("-"), see the header. The name is a normalised *.ts.net name, never a flag. */
+export const SIGNER_TLS_CERT_ARGS = (name: string): string[] => ['cert', '--cert-file', '-', '--key-file', '-', name]
+
+/** The only variables of this process's env a Tailscale CLI call here may see (each only if set), plus SHLVL. */
+export const SIGNER_TLS_CLI_ENV_KEYS: readonly string[] = Object.freeze(['HOME', 'PATH', 'TMPDIR', 'USER', 'LANG'])
+
+/**
+ * The explicit, minimal env for this module's Tailscale CLI calls (status and cert): HOME, PATH, TMPDIR,
+ * USER and LANG when set, and SHLVL=1. Never this process's env: it carries the .env values (the API token,
+ * provider keys), which the Tailscale child has no need for.
+ *
+ * WHY SHLVL (probe 2026-11-10, operations/personal/wk46_2026/phone_key_build_20261110/
+ * probe_tailscale_launchd/RESULT.md): the macsys Tailscale binary is the GUI app's own executable, and it
+ * decides whether it runs as the CLI from SHLVL. A child of the LaunchAgent's node has none, so it tries to
+ * be the GUI and prints "The Tailscale GUI failed to start (Tailscale.CLIError error 3)" with exit 0 in
+ * 0.04 s; with SHLVL=1 (and only HOME and PATH besides) it answers the JSON in 0.03 s.
+ *
+ * Only this module's calls get it. tailscale-nudge has the same gap and is left unchanged: fixing it changes
+ * behaviour for everyone and needs Miles's go (side finding, COORD 2026-11-10 12:03).
+ */
+export function signerTlsCliEnv(base: NodeJS.ProcessEnv = process.env): NodeJS.ProcessEnv {
+  const env: NodeJS.ProcessEnv = {}
+  for (const key of SIGNER_TLS_CLI_ENV_KEYS) {
+    const value = base[key]
+    if (typeof value === 'string' && value !== '') env[key] = value
+  }
+  env.SHLVL = '1'
+  return env
+}
+
+/** A runner for this module's CLI calls: like tailscale-nudge's TailscaleRunner, plus the env the child gets. */
+export type SignerTlsRunner = (bin: string, args: readonly string[], timeoutMs: number, env: NodeJS.ProcessEnv) => Promise<RunResult>
+
+/**
+ * The default runner: a binary path and an argument array (never a shell), with `env` as the child's whole
+ * env. The result mapping is tailscale-nudge's execFileRunner's, copied so that file stays byte-identical.
+ */
+export const signerTlsCliRunner: SignerTlsRunner = (bin, args, timeoutMs, env) => new Promise(resolve => {
+  const started = Date.now()
+  execFile(bin, [...args], { timeout: timeoutMs, killSignal: 'SIGKILL', maxBuffer: 1024 * 1024, windowsHide: true, encoding: 'utf8', env },
+    (error, stdout, stderr) => {
+      const err = error as (NodeJS.ErrnoException & { code?: number | string; killed?: boolean; signal?: string }) | null
+      const spawnError = err && typeof err.code === 'string' ? err.code : undefined
+      resolve({
+        code: err ? (typeof err.code === 'number' ? err.code : null) : 0,
+        signal: err?.signal ?? null,
+        killed: !!err?.killed,
+        stdout: String(stdout ?? ''),
+        stderr: String(stderr ?? ''),
+        ...(spawnError ? { spawnError } : {}),
+        elapsedMs: Date.now() - started,
+      })
+    })
+})
+/** What the macOS CLI prints, with exit 0, when it cannot reach the GUI app (no SHLVL; see signerTlsCliEnv). */
+const CLI_GUI_UNAVAILABLE = /Tailscale GUI failed to start|Tailscale\.CLIError error 3\b/i
+
+/**
+ * SR-11: every reason the status can carry, in one list. The type is derived from it, and COS Control's
+ * golden fixture is generated from it (Control Tests/fixtures/web-key/gen_control_golden.mts), so Control's
+ * sentences cannot drift from the server. Adding or renaming one means regenerating that fixture.
+ */
+export const SIGNER_TLS_REASONS = Object.freeze([
+  'ok',
+  'pending',
+  'tailnet_name_unknown',
+  'tailnet_https_certs_off',
+  'cert_domain_mismatch',
+  'tailscale_not_running',
+  'tailscale_cli_missing',
+  'tailscale_cli_timeout',
+  'tailscale_cli_gui_unavailable',
+  'tailscale_output_too_large',
+  'tailscale_status_failed',
+  'tailscale_cert_failed',
+  'cert_invalid',
+  'cert_expired',
+  'cert_not_yet_valid',
+  'cert_untrusted',
+  'cert_store_failed',
+  'cert_daily_cap',
+  'stored_symlink',
+  'stored_not_file',
+  'stored_owner',
+  'stored_writable',
+  'stored_too_large',
+] as const)
+
+export type SignerTlsReason = typeof SIGNER_TLS_REASONS[number]
+
+export interface SignerTlsStatus {
+  /**
+   * 3144 presents a publicly trusted certificate for this Mac's tailnet name: it chains, link by link, to one
+   * of Node's bundled public roots (Mozilla's list, `tls.rootCertificates`), it is inside its validity window,
+   * and its key is the one served. No other pair is ever served by SNI (SR-01, review 2026-11-10).
