@agfpd/iapeer 0.2.58 → 0.2.60

package/README.md

@@ -0,0 +1,11 @@
+# iapeer
+
+Ядро мультиагентной системы на одном хосте: реестр пиров, единый демон-роутер для обмена сообщениями между ними и команды управления — создание, запуск, усыпление, перезапуск, подключение к внешним каналам. Пиром может быть AI-агент (Claude, Codex), человек (через Telegram) или служебный процесс; все адресуются одинаково через один протокол.
+
+## Документация
+
+Документация проекта — в [`docs/`](docs/README.md): продуктовая часть (что это и как пользоваться) и механика изнутри ([`docs/internals/`](docs/internals/README.md)). Этот репозиторий — кодовая реализация.
+
+## Статус
+
+Версия — **0.2.58**. Платформа — macOS/launchd.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@agfpd/iapeer",
-  "version": "0.2.58",
+  "version": "0.2.60",
   "description": "Foundation core for the iapeer multi-agent ecosystem: identity, registry, storage, codec.",
   "type": "module",
   "bin": {

package/src/cli/index.ts

@@ -937,6 +937,8 @@ const USAGE = `usage: iapeer <verb> [args]
   native-memory <off|on> (--peer <p> | --all)                    gate/restore runtimes' native memory (canonized lever; контракт «Слот памяти»)
   trust-hooks <hooks.json> [--check]                             pre-seed codex hooks trust for a file-form hooks.json (no modal); --check = drift report
   shadow [--once | --minutes <m>] [--interval <ms>]              READ-ONLY tmux→pty fidelity observer (migration burn-in): compares pty-model verdicts vs live tmux on the warm fleet → ~/.iapeer/logs/iapeer/shadow-fidelity.json. Never delivers/mutates.
+  shadow-install                                                 install + load the always-on shadow burn-in launchd job (com.iapeer.shadow-fidelity); idempotent, re-cycled by update/rollback
+  shadow-uninstall                                               bootout + remove the shadow burn-in launchd job (the correct way to stop it; STOP file is respawned past by KeepAlive)
 `
 
 export async function runCli(argv: string[], env: NodeJS.ProcessEnv = process.env): Promise<number> {
@@ -1772,6 +1774,33 @@ export async function runCli(argv: string[], env: NodeJS.ProcessEnv = process.en
         out(`shadow-fidelity → ${cfg.eventLogDir}/shadow-fidelity.json\n`)
         return 0
       }
+      case 'shadow-install': {
+        // Code-managed install of the always-on shadow burn-in launchd job
+        // (com.iapeer.shadow-fidelity) — reproducible, ownership-sentinel-guarded, and
+        // re-cycled automatically by `iapeer update` / `rollback`. Replaces a hand-written
+        // plist. Idempotent (no-op on an unchanged, loaded job). install.ts carries NO @xterm
+        // — dynamic-imported anyway to keep the CLI's common path light.
+        const { installShadowJob } = await import('../shadow/install.ts')
+        const r = installShadowJob(env)
+        if (r.action === 'failed' || r.action === 'refused-foreign') {
+          errOut(`shadow-install: ${r.action}${r.detail ? ` — ${r.detail}` : ''}\n  plist: ${r.path}\n`)
+          return 1
+        }
+        out(`shadow burn-in job ${r.action} (plist ${r.changed ? 'written' : 'unchanged'}): ${r.path}\n`)
+        return 0
+      }
+      case 'shadow-uninstall': {
+        // Stop + remove the burn-in job (bootout + rm) — the CORRECT halt (KeepAlive would
+        // respawn past the STOP sentinel). Refuses a non-foundation plist (ownership guard).
+        const { uninstallShadowJob } = await import('../shadow/install.ts')
+        const r = uninstallShadowJob(env)
+        if (r.action === 'failed' || r.action === 'refused-foreign') {
+          errOut(`shadow-uninstall: ${r.action}${r.detail ? ` — ${r.detail}` : ''}\n  plist: ${r.path}\n`)
+          return 1
+        }
+        out(`shadow burn-in job ${r.action}: ${r.path}\n`)
+        return 0
+      }
       default: {
         // `iapeer <runtime>` (launch) — folder-launch the cwd's peer, ALWAYS fresh.
         if (verb && isRuntime(verb)) {