+   */
+  trusted: boolean
+  /** `ok` when trusted; otherwise why not. */
+  reason: SignerTlsReason
+  /** The name the served certificate covers, or null. */
+  host: string | null
+  notAfter: string | null
+  /** The last check's failure, or null after a clean check (a renewal can fail while a pair still serves). */
+  lastError: SignerTlsReason | null
+  checkedAt: string | null
+  nextCheckAt: string | null
+  /**
+   * SR-08: why the last read of the stored pair was refused (a `stored_*` hygiene reason, or the pair's own
+   * check: `cert_untrusted`, `cert_invalid`, `cert_expired`, `cert_not_yet_valid`), or null when the stored
+   * pair was taken, is absent, or only lasts less long than the one serving.
+   */
+  storedRefused: SignerTlsReason | null
+}
+
+export interface SignerTlsPair {
+  host: string
+  certPem: string
+  keyPem: string
+  notAfterMs: number
+  context: SecureContext
+}
+
+export type Timer = { cancel: () => void }
+
+export interface SignerTlsDeps {
+  /** <data dir>/web-key-tls. Created only when a pair is first stored. */
+  dir: string
+  /** This Mac's tailnet name (index.ts: tailnetSelf.refresh()). Null: unknown. */
+  selfDnsName: () => Promise<string | null>
+  /** Default signerTlsCliRunner. Every call gets signerTlsCliEnv() as the child's env. */
+  runner?: SignerTlsRunner
+  resolveBin?: () => string | null
+  now?: () => number
+  log?: (line: string) => void
+  /** Test seam for the check timer. */
+  schedule?: (fn: () => void, ms: number) => Timer
+  /** SR-01 test seam: the public roots a pair must chain to. Default: Node's bundled roots (signerTlsPublicRoots). */
+  roots?: readonly X509Certificate[]
+  /** SR-08 test seam: the owner every stored file must have. Default: this process's uid. */
+  ownerUid?: number | null
+}
+
+const defaultSchedule = (fn: () => void, ms: number): Timer => {
+  const t = setTimeout(fn, ms)
+  t.unref?.()
+  return { cancel: () => clearTimeout(t) }
+}
+
+/** The cooldown after `consecutive` failed checks (1 = the first failure). */
+export function signerTlsRetryMs(consecutive: number): number {
+  const n = Math.max(1, Math.floor(consecutive))
+  return Math.min(SIGNER_TLS_RETRY_BASE_MS * 2 ** (n - 1), SIGNER_TLS_RETRY_MAX_MS)
+}
+
+/** Whether a pair with this notAfter is due for renewal at `now`. */
+export function signerTlsRenewalDue(notAfterMs: number, now: number): boolean {
+  return notAfterMs - now < SIGNER_TLS_RENEW_BEFORE_MS
+}
+
+const PEM_BLOCK = /-----BEGIN ([A-Z0-9 ]+)-----\r?\n[A-Za-z0-9+/=\r\n]+?-----END \1-----/g
+
+/** The most issuer links followed from the leaf to a bundled root. */
+export const SIGNER_TLS_MAX_CHAIN = 6
+
+/** Parse root certificates, skipping any that do not parse. */
+export function parseRootPems(pems: readonly string[]): X509Certificate[] {
+  const out: X509Certificate[] = []
+  for (const pem of pems) {
+    try { out.push(new X509Certificate(pem)) } catch { /* not a certificate: never a root */ }
+  }
+  return out
+}
+
+let bundledRoots: readonly X509Certificate[] | null = null
+/** SR-01: Node's bundled public roots (Mozilla's list, `tls.rootCertificates`), parsed once. */
+export function signerTlsPublicRoots(): readonly X509Certificate[] {
+  if (!bundledRoots) bundledRoots = Object.freeze(parseRootPems(rootCertificates))
+  return bundledRoots
+}
+
+function validAt(cert: X509Certificate, now: number): boolean {
+  const from = Date.parse(cert.validFrom)
+  const to = Date.parse(cert.validTo)
+  return Number.isFinite(from) && Number.isFinite(to) && from <= now && now < to
+}
+
+function issuedBy(child: X509Certificate, issuer: X509Certificate): boolean {
+  try { return child.checkIssued(issuer) && child.verify(issuer.publicKey) } catch { return false }
+}
+
+/**
+ * SR-01: whether `chain[0]` chains to one of `roots`, link by link. Each link is issued by the next
+ * (`checkIssued`: names, key identifiers, key usage) and its signature verifies with that issuer's key;
+ * every certificate on the path below the root is inside its validity window at `now`; every issuer below
+ * the root is a CA. The order the chain arrives in is not trusted: each step looks for the issuer among
+ * the certificates not yet used. A root's own validity is not checked (a trust anchor, as in browsers).
+ */
+export function chainsToRoot(chain: readonly X509Certificate[], roots: readonly X509Certificate[], now: number): boolean {
+  if (chain.length === 0) return false
+  let cur = chain[0]
+  const unused = chain.slice(1)
+  for (let step = 0; step <= SIGNER_TLS_MAX_CHAIN; step++) {
+    if (!validAt(cur, now)) return false
+    if (roots.some(root => issuedBy(cur, root))) return true
+    const next = unused.findIndex(c => c.ca && issuedBy(cur, c))
+    if (next < 0) return false
+    cur = unused.splice(next, 1)[0]
+  }
+  return false
+}
+
+export type SignerTlsPairRefusal = 'cert_invalid' | 'cert_expired' | 'cert_not_yet_valid' | 'cert_untrusted'
+export type SignerTlsPairCheck = { ok: true; pair: SignerTlsPair } | { ok: false; reason: SignerTlsPairRefusal }
+
+/**
+ * Validate a cert chain plus key for `name` (or, with name null, for the *.ts.net name the leaf covers),
+ * the same way for every source: the fetch, the stored pair at boot, at each check and in the pickup.
+ * Refused with a reason: no single key or no certificate, the leaf does not name the host, or the key is
+ * not the leaf's (`cert_invalid`); past notAfter (`cert_expired`); before notBefore (`cert_not_yet_valid`);
+ * not chaining to `roots`, Node's bundled public roots by default (`cert_untrusted`). Never throws.
+ */
+export function checkSignerTlsPair(text: string, name: string | null, now: number, roots: readonly X509Certificate[] = signerTlsPublicRoots()): SignerTlsPairCheck {
+  const refuse = (reason: SignerTlsPairRefusal): SignerTlsPairCheck => ({ ok: false, reason })
+  try {
+    const certs: string[] = []
+    const keys: string[] = []
+    for (const m of text.matchAll(PEM_BLOCK)) {
+      if (m[1] === 'CERTIFICATE') certs.push(m[0])
+      else if (/(^| )PRIVATE KEY$/.test(m[1])) keys.push(m[0])
+    }
+    if (certs.length === 0 || keys.length !== 1) return refuse('cert_invalid')
+    const leaf = new X509Certificate(certs[0])
+    const host = name ?? leafTailnetName(leaf)
+    if (!host || normalizeTailscaleDnsName(host) !== host) return refuse('cert_invalid')
+    if (leaf.checkHost(host, { subject: 'never' })?.toLowerCase() !== host) return refuse('cert_invalid')
+    const notAfterMs = Date.parse(leaf.validTo)
+    if (!Number.isFinite(notAfterMs) || notAfterMs <= now) return refuse('cert_expired')
+    // SR-01: never before the leaf's notBefore.
+    const notBeforeMs = Date.parse(leaf.validFrom)
+    if (!Number.isFinite(notBeforeMs) || notBeforeMs > now) return refuse('cert_not_yet_valid')
+    // SR-01: the key is the leaf's. createSecureContext alone does not refuse an EC certificate with an RSA
+    // key (probe 2026-11-11 on Node 24.1: no throw), so compare the public keys' SPKI bytes.
+    const spki = createPublicKey(createPrivateKey(keys[0])).export({ type: 'spki', format: 'der' })
+    if (!spki.equals(leaf.publicKey.export({ type: 'spki', format: 'der' }))) return refuse('cert_invalid')
+    // SR-01: publicly trusted: the leaf chains, link by link, to one of the roots.
+    if (!chainsToRoot(certs.map(c => new X509Certificate(c)), roots, now)) return refuse('cert_untrusted')
+    const certPem = `${certs.join('\n')}\n`
+    const keyPem = `${keys[0]}\n`
+    // Throws when the key does not belong to the leaf: a torn or swapped pair never serves.