package/src/shadow/gate.test.ts

@@ -0,0 +1,143 @@
+import { describe, expect, test } from 'bun:test'
+import {
+  GateTracker,
+  classifyDirection,
+  classifySampleState,
+  type PeerSample,
+} from './gate.ts'
+
+// SYNTHETIC ACCEPTANCE (boris's Block-1 gate): prove the instrument classifies known-behind vs
+// known-ahead and records coverage CORRECTLY — the live proof that it measures right before we
+// trust any pty verdict for a flip. Pure: no tmux, no @xterm, booleans in, gate stats out.
+
+const mkSample = (peer: string, runtime: string, over: Partial<PeerSample> = {}): PeerSample => ({
+  peer,
+  runtime,
+  occupancy: { pty: false, tmux: false },
+  readyGate: { pty: false, tmux: false },
+  liveness: { pty: true, tmux: true },
+  bytesFed: 0,
+  hasHistory: false,
+  ...over,
+})
+
+describe('classifyDirection (tmux authoritative; true = active/ready/alive pole)', () => {
+  test('equal → agree', () => {
+    expect(classifyDirection(true, true)).toBe('agree')
+    expect(classifyDirection(false, false)).toBe('agree')
+  })
+  test('pty under-reports (false vs true) → pty-behind (the stale class)', () => {
+    expect(classifyDirection(false, true)).toBe('pty-behind')
+  })
+  test('pty over-reports (true vs false) → pty-ahead', () => {
+    expect(classifyDirection(true, false)).toBe('pty-ahead')
+  })
+})
+
+describe('classifySampleState (coverage regimes — idle must be distinguishable, false-green lesson)', () => {
+  test('tmux occupancy → composer-busy (regardless of bytes)', () => {
+    expect(classifySampleState(true, 0)).toBe('composer-busy')
+    expect(classifySampleState(true, 999)).toBe('composer-busy')
+  })
+  test('no occupancy + output churn → model-advancing (non-trivial regime)', () => {
+    expect(classifySampleState(false, 4096)).toBe('model-advancing')
+  })
+  test('no occupancy + no churn → idle', () => {
+    expect(classifySampleState(false, 0)).toBe('idle')
+  })
+})
+
+describe('GateTracker — direction is scored hard vs benign per the adjudicated criterion', () => {
+  test('pty-BEHIND is ALWAYS hard (zero tolerance) and closes the verdict gate', () => {
+    const g = new GateTracker()
+    g.sample(0, mkSample('p', 'claude', { occupancy: { pty: false, tmux: true } }))
+    const s = g.summary()
+    expect(s.verdicts.occupancy.behind).toBe(1)
+    expect(s.verdicts.occupancy.gateClean).toBe(false)
+    expect(s.hardViolations.some(v => v.kind === 'behind' && v.verdict === 'occupancy')).toBe(true)
+  })
+
+  test('boot-window pty-AHEAD that self-heals within budget → BENIGN (latency recorded)', () => {
+    const g = new GateTracker({ settleTicks: 3, convergeBudgetTicks: 3 })
+    // tick0: readyGate ahead (pty ready, tmux not) — peer just born → boot window
+    g.sample(0, mkSample('p', 'claude', { readyGate: { pty: true, tmux: false } }))
+    // tick1: readyGate agrees (both ready) → transient closes, latency 1, boot+within budget → benign
+    g.sample(1, mkSample('p', 'claude', { readyGate: { pty: true, tmux: true } }))
+    const rg = g.summary().verdicts.readyGate
+    expect(rg.aheadBenign).toBe(1)
+    expect(rg.aheadHard).toBe(0)
+    expect(rg.gateClean).toBe(true)
+    expect(rg.selfHeal.count).toBe(1)
+    expect(rg.selfHeal.max).toBe(1)
+  })
+
+  test('READY-GATE non-converging pty-AHEAD is HARD, never benign (refinement #3: false-ready = msg-loss)', () => {
+    const g = new GateTracker({ settleTicks: 3, convergeBudgetTicks: 3 })