+    const context = createSecureContext({ cert: certPem, key: keyPem })
+    return { ok: true, pair: { host, certPem, keyPem, notAfterMs, context } }
+  } catch {
+    return refuse('cert_invalid')
+  }
+}
+
+/** checkSignerTlsPair without the reason: the pair, or null. */
+export function parseSignerTlsPair(text: string, name: string | null, now: number, roots?: readonly X509Certificate[]): SignerTlsPair | null {
+  const checked = checkSignerTlsPair(text, name, now, roots)
+  return checked.ok ? checked.pair : null
+}
+
+function leafTailnetName(leaf: X509Certificate): string | null {
+  const san = leaf.subjectAltName ?? ''
+  for (const part of san.split(',')) {
+    const m = /^\s*DNS:(.+)$/.exec(part)
+    const n = m ? normalizeTailscaleDnsName(m[1]) : null
+    if (n) return n
+  }
+  return null
+}
+
+function classifyCertFailure(r: RunResult): SignerTlsReason {
+  if (r.spawnError) return r.spawnError === MAXBUFFER ? 'tailscale_output_too_large' : 'tailscale_cli_missing'
+  if (r.killed || r.signal) return 'tailscale_cli_timeout'
+  const text = `${r.stderr}\n${r.stdout.includes('PRIVATE KEY') ? '' : r.stdout}`
+  if (/does not support getting TLS certs|HTTPS cert(ificate)?s? (support )?(is |are )?not enabled|not enabled\/configured/i.test(text)) return 'tailnet_https_certs_off'
+  if (/invalid domain|must be one of/i.test(text)) return 'cert_domain_mismatch'
+  return 'tailscale_cert_failed'
+}
+
+function firstLine(s: string): string {
+  return (s.split('\n').find(l => l.trim()) ?? '').replace(/[^\x20-\x7e]/g, '').slice(0, 160)
+}
+
+type StoredRefusal = 'stored_symlink' | 'stored_not_file' | 'stored_owner' | 'stored_writable' | 'stored_too_large'
+type StoredRead = { kind: 'none' } | { kind: 'refused'; reason: StoredRefusal } | { kind: 'ok'; text: string }
+
+/** SR-08: the owner and mode rules shared by the folder and each file. */
+function ownerOrModeRefusal(uid: number, mode: number, ownerUid: number | null): StoredRefusal | null {
+  if (ownerUid !== null && uid !== ownerUid) return 'stored_owner'
+  if ((mode & 0o022) !== 0) return 'stored_writable'
+  return null
+}
+
+/**
+ * SR-08: the stored-pair folder: absent (nothing stored), refused, or fine. Not a symlink, a directory,
+ * owned by this server's user, and not writable by group or others.
+ */
+function storedDirRefusal(dir: string, ownerUid: number | null): 'none' | StoredRefusal | null {
+  const st = lstatSync(dir, { throwIfNoEntry: false })
+  if (!st) return 'none'
+  if (st.isSymbolicLink()) return 'stored_symlink'
+  if (!st.isDirectory()) return 'stored_not_file'
+  return ownerOrModeRefusal(st.uid, st.mode, ownerUid)
+}
+
+/**
+ * SR-08: read one stored file, or refuse it. Opened with O_NOFOLLOW and judged on the open descriptor
+ * (no swap between the check and the read): a regular file, owned by this server's user, not writable by
+ * group or others, at most SIGNER_TLS_STORED_MAX_BYTES (read with that cap, never the whole file).
+ */
+function readStoredFile(path: string, ownerUid: number | null): StoredRead {
+  const pre = lstatSync(path, { throwIfNoEntry: false })
+  if (!pre) return { kind: 'none' }
+  if (pre.isSymbolicLink()) return { kind: 'refused', reason: 'stored_symlink' }
+  let fd: number
+  try {
+    fd = openSync(path, fsConstants.O_RDONLY | fsConstants.O_NOFOLLOW | fsConstants.O_NONBLOCK)
+  } catch (error) {
+    const code = (error as NodeJS.ErrnoException).code
+    if (code === 'ENOENT') return { kind: 'none' }
+    if (code === 'ELOOP') return { kind: 'refused', reason: 'stored_symlink' }
+    throw error
+  }
+  try {
+    const st = fstatSync(fd)
+    if (!st.isFile()) return { kind: 'refused', reason: 'stored_not_file' }
+    const refusal = ownerOrModeRefusal(st.uid, st.mode, ownerUid)
+    if (refusal) return { kind: 'refused', reason: refusal }
+    if (st.size > SIGNER_TLS_STORED_MAX_BYTES) return { kind: 'refused', reason: 'stored_too_large' }
+    const buf = Buffer.alloc(SIGNER_TLS_STORED_MAX_BYTES + 1)
+    let total = 0
+    for (;;) {
+      const n = readSync(fd, buf, total, buf.length - total, total)
+      if (n <= 0) break
+      total += n
+      if (total > SIGNER_TLS_STORED_MAX_BYTES) return { kind: 'refused', reason: 'stored_too_large' }
+    }
+    return { kind: 'ok', text: buf.subarray(0, total).toString('utf8') }
+  } finally {
+    closeSync(fd)
+  }
+}
+
+/** SR-05: what survives a restart. */
+interface FetchState {
+  v: 1
+  /** The UTC day `certCalls` counts (YYYY-MM-DD). */
+  day: string
+  certCalls: number
+  lastCertAt: number | null
+  failures: number
+  cooldownUntil: number | null
+  lastError: SignerTlsReason | null
+}
+const utcDay = (ms: number): string => new Date(ms).toISOString().slice(0, 10)
+const finiteOrNull = (v: unknown): number | null => (typeof v === 'number' && Number.isFinite(v) ? v : null)
+
+export class SignerTls {
+  private readonly now: () => number
+  private readonly runner: SignerTlsRunner
+  private readonly resolveBin: () => string | null
+  private readonly log: (line: string) => void
+  private readonly schedule: (fn: () => void, ms: number) => Timer
+  private readonly roots: readonly X509Certificate[]
+  private readonly ownerUid: number | null
+  private current: SignerTlsPair | null = null
+  private storedRefused: SignerTlsReason | null = null
+  private name: string | null = null
+  private lastError: SignerTlsReason | null = null
+  private checkedAt: number | null = null
+  private nextCheckAt: number | null = null
+  private failures = 0
+  private cooldownUntil: number | null = null
+  private inflight: Promise<SignerTlsStatus> | null = null
+  private timer: Timer | null = null
+  private stopped = false
+  /** The stored files' `stat` signature the prompt pickup last read (null: never read). */
+  private pickupSeen: string | null = null
+  /** SR-04: when a status read last started a re-check, and whether the next check may skip the cooldown. */
+  private lastRecheckAt: number | null = null
+  private bypassCooldown = false
+  /** SR-05: the saved breaker and day's count. */
+  private fetchState: FetchState
+
+  constructor(private readonly deps: SignerTlsDeps) {
+    this.now = deps.now ?? Date.now
+    this.runner = deps.runner ?? signerTlsCliRunner
+    this.resolveBin = deps.resolveBin ?? (() => resolveTailscaleBin())
+    this.log = deps.log ?? (line => console.log(line))
+    this.schedule = deps.schedule ?? defaultSchedule
+    this.roots = deps.roots ?? signerTlsPublicRoots()
+    this.ownerUid = deps.ownerUid !== undefined ? deps.ownerUid : (typeof process.getuid === 'function' ? process.getuid() : null)
+    this.fetchState = { v: 1, day: utcDay(this.now()), certCalls: 0, lastCertAt: null, failures: 0, cooldownUntil: null, lastError: null }
+    this.loadFetchState()
+    this.loadStored()
+  }
+
+  /** For `https.createServer({ SNICallback })` on 3144. Unmatched: the server's default (mkcert) context. */
+  readonly sniCallback = (servername: string, cb: (err: Error | null, ctx?: SecureContext) => void): void => {
+    if (!this.sniMatches(this.current, servername)) this.pickUpStored()
+    // Read again after the pickup: a pair stored just now serves in this same handshake.
+    const pair = this.current
+    if (pair && this.sniMatches(pair, servername)) {
+      cb(null, pair.context)
+      return
+    }
+    cb(null, undefined)
+  }
+
+  private sniMatches(pair: SignerTlsPair | null, servername: string): boolean {
+    return pair !== null && typeof servername === 'string' && servername.toLowerCase() === pair.host && this.now() < pair.notAfterMs
+      // SR-09: and only for this Mac's current name once it is known (after a rename, never the old name's pair).
+      && (this.name === null || this.name === pair.host)
+  }
+
+  /** The pair being served, for tests and diagnostics. */
+  served(): SignerTlsPair | null { return this.current }
+
+  /** A trusted pair serves: unexpired, and for this Mac's tailnet name once that is known. */
+  private live(at: number): boolean {
+    const pair = this.current
+    return pair !== null && at < pair.notAfterMs && (this.name === null || this.name === pair.host)
+  }
+
+  /**
+   * The prompt pickup (see the header). Only while nothing trusted serves. Two `stat` calls; the files
+   * are read and validated only when their signature changed since the last pickup, so repeated
+   * handshakes or status reads cost no parsing. Scoped to this Mac's name like the check's re-read, and
+   * never a step back (loadStored). Never a CLI call, never a write, never throws.