+    // hold readyGate ahead (pty ready, tmux NEVER ready → stays in boot window, never converges)
+    for (let t = 0; t <= 5; t++) g.sample(t, mkSample('p', 'codex', { readyGate: { pty: true, tmux: false } }))
+    const rg = g.summary().verdicts.readyGate
+    expect(rg.aheadHard).toBeGreaterThanOrEqual(1)
+    expect(rg.aheadBenign).toBe(0)
+    expect(rg.gateClean).toBe(false)
+    expect(g.summary().hardViolations.some(v => v.kind === 'ahead-nonconverged' && v.verdict === 'readyGate')).toBe(true)
+  })
+
+  test('STEADY-STATE pty-AHEAD is HARD immediately (cannot be a boot self-heal)', () => {
+    const g = new GateTracker({ settleTicks: 2, convergeBudgetTicks: 3 })
+    // boot the peer: readyGate true for settleTicks
+    g.sample(0, mkSample('p', 'claude', { readyGate: { pty: true, tmux: true } }))
+    g.sample(1, mkSample('p', 'claude', { readyGate: { pty: true, tmux: true } })) // bootDone now
+    // steady-state occupancy ahead → hard 'ahead-steady'
+    g.sample(2, mkSample('p', 'claude', { readyGate: { pty: true, tmux: true }, occupancy: { pty: true, tmux: false } }))
+    const occ = g.summary().verdicts.occupancy
+    expect(occ.aheadHard).toBeGreaterThanOrEqual(1)
+    expect(g.summary().hardViolations.some(v => v.kind === 'ahead-steady' && v.verdict === 'occupancy')).toBe(true)
+  })
+
+  test('full agreement → clean gate, zero hard violations', () => {
+    const g = new GateTracker()
+    for (let t = 0; t < 5; t++) g.sample(t, mkSample('p', 'claude', { readyGate: { pty: true, tmux: true } }))
+    const s = g.summary()
+    expect(s.hardViolations).toHaveLength(0)
+    expect(s.verdicts.readyGate.gateClean).toBe(true)
+    expect(s.verdicts.occupancy.gateClean).toBe(true)
+    expect(s.verdicts.liveness.gateClean).toBe(true)
+  })
+})
+
+describe('GateTracker — coverage matrix (clean counts only over traversed cells)', () => {
+  test('births split fresh vs wake-resume by history-at-first-sight', () => {
+    const g = new GateTracker()
+    g.tickStart(1)
+    g.sample(0, mkSample('fresh-peer', 'claude', { hasHistory: false }))
+    g.tickStart(2)
+    g.sample(1, mkSample('woke-peer', 'codex', { hasHistory: true })) // history at first sight = wake-resume
+    const cov = g.summary().coverage
+    expect(cov.births['claude/fresh']).toBe(1)
+    expect(cov.births['codex/wake-resume']).toBe(1)
+  })
+
+  test('a peer that died and reappeared is a restart', () => {
+    const g = new GateTracker()
+    g.sample(0, mkSample('p', 'claude'))
+    g.peerGone('p')
+    g.sample(5, mkSample('p', 'claude'))
+    expect(g.summary().coverage.restarts['claude']).toBe(1)
+  })
+
+  test('state cells, fleet width, and history samples are recorded', () => {
+    const g = new GateTracker()
+    g.tickStart(3) // fleet width 3 this tick
+    g.sample(0, mkSample('p', 'claude', { occupancy: { pty: true, tmux: true } })) // composer-busy
+    g.sample(1, mkSample('p', 'claude', { bytesFed: 4096 })) // model-advancing
+    g.sample(2, mkSample('p', 'claude', { bytesFed: 0 })) // idle
+    g.sample(3, mkSample('h', 'codex', { hasHistory: true }))
+    const cov = g.summary().coverage
+    expect(cov.fleetWidthMax).toBe(3)
+    expect(cov.byRuntimeState['claude/composer-busy']).toBe(1)
+    expect(cov.byRuntimeState['claude/model-advancing']).toBe(1)
+    expect(cov.byRuntimeState['claude/idle']).toBe(1)
+    expect(cov.historySamples['codex']).toBe(1)
+    expect(cov.runtimes).toEqual(['claude', 'codex'])
+  })
+})