+   */
+  private pickUpStored(): void {
+    try {
+      if (this.live(this.now())) return
+      const sig = this.storedSignature()
+      if (sig === null || sig === this.pickupSeen) return
+      this.pickupSeen = sig
+      const before = this.current
+      this.loadStored(this.name)
+      const pair = this.current
+      if (pair !== null && pair !== before) this.log(`[signer-tls] serving a stored certificate on 3144 until ${new Date(pair.notAfterMs).toISOString()}`)
+      else this.log(`[signer-tls] a stored certificate was not taken: ${this.storedRefused ?? 'not newer than the one serving'}`)
+    } catch { /* the next handshake, status read or check tries again */ }
+  }
+
+  /** inode, size, mtime and ctime of both stored files (lstat: a link's own change); null when either is missing (any other error throws to the caller's catch). */
+  private storedSignature(): string | null {
+    const parts: string[] = []
+    for (const file of [SIGNER_TLS_CERT_FILE, SIGNER_TLS_KEY_FILE]) {
+      const st = lstatSync(join(this.deps.dir, file), { bigint: true, throwIfNoEntry: false })
+      if (!st) return null
+      parts.push(`${st.dev}:${st.ino}:${st.size}:${st.mtimeNs}:${st.ctimeNs}`)
+    }
+    return parts.join('|')
+  }
+
+  /** Also the prompt pickup's second trigger: while nothing trusted serves, a status read looks at the stored files first. */
+  status(): SignerTlsStatus {
+    this.pickUpStored()
+    const pair = this.current
+    const at = this.now()
+    const live = this.live(at)
+    let reason: SignerTlsReason = 'ok'
+    if (!live) reason = this.lastError ?? (pair && at >= pair.notAfterMs ? 'cert_expired' : 'pending')
+    return {
+      trusted: live,
+      reason,
+      host: pair ? pair.host : null,
+      notAfter: pair ? new Date(pair.notAfterMs).toISOString() : null,
+      lastError: this.lastError,
+      checkedAt: this.checkedAt === null ? null : new Date(this.checkedAt).toISOString(),
+      nextCheckAt: this.nextCheckAt === null ? null : new Date(this.nextCheckAt).toISOString(),
+      storedRefused: this.storedRefused,
+    }
+  }
+
+  /**
+   * SR-04: what the authenticated status route calls (index.ts). When the last check failed only because
+   * HTTPS Certificates was off for the tailnet or the tailnet name was unknown, both fixed by the user outside
+   * this server, start one check now instead of waiting out the backoff (up to SIGNER_TLS_RETRY_MAX_MS): at
+   * most once per SIGNER_TLS_RECHECK_FLOOR_MS, never while a check runs, never after stop(). It does not wait
+   * for that check; the next read shows its result. Any other failure keeps the breaker as it is.
+   */
+  statusRead(): SignerTlsStatus {
+    try {
+      const at = this.now()
+      if (!this.stopped && this.inflight === null && this.lastError !== null && RECHECK_ON_READ.has(this.lastError)
+        && (this.lastRecheckAt === null || at - this.lastRecheckAt >= SIGNER_TLS_RECHECK_FLOOR_MS)) {
+        this.lastRecheckAt = at
+        this.bypassCooldown = true
+        void this.check()
+      }
+    } catch { /* the timer checks again */ }
+    return this.status()
+  }
+
+  /** The boot check, then one every interval (or after the cooldown while failing). */
+  start(): void {
+    this.stopped = false
+    void this.check()
+  }
+
+  stop(): void {
+    this.stopped = true
+    this.timer?.cancel()
+    this.timer = null
+  }
+
+  /** One check (single-flight). Never throws. */
+  check(): Promise<SignerTlsStatus> {
+    if (this.inflight) return this.inflight
+    this.inflight = this.runCheck()
+      .catch(() => { this.fail('tailscale_cert_failed') })
+      .then(() => {
+        try { this.reschedule() } catch { /* the next boot checks again */ }
+        return this.status()
+      })
+      .finally(() => { this.inflight = null })
+    return this.inflight
+  }
+
+  private reschedule(): void {
+    this.timer?.cancel()
+    this.timer = null
+    if (this.stopped) return
+    const at = this.now()
+    const next = this.failures > 0 && this.cooldownUntil !== null ? this.cooldownUntil : at + SIGNER_TLS_CHECK_INTERVAL_MS
+    this.nextCheckAt = next
+    this.timer = this.schedule(() => { void this.check() }, Math.max(0, next - at))
+  }
+
+  private succeed(storeFailed: boolean): void {
+    this.failures = 0
+    this.cooldownUntil = null
+    this.lastError = storeFailed ? 'cert_store_failed' : null
+    this.checkedAt = this.now()
+  }
+
+  private fail(reason: SignerTlsReason, facts = ''): void {
+    const at = this.now()
+    this.failures += 1
+    this.cooldownUntil = at + signerTlsRetryMs(this.failures)
+    this.lastError = reason
+    this.checkedAt = at
+    this.log(`[signer-tls] no trusted certificate: ${reason}${facts ? ` (${facts})` : ''}; next try in ${Math.round(signerTlsRetryMs(this.failures) / 1000)}s`)
+  }
+
+  private async runCheck(): Promise<void> {
+    // SR-04: a re-check a status read started may skip the cooldown, once.
+    const bypass = this.bypassCooldown
+    this.bypassCooldown = false
+    let name: string | null = null
+    try { name = normalizeTailscaleDnsName(await this.deps.selfDnsName()) } catch { name = null }
+    if (!name) { this.fail('tailnet_name_unknown'); return }
+    this.name = name
+    // A pair stored by another process (e.g. COS Control's helper, if this LaunchAgent's CLI cannot
+    // answer) is picked up here, validated, without a restart and without a CLI call.
+    this.loadStored(name)
+    const at = this.now()
+    const pair = this.current
+    if (pair && pair.host === name && !signerTlsRenewalDue(pair.notAfterMs, at)) { this.succeed(false); return }
+    // The breaker: no CLI call inside the cooldown after a failure.
+    if (this.cooldownUntil !== null && at < this.cooldownUntil && !bypass) return
+    const bin = this.resolveBin()
+    if (!bin) { this.fail('tailscale_cli_missing'); return }
+    // The minimal env with SHLVL=1 for both calls (signerTlsCliEnv): never this process's env.
+    const cliEnv = signerTlsCliEnv()
+
+    const st = await this.runner(bin, SIGNER_TLS_STATUS_ARGS, SIGNER_TLS_STATUS_TIMEOUT_MS, cliEnv)
+    if (st.spawnError) { this.fail(st.spawnError === MAXBUFFER ? 'tailscale_output_too_large' : 'tailscale_cli_missing', `spawn ${st.spawnError}`); return }
+    if (st.killed || st.signal) { this.fail('tailscale_cli_timeout', `status killed=${st.killed} signal=${st.signal ?? ''} ${st.elapsedMs}ms`); return }
+    const parsed = st.code === 0 ? parseTailscaleStatus(st.stdout, this.now()) : null
+    if (!parsed) {
+      // The CLI's "GUI failed to start" answer (exit 0, see signerTlsCliEnv) gets its own reason, so the status names it.
+      this.fail(CLI_GUI_UNAVAILABLE.test(`${st.stdout}\n${st.stderr}`) ? 'tailscale_cli_gui_unavailable' : 'tailscale_status_failed', `status code=${st.code} ${firstLine(st.stderr)}`)
+      return
+    }
+    if (parsed.backendState !== 'Running') { this.fail('tailscale_not_running', parsed.backendState); return }
+    const domains = parsed.certDomains ?? []
+    if (domains.length === 0) { this.fail('tailnet_https_certs_off'); return }
+    if (!domains.includes(name)) { this.fail('cert_domain_mismatch'); return }
+
+    // SR-05: at most SIGNER_TLS_CERT_CALLS_PER_DAY cert calls per UTC day. The call is counted and SAVED before
+    // it runs, so a crash during the call still counts; if the count cannot be saved, no call is made.
+    const today = utcDay(this.now())
+    if (this.fetchState.day !== today) this.fetchState = { ...this.fetchState, day: today, certCalls: 0 }
+    if (this.fetchState.certCalls >= SIGNER_TLS_CERT_CALLS_PER_DAY) { this.fail('cert_daily_cap', `${this.fetchState.certCalls} today`); return }
+    this.fetchState = { ...this.fetchState, certCalls: this.fetchState.certCalls + 1, lastCertAt: this.now() }
+    if (!this.saveFetchState()) { this.fail('cert_store_failed', 'fetch state not saved; no cert call'); return }
+
+    const r = await this.runner(bin, SIGNER_TLS_CERT_ARGS(name), SIGNER_TLS_CERT_TIMEOUT_MS, cliEnv)
+    if (r.spawnError || r.killed || r.signal || r.code !== 0) {
+      this.fail(classifyCertFailure(r), `cert code=${r.code} killed=${r.killed} signal=${r.signal ?? ''} ${r.elapsedMs}ms ${firstLine(r.stderr)}`)
+      this.recordCertOutcome()
+      return
+    }
+    // SR-01: the same checks as a stored pair, including the public-root chain; the refusal names why.
+    const checked = checkSignerTlsPair(r.stdout, name, this.now(), this.roots)
+    if (!checked.ok) { this.fail(checked.reason); this.recordCertOutcome(); return }
+    const fresh = checked.pair
+    let storeFailed = false
+    try { this.store(fresh) } catch { storeFailed = true }
+    // The hot swap: the next 3144 handshake reads this pair.
+    this.current = fresh
+    this.succeed(storeFailed)
+    this.recordCertOutcome()
+    this.log(`[signer-tls] serving a trusted certificate on 3144 until ${new Date(fresh.notAfterMs).toISOString()}${storeFailed ? ' (not saved: cert_store_failed)' : ''}`)
+  }
+
+  private store(pair: SignerTlsPair): void {
+    mkdirSync(this.deps.dir, { recursive: true, mode: 0o700 })
+    chmodSync(this.deps.dir, 0o700)
+    // Key first: a crash between the two writes leaves a pair that fails to load (key mismatch) and is fetched again.
+    durableAtomicWriteFileSync(join(this.deps.dir, SIGNER_TLS_KEY_FILE), pair.keyPem, { mode: SIGNER_TLS_KEY_MODE })
+    durableAtomicWriteFileSync(join(this.deps.dir, SIGNER_TLS_CERT_FILE), pair.certPem, { mode: SIGNER_TLS_CERT_MODE })
+  }
+
+  /**
+   * SR-05: restore the day's cert-call count and, while it has not run out, the breaker (failures, cooldown,
+   * last error), so a restart or a crash loop cannot call `tailscale cert` sooner than the backoff allows.