package/src/shadow/gate.ts

@@ -0,0 +1,306 @@
+// Gate-grade instrumentation for the tmux→pty fidelity burn-in (cutover Block 1).
+//
+// PURE classifiers + one stateful tracker, with NO tmux and NO @xterm import — the observer
+// loop feeds it booleans, so all classification is hermetically unit-testable. That is exactly
+// the synthetic acceptance boris requires: PROVE the instrument classifies known-behind vs
+// known-ahead and records coverage correctly BEFORE we trust any pty verdict for a flip.
+//
+// It operationalizes the adjudicated gate criterion (NOT a flat sample count):
+//   • DIRECTION — pty-BEHIND (pty=False,tmux=True: a stale frame while live tmux moved on) is
+//     the reattach-stale class → ZERO tolerance, always hard. pty-AHEAD (pty=True,tmux=False)
+//     is benign ONLY when it (a) opened in the boot window, (b) self-heals within the converge
+//     budget, and (c) is not steady-state. Anything else → hard. For READY-GATE a non-converging
+//     ahead is a FALSE-ready (early-delivery → msg-loss class), so it is hard, never benign.
+//   • COVERAGE — which matrix cells the burn-in actually traversed {runtime × {idle,
+//     model-advancing, composer-busy}} + births (fresh vs wake-resume-with-history) + restarts +
+//     history-bearing samples + fleet width. "Clean" counts only over traversed cells (idle
+//     alone is a false green).
+
+export type Verdict = 'occupancy' | 'readyGate' | 'liveness'
+export const VERDICTS: Verdict[] = ['occupancy', 'readyGate', 'liveness']
+
+export type Direction = 'agree' | 'pty-behind' | 'pty-ahead'
+
+/**
+ * Direction of a (pty, tmux) verdict pair. tmux is AUTHORITATIVE; `true` is the
+ * active/ready/alive pole. pty UNDER-reporting it (pty=false, tmux=true) is BEHIND — the
+ * stale-frame class (zero tolerance). pty OVER-reporting (pty=true, tmux=false) is AHEAD.
+ */
+export function classifyDirection(ptyVal: boolean, tmuxVal: boolean): Direction {
+  if (ptyVal === tmuxVal) return 'agree'
+  return tmuxVal ? 'pty-behind' : 'pty-ahead'
+}
+
+export type SampleState = 'idle' | 'model-advancing' | 'composer-busy'
+
+/**
+ * Per-sample fleet state for the coverage matrix. composer-busy when tmux (authoritative) sees
+ * human occupancy; else model-advancing when output churned this tick (the model is actively
+ * rendering — the NON-trivial fidelity regime, not at rest); else idle. The split is what keeps
+ * an idle burn-in from reading as a false green (boris's acceptance lesson).
+ */
+export function classifySampleState(occTmux: boolean, bytesFedThisTick: number): SampleState {
+  if (occTmux) return 'composer-busy'
+  if (bytesFedThisTick > 0) return 'model-advancing'
+  return 'idle'
+}
+
+export interface VerdictSample {
+  /** null when the tmux side was unavailable this tick (capture failed) — skipped, not scored. */
+  pty: boolean
+  tmux: boolean | null
+}
+
+export interface PeerSample {
+  peer: string
+  runtime: string
+  occupancy: VerdictSample
+  readyGate: VerdictSample
+  liveness: VerdictSample
+  /** Bytes appended to this peer's pane-log since the previous tick (drives state coverage). */
+  bytesFed: number
+  /** The model already carries scrollback (transcript taller than the pane) — a history-bearing
+   *  peer; at first sight this means it woke WITH resume-history (vs a fresh empty birth). */
+  hasHistory: boolean
+}
+
+export interface GateConfig {
+  /** Consecutive tmux-ready ticks after which a peer is considered booted (out of the boot
+   *  window). Default 3. */
+  settleTicks?: number
+  /** Max ticks an ahead-transient may stay open and still count as a benign self-heal.
+   *  Default 3. */
+  convergeBudgetTicks?: number
+}
+
+export interface HardViolation {
+  peer: string
+  runtime: string
+  verdict: Verdict
+  tick: number
+  kind: 'behind' | 'ahead-steady' | 'ahead-nonconverged'
+  detail?: string
+}
+
+interface OpenAhead {
+  startTick: number
+  bootWindow: boolean
+  flaggedNonConverged: boolean
+}
+
+interface PeerState {
+  runtime: string
+  firstTick: number
+  readyStreak: number
+  bootDone: boolean
+  open: Record<Verdict, OpenAhead | null>
+}
+
+interface VerdictTally {
+  samples: number
+  agree: number
+  behind: number
+  aheadBenign: number
+  aheadHard: number
+  aheadLatencies: number[]
+}
+
+export class GateTracker {
+  private readonly settleTicks: number
+  private readonly budget: number
+  private peers = new Map<string, PeerState>()
+  private goneOnce = new Set<string>()
+  private tally: Record<Verdict, VerdictTally>
+  private hard: HardViolation[] = []
+  // coverage
+  private fleetWidthMax = 0
+  private stateCells = new Map<string, number>() // `${runtime}/${state}` → count
+  private births = new Map<string, number>() // `${runtime}/${fresh|wake-resume}` → count
+  private restarts = new Map<string, number>() // runtime → count
+  private historySamples = new Map<string, number>() // runtime → count
+  private runtimesSeen = new Set<string>()
+
+  constructor(cfg: GateConfig = {}) {
+    this.settleTicks = cfg.settleTicks ?? 3
+    this.budget = cfg.convergeBudgetTicks ?? 3
+    const mk = (): VerdictTally => ({ samples: 0, agree: 0, behind: 0, aheadBenign: 0, aheadHard: 0, aheadLatencies: [] })
+    this.tally = { occupancy: mk(), readyGate: mk(), liveness: mk() }
+  }
+
+  /** Once per tick, before the per-peer samples — records the concurrent fleet width. */
+  tickStart(nPeers: number): void {
+    if (nPeers > this.fleetWidthMax) this.fleetWidthMax = nPeers
+  }
+
+  /** The observer dropped a peer (session gone) — a later reappearance is a restart. */
+  peerGone(addr: string): void {
+    if (this.peers.delete(addr)) this.goneOnce.add(addr)
+  }
+
+  private inc(map: Map<string, number>, key: string): void {
+    map.set(key, (map.get(key) ?? 0) + 1)
+  }
+
+  /** Record one peer's sample at `tick`. Updates boot state, classifies every verdict's
+   *  direction, manages ahead-transients, and accumulates the coverage matrix. */
+  sample(tick: number, s: PeerSample): void {
+    this.runtimesSeen.add(s.runtime)
+    let ps = this.peers.get(s.peer)
+    if (!ps) {
+      // First sight this lifetime → a birth (or a restart if we saw it die before).
+      const birthType = s.hasHistory ? 'wake-resume' : 'fresh'
+      this.inc(this.births, `${s.runtime}/${birthType}`)
+      if (this.goneOnce.has(s.peer)) {
+        this.inc(this.restarts, s.runtime)
+        this.goneOnce.delete(s.peer)
+      }
+      ps = { runtime: s.runtime, firstTick: tick, readyStreak: 0, bootDone: false, open: { occupancy: null, readyGate: null, liveness: null } }
+      this.peers.set(s.peer, ps)
+    }
+
+    // Boot-window: a peer is booted once tmux-ready has held true for settleTicks in a row.
+    if (s.readyGate.tmux === true) {
+      ps.readyStreak++
+      if (ps.readyStreak >= this.settleTicks) ps.bootDone = true
+    } else if (s.readyGate.tmux === false) {
+      ps.readyStreak = 0
+    }
+    const bootWindow = !ps.bootDone
+
+    // Coverage.
+    const state = classifySampleState(s.occupancy.tmux === true, s.bytesFed)
+    this.inc(this.stateCells, `${s.runtime}/${state}`)
+    if (s.hasHistory) this.inc(this.historySamples, s.runtime)
+
+    // Per-verdict direction + transient bookkeeping.
+    this.scoreVerdict('occupancy', tick, ps, bootWindow, s.runtime, s.peer, s.occupancy)
+    this.scoreVerdict('readyGate', tick, ps, bootWindow, s.runtime, s.peer, s.readyGate)
+    this.scoreVerdict('liveness', tick, ps, bootWindow, s.runtime, s.peer, s.liveness)
+  }
+
+  private scoreVerdict(
+    verdict: Verdict,
+    tick: number,
+    ps: PeerState,
+    bootWindow: boolean,
+    runtime: string,
+    peer: string,
+    vs: VerdictSample,
+  ): void {
+    if (vs.tmux === null) return // tmux side unavailable → not scored
+    const t = this.tally[verdict]
+    t.samples++
+    const dir = classifyDirection(vs.pty, vs.tmux)
+
+    // An open ahead-transient that exceeds the converge budget is a hard violation the moment
+    // it overruns (logged once), regardless of how it later resolves.
+    const open = ps.open[verdict]
+    if (open && !open.flaggedNonConverged && tick - open.startTick > this.budget) {
+      open.flaggedNonConverged = true
+      t.aheadHard++
+      this.hard.push({ peer, runtime, verdict, tick, kind: 'ahead-nonconverged', detail: `ahead open ${tick - open.startTick} ticks > budget ${this.budget}` })
+    }
+
+    if (dir === 'agree') {
+      t.agree++
+      if (open) {
+        // transient closed → measure self-heal latency; benign iff it opened in the boot
+        // window AND healed within budget (and was not already flagged non-converged).
+        const latency = tick - open.startTick
+        t.aheadLatencies.push(latency)
+        if (!open.flaggedNonConverged) {
+          if (open.bootWindow && latency <= this.budget) t.aheadBenign++
+          else {
+            t.aheadHard++
+            this.hard.push({ peer, runtime, verdict, tick, kind: open.bootWindow ? 'ahead-nonconverged' : 'ahead-steady', detail: `latency ${latency}, boot=${open.bootWindow}` })
+          }
+        }
+        ps.open[verdict] = null
+      }
+      return
+    }
+
+    if (dir === 'pty-behind') {
+      // The stale-frame class — always hard, zero tolerance. A behind also closes any open
+      // ahead-transient as an abnormal resolution (not a benign heal).
+      t.behind++
+      this.hard.push({ peer, runtime, verdict, tick, kind: 'behind' })
+      ps.open[verdict] = null
+      return
+    }
+
+    // pty-ahead: open a transient if none is open. Steady-state ahead is hard immediately
+    // (it cannot be a boot self-heal); boot-window ahead waits to see if it converges.
+    if (!open) {
+      if (!bootWindow) {
+        t.aheadHard++
+        this.hard.push({ peer, runtime, verdict, tick, kind: 'ahead-steady' })
+        // still open it so we can measure when it clears, but it is already counted hard
+        ps.open[verdict] = { startTick: tick, bootWindow: false, flaggedNonConverged: true }
+      } else {
+        ps.open[verdict] = { startTick: tick, bootWindow: true, flaggedNonConverged: false }
+      }
+    }
+  }
+
+  /** True iff no hard divergence has been recorded for `verdict` (the per-verdict gate bar). */
+  verdictGateClean(verdict: Verdict): boolean {
+    const t = this.tally[verdict]
+    return t.behind === 0 && t.aheadHard === 0
+  }
+
+  summary(): GateSummary {
+    const latStats = (xs: number[]): { count: number; max: number; mean: number } =>
+      xs.length ? { count: xs.length, max: Math.max(...xs), mean: +(xs.reduce((a, b) => a + b, 0) / xs.length).toFixed(2) } : { count: 0, max: 0, mean: 0 }
+    const verdicts = {} as Record<Verdict, VerdictGateStat>
+    for (const v of VERDICTS) {
+      const t = this.tally[v]
+      verdicts[v] = {
+        samples: t.samples,
+        agree: t.agree,
+        behind: t.behind,
+        aheadBenign: t.aheadBenign,
+        aheadHard: t.aheadHard,
+        selfHeal: latStats(t.aheadLatencies),
+        gateClean: this.verdictGateClean(v),
+      }
+    }
+    return {
+      verdicts,
+      hardViolations: this.hard.slice(),
+      coverage: {
+        fleetWidthMax: this.fleetWidthMax,
+        runtimes: [...this.runtimesSeen].sort(),
+        byRuntimeState: Object.fromEntries([...this.stateCells].sort()),
+        births: Object.fromEntries([...this.births].sort()),
+        restarts: Object.fromEntries([...this.restarts].sort()),
+        historySamples: Object.fromEntries([...this.historySamples].sort()),
+      },
+    }
+  }
+}
+
+export interface VerdictGateStat {
+  samples: number
+  agree: number
+  behind: number
+  aheadBenign: number
+  aheadHard: number
+  selfHeal: { count: number; max: number; mean: number }
+  gateClean: boolean
+}
+
+export interface GateCoverage {
+  fleetWidthMax: number
+  runtimes: string[]
+  byRuntimeState: Record<string, number>
+  births: Record<string, number>
+  restarts: Record<string, number>
+  historySamples: Record<string, number>
+}
+
+export interface GateSummary {
+  verdicts: Record<Verdict, VerdictGateStat>
+  hardViolations: HardViolation[]
+  coverage: GateCoverage
+}