+   * Read with the same hygiene as the pair (SR-08). Missing, refused or malformed: nothing restored.
+   */
+  private loadFetchState(): void {
+    try {
+      if (storedDirRefusal(this.deps.dir, this.ownerUid) !== null) return
+      const read = readStoredFile(join(this.deps.dir, SIGNER_TLS_STATE_FILE), this.ownerUid)
+      if (read.kind !== 'ok') return
+      const raw = JSON.parse(read.text) as Record<string, unknown>
+      if (!raw || raw.v !== 1 || typeof raw.day !== 'string' || !/^\d{4}-\d{2}-\d{2}$/.test(raw.day)) return
+      const certCalls = typeof raw.certCalls === 'number' && Number.isInteger(raw.certCalls) && raw.certCalls >= 0 ? raw.certCalls : 0
+      const failures = typeof raw.failures === 'number' && Number.isInteger(raw.failures) && raw.failures >= 0 ? raw.failures : 0
+      const cooldownUntil = finiteOrNull(raw.cooldownUntil)
+      const lastError = typeof raw.lastError === 'string' && (SIGNER_TLS_REASONS as readonly string[]).includes(raw.lastError) ? raw.lastError as SignerTlsReason : null
+      this.fetchState = { v: 1, day: raw.day, certCalls, lastCertAt: finiteOrNull(raw.lastCertAt), failures, cooldownUntil, lastError }
+      if (failures > 0 && cooldownUntil !== null && cooldownUntil > this.now()) {
+        this.failures = failures
+        this.cooldownUntil = cooldownUntil
+        this.lastError = lastError
+      }
+    } catch { /* nothing restored */ }
+  }
+
+  /** SR-05: save the state (0600, atomic) in the pair's folder (0700). False when it cannot be written. */
+  private saveFetchState(): boolean {
+    try {
+      mkdirSync(this.deps.dir, { recursive: true, mode: 0o700 })
+      chmodSync(this.deps.dir, 0o700)
+      durableAtomicWriteFileSync(join(this.deps.dir, SIGNER_TLS_STATE_FILE), `${JSON.stringify(this.fetchState)}\n`, { mode: 0o600 })
+      return true
+    } catch {
+      return false
+    }
+  }
+
+  /** SR-05: after a cert call, save how it ended (the breaker as it now stands). */
+  private recordCertOutcome(): void {
+    this.fetchState = { ...this.fetchState, failures: this.failures, cooldownUntil: this.cooldownUntil, lastError: this.lastError }
+    this.saveFetchState()
+  }
+
+  /** With a name: only a pair for exactly that name. Without (at construction): the *.ts.net name the leaf covers. */
+  private loadStored(name: string | null = null): void {
+    try {
+      // SR-08: the folder, then each file, pass the hygiene checks before anything is read.
+      const dir = storedDirRefusal(this.deps.dir, this.ownerUid)
+      if (dir === 'none') { this.storedRefused = null; return }
+      if (dir !== null) { this.storedRefused = dir; return }
+      const cert = readStoredFile(join(this.deps.dir, SIGNER_TLS_CERT_FILE), this.ownerUid)
+      const key = readStoredFile(join(this.deps.dir, SIGNER_TLS_KEY_FILE), this.ownerUid)
+      if (cert.kind === 'none' || key.kind === 'none') { this.storedRefused = null; return }
+      if (cert.kind === 'refused') { this.storedRefused = cert.reason; return }
+      if (key.kind === 'refused') { this.storedRefused = key.reason; return }
+      const checked = checkSignerTlsPair(`${cert.text}\n${key.text}`, name, this.now(), this.roots)
+      if (!checked.ok) { this.storedRefused = checked.reason; return }
+      this.storedRefused = null
+      const pair = checked.pair
+      // Never a step back: only when nothing serves, the serving pair names another host, or this one lasts longer.
+      if (pair && (this.current === null || this.current.host !== pair.host || pair.notAfterMs > this.current.notAfterMs)) this.current = pair
+    } catch { /* nothing stored that serves: the check fetches */ }
+  }
+}
+
+/**
+ * index.ts's one door: a SignerTls only with COS_WEB_KEY_AUTHORITY on AND the 3144 listener planned.
+ * Otherwise null, and nothing is read, written or run.
+ */
+export function createSignerTlsForBoot(opts: { authorityEnabled: boolean; signerListener: boolean }, deps: SignerTlsDeps): SignerTls | null {
+  if (opts.authorityEnabled !== true || opts.signerListener !== true) return null
+  return new SignerTls(deps)
+}
+
+/** The 3144 server's TLS options: the mkcert pair as the default, plus SNI selection when a SignerTls exists. */
+export function signerHttpsOptions(base: { cert: Buffer | string; key: Buffer | string }, tls: SignerTls | null): HttpsServerOptions {
+  return tls ? { ...base, SNICallback: tls.sniCallback } : { ...base }
+}