package/src/shadow/index.ts

@@ -19,6 +19,7 @@ import { tmux, composerCaptureHasHumanInput, listOnlinePeers } from '../transpor
 import { getAdapter } from '../launch/index.ts'
 import { buildSocketPath, buildProcessAddress } from '../core/socket.ts'
 import type { Runtime } from '../core/constants.ts'
+import { GateTracker, classifyDirection, VERDICTS } from './gate.ts'
 
 const SEED_BYTES = 4 * 1024 * 1024
 const promptGlyph = (rt: string): string => (rt === 'claude' ? '❯' : '›')
@@ -72,16 +73,18 @@ export interface ShadowOptions {
   log?: (s: string) => void
 }
 
-function feed(o: Obs): Promise<void> {
+/** Feed the pane-log delta into the model; resolves with the byte count appended this tick
+ *  (drives the coverage state classifier: >0 ⇒ the model is advancing, not idle). */
+function feed(o: Obs): Promise<number> {
   const sz = fstatSync(o.fd).size
-  if (sz <= o.offset) return Promise.resolve()
+  if (sz <= o.offset) return Promise.resolve(0)
   const buf = Buffer.alloc(sz - o.offset)
   readSync(o.fd, buf, 0, buf.length, o.offset)
   o.offset = sz
-  return new Promise(res => o.term.write(buf, () => res()))
+  return new Promise(res => o.term.write(buf, () => res(buf.length)))
 }
 
-function enumerate(models: Map<string, Obs>, opts: ShadowOptions): void {
+function enumerate(models: Map<string, Obs>, opts: ShadowOptions, tracker?: GateTracker): void {
   let peers: { runtime: Runtime; personality: string }[] = []
   try { peers = listOnlinePeers(opts.sockDir).filter(p => p.runtime === 'claude' || p.runtime === 'codex') } catch { return }
   const live = new Set<string>()
@@ -98,7 +101,7 @@ function enumerate(models: Map<string, Obs>, opts: ShadowOptions): void {
       models.set(addr, { addr, sock, runtime: p.runtime, cols, rows, term, fd, offset })
     } catch { /* a peer racing spawn/teardown — skip this round */ }
   }
-  for (const [addr, o] of models) if (!live.has(addr)) { try { closeSync(o.fd) } catch { /* */ }; models.delete(addr) }
+  for (const [addr, o] of models) if (!live.has(addr)) { try { closeSync(o.fd) } catch { /* */ }; tracker?.peerGone(addr); models.delete(addr) }
 }
 
 /** Run the read-only fidelity observer. Resolves when once/maxMinutes/STOP completes. */
@@ -110,19 +113,20 @@ export async function runShadowFidelity(opts: ShadowOptions): Promise<void> {
   const interval = opts.intervalMs ?? 5000
   const log = opts.log ?? (() => { /* */ })
   const models = new Map<string, Obs>()
-  const tot = { samples: 0, occDiv: 0, occNt: 0, rgDiv: 0, livDiv: 0 }
+  const tracker = new GateTracker()
   const startedISO = new Date().toISOString(), t0 = Date.now()
-  enumerate(models, opts)
+  enumerate(models, opts, tracker)
   for (const o of models.values()) await feed(o)  // seed each model
   log(`shadow-fidelity: ${models.size} warm peers, interval ${interval}ms\n`)
   let i = 0
   for (;;) {
     if (opts.maxMinutes && Date.now() - t0 >= opts.maxMinutes * 60000) break
     if (existsSync(stopFile)) break
-    if (i > 0 && i % 15 === 0) enumerate(models, opts)
+    if (i > 0 && i % 15 === 0) enumerate(models, opts, tracker)
+    tracker.tickStart(models.size)
     for (const o of models.values()) {
       try {
-        await feed(o)
+        const bytesFed = await feed(o)
         const adapter = getAdapter(o.runtime)
         const capE = tmux(o.sock, 'capture-pane', '-e', '-p', '-t', o.addr); if (!capE.ok) continue
         const occTmux = composerCaptureHasHumanInput(capE.out, adapter.deliveryMarkers)
@@ -134,34 +138,49 @@ export async function runShadowFidelity(opts: ShadowOptions): Promise<void> {
         const ppid = tmux(o.sock, 'display', '-p', '-t', o.addr, '#{pane_pid}')
         const pid = ppid.ok ? Number(ppid.out.trim()) : NaN
         const livPty = Number.isFinite(pid) ? pidAlive(pid) : false
+        const hasHistory = o.term.buffer.active.baseY > 0
 
-        tot.samples++
-        const occDiv = occPty !== occTmux, rgDiv = rgTmux !== null && rgPty !== rgTmux, livDiv = livPty !== livTmux
-        const occNt = occPty || occTmux
-        if (occNt) tot.occNt++
-        if (occDiv) tot.occDiv++
-        if (rgDiv) tot.rgDiv++
-        if (livDiv) tot.livDiv++
-        if (occDiv || rgDiv || livDiv) {
+        // Feed the gate tracker — it scores DIRECTION (behind=hard / ahead=benign-if-boot-self-heal)
+        // and accumulates the COVERAGE matrix. The boolean booleans never leave this call.
+        tracker.sample(i, {
+          peer: o.addr, runtime: o.runtime,
+          occupancy: { pty: occPty, tmux: occTmux },
+          readyGate: { pty: rgPty, tmux: rgTmux },
+          liveness: { pty: livPty, tmux: livTmux },
+          bytesFed, hasHistory,
+        })
+
+        // DIRECTION-AWARE per-event divergence trail (gate scoring lives in the tracker; this is
+        // the forensic log). Occupancy carries the raw capture + model frames for root-cause.
+        const occDir = classifyDirection(occPty, occTmux)
+        const rgDir = rgTmux === null ? 'agree' : classifyDirection(rgPty, rgTmux)
+        const livDir = classifyDirection(livPty, livTmux)
+        if (occDir !== 'agree' || rgDir !== 'agree' || livDir !== 'agree') {
           appendFileSync(jsonl, JSON.stringify({
-            ts: new Date().toISOString(), peer: o.addr, geom: `${o.cols}x${o.rows}`,
-            occupancy: { pty: occPty, tmux: occTmux, diverged: occDiv, nontrivial: occNt },
-            readyGate: { pty: rgPty, tmux: rgTmux, diverged: rgDiv },
-            liveness: { pty: livPty, tmux: livTmux, diverged: livDiv },
-            ...(occDiv ? { capE: capE.out.slice(-2500), model: modelToPlainText(o.term, o.cols, o.rows).slice(-2500) } : {}),
+            ts: new Date().toISOString(), peer: o.addr, geom: `${o.cols}x${o.rows}`, hasHistory,
+            occupancy: { pty: occPty, tmux: occTmux, dir: occDir },
+            readyGate: { pty: rgPty, tmux: rgTmux, dir: rgDir },
+            liveness: { pty: livPty, tmux: livTmux, dir: livDir },
+            ...(occDir !== 'agree' ? { capE: capE.out.slice(-2500), model: modelToPlainText(o.term, o.cols, o.rows).slice(-2500) } : {}),
           }) + '\n')
-          log(`DIVERGE ${o.addr} occ=${occDiv} rg=${rgDiv} liv=${livDiv}${occNt && occDiv ? ' *** NON-TRIVIAL occ ***' : ''}\n`)
+          log(`DIVERGE ${o.addr} occ=${occDir} rg=${rgDir} liv=${livDir}\n`)
         }
       } catch { /* an observer error must NEVER escape — it is a passive sidecar */ }
     }
     i++
+    const gs = tracker.summary()
+    const totalSamples = VERDICTS.reduce((n, v) => n + gs.verdicts[v].samples, 0)
     writeFileSync(statsFile, JSON.stringify({
       startedISO, lastISO: new Date().toISOString(), elapsedMin: +((Date.now() - t0) / 60000).toFixed(1),
-      peers: models.size, samples: tot.samples,
-      occupancy: { divergences: tot.occDiv, nonTrivialSamples: tot.occNt },
-      readyGate: { divergences: tot.rgDiv },
-      liveness: { divergences: tot.livDiv },
-      fidelity: tot.samples ? +(1 - (tot.occDiv + tot.rgDiv + tot.livDiv) / (tot.samples * 3)).toFixed(6) : 1,
+      peers: models.size, samples: totalSamples,
+      // Gate (boris's criterion): clean iff zero hard divergences across all verdicts. Hard =
+      // every pty-behind + non-benign pty-ahead (steady-state or non-converging). Benign =
+      // boot-window ahead that self-heals within budget. Coverage proves the matrix was traversed.
+      gateClean: VERDICTS.every(v => gs.verdicts[v].gateClean),
+      hardTotal: gs.hardViolations.length,
+      verdicts: gs.verdicts,
+      coverage: gs.coverage,
+      recentHard: gs.hardViolations.slice(-20),
     }, null, 2))
     if (opts.once) break
     await new Promise(r => setTimeout(r, interval))

package/src/shadow/install.test.ts

@@ -0,0 +1,153 @@
+import { afterEach, describe, expect, test } from 'bun:test'
+import { existsSync, mkdirSync, mkdtempSync, readFileSync, rmSync, statSync, writeFileSync } from 'node:fs'
+import { spawnSync } from 'node:child_process'
+import { tmpdir } from 'node:os'
+import { join } from 'node:path'
+import { isFoundationOwnedPlist } from '../launch/launchd.ts'
+import {
+  SHADOW_PERSONALITY,
+  SHADOW_PLIST_LABEL,
+  buildShadowPlistSpec,
+  cycleShadowJob,
+  installShadowJob,
+  installShadowPlist,
+  shadowPlistPath,
+  uninstallShadowJob,
+} from './install.ts'
+
+// Hermetic: every plist write lands under IAPEER_LAUNCHAGENTS_DIR / IAPEER_ROOT overrides, and
+// the suite runs with IAPEER_TEST_SANDBOX=1 (process.env) so the launchctl-touching paths
+// (installShadowJob / cycleShadowJob / uninstallShadowJob) fail-CLOSED to skipped-sandbox — they
+// NEVER load/unload a real launchd job. No tmux, no @xterm (install.ts is pure plumbing).
+
+const tmpDirs: string[] = []
+function mkEnv(extra: Record<string, string> = {}): NodeJS.ProcessEnv {
+  const root = mkdtempSync(join(tmpdir(), 'iapeer-shadow-install-'))
+  tmpDirs.push(root)
+  return {
+    IAPEER_LAUNCHAGENTS_DIR: join(root, 'LaunchAgents'),
+    IAPEER_ROOT: join(root, 'iapeer'),
+    HOME: root,
+    PATH: '/usr/bin:/bin',
+    ...extra,
+  } as NodeJS.ProcessEnv
+}
+afterEach(() => {
+  while (tmpDirs.length) rmSync(tmpDirs.pop()!, { recursive: true, force: true })
+})
+const plutilLint = (path: string): boolean => spawnSync('plutil', ['-lint', path], { encoding: 'utf8' }).status === 0
+
+describe('buildShadowPlistSpec', () => {
+  test('label is com.iapeer.shadow-fidelity and runs the INSTALLED `iapeer shadow`', () => {
+    const spec = buildShadowPlistSpec({ HOME: '/Users/x' } as NodeJS.ProcessEnv)
+    expect(spec.label).toBe(SHADOW_PLIST_LABEL)
+    expect(SHADOW_PLIST_LABEL).toBe('com.iapeer.shadow-fidelity')
+    expect(spec.programArguments).toEqual(['/Users/x/.local/bin/iapeer', 'shadow'])
+  })
+  test('launchd-minimal PATH includes /opt/homebrew/bin (tmux lives there — the observer shells read-only tmux)', () => {
+    expect(buildShadowPlistSpec({ HOME: '/Users/x' } as NodeJS.ProcessEnv).environment.PATH).toContain('/opt/homebrew/bin')
+  })
+  test('stdout/stderr land under ~/.iapeer/logs/iapeer', () => {
+    const spec = buildShadowPlistSpec(mkEnv())
+    expect(spec.stdoutPath).toContain(join('logs', 'iapeer', 'shadow-fidelity.launchd-stdout.log'))
+    expect(spec.stderrPath).toContain(join('logs', 'iapeer', 'shadow-fidelity.launchd-stderr.log'))
+  })
+})
+
+describe('installShadowPlist', () => {
+  test('writes a valid, foundation-owned (sentinel) always-on plist under the override dir', () => {
+    const env = mkEnv()
+    const r = installShadowPlist(env)
+    expect(r.path).toBe(shadowPlistPath(env))
+    expect(r.changed).toBe(true)
+    expect(existsSync(r.path)).toBe(true)
+    expect(isFoundationOwnedPlist(r.path)).toBe(true) // carries the ownership sentinel
+    expect(plutilLint(r.path)).toBe(true)
+    const xml = readFileSync(r.path, 'utf8')
+    expect(xml).toContain(`<string>${SHADOW_PLIST_LABEL}</string>`)
+    expect(xml).toContain('<string>shadow</string>')
+    expect(xml).toContain('<key>RunAtLoad</key>')
+    expect(xml).toContain('<key>KeepAlive</key>')
+  })
+
+  test('idempotent by content: an unchanged re-install does NOT rewrite (changed:false, mtime stable); a changed PATH DOES', () => {
+    const env = mkEnv()
+    const first = installShadowPlist(env)
+    expect(first.changed).toBe(true)
+    const mtime1 = statSync(first.path).mtimeMs
+    const second = installShadowPlist(env)
+    expect(second.changed).toBe(false)
+    expect(statSync(second.path).mtimeMs).toBe(mtime1) // file untouched (no BTM spam)
+    // a different rendered plist (PATH override) rewrites
+    const third = installShadowPlist(mkEnv({ IAPEER_SHADOW_PATH: '/opt/homebrew/bin:/usr/bin:/bin', IAPEER_LAUNCHAGENTS_DIR: env.IAPEER_LAUNCHAGENTS_DIR!, IAPEER_ROOT: env.IAPEER_ROOT!, HOME: env.HOME! }))
+    expect(third.changed).toBe(true)
+  })
+
+  test('collision guard: REFUSES to overwrite a foreign (no-sentinel) plist at the label', () => {
+    const env = mkEnv()
+    const laDir = env.IAPEER_LAUNCHAGENTS_DIR!
+    mkdirSync(laDir, { recursive: true })
+    const path = shadowPlistPath(env)
+    const foreign = '<?xml version="1.0"?>\n<plist><dict><key>Label</key><string>com.iapeer.shadow-fidelity</string></dict></plist>\n'
+    writeFileSync(path, foreign)
+    expect(() => installShadowPlist(env)).toThrow(/foundation-managed|refus/i)
+    expect(readFileSync(path, 'utf8')).toBe(foreign) // untouched — guard fires before any write
+  })
+})
+
+describe('installShadowJob (sandbox fail-closed)', () => {
+  test('writes the plist FILE but NEVER touches launchctl under IAPEER_TEST_SANDBOX → skipped-sandbox', () => {
+    const env = mkEnv()
+    const r = installShadowJob(env)
+    expect(r.action).toBe('skipped-sandbox')
+    expect(r.path).toBe(shadowPlistPath(env))
+    expect(r.changed).toBe(true)
+    expect(existsSync(r.path)).toBe(true) // file written to the override dir
+    expect(isFoundationOwnedPlist(r.path)).toBe(true)
+  })
+})
+
+describe('cycleShadowJob (deploy re-cycle hook)', () => {
+  test('returns null when no plist exists (nothing to recycle)', () => {
+    expect(cycleShadowJob(mkEnv())).toBeNull()
+  })
+  test('a written foundation plist recycles to skipped-sandbox under the test flag (never a real launchctl cycle)', () => {
+    const env = mkEnv()
+    installShadowPlist(env)
+    expect(cycleShadowJob(env)?.state).toBe('skipped-sandbox')
+  })
+  test('a foreign plist at the label is NOT recycled (ownership guard) → null', () => {
+    const env = mkEnv()
+    mkdirSync(env.IAPEER_LAUNCHAGENTS_DIR!, { recursive: true })
+    writeFileSync(shadowPlistPath(env), '<?xml version="1.0"?>\n<plist><dict/></plist>\n')
+    expect(cycleShadowJob(env)).toBeNull()
+  })
+})
+
+describe('uninstallShadowJob', () => {
+  test('absent plist → absent (no-op)', () => {
+    expect(uninstallShadowJob(mkEnv()).action).toBe('absent')
+  })
+  test('foundation plist under sandbox → skipped-sandbox, plist left in place (no real bootout)', () => {
+    const env = mkEnv()
+    installShadowPlist(env)
+    const r = uninstallShadowJob(env)
+    expect(r.action).toBe('skipped-sandbox')
+    expect(existsSync(shadowPlistPath(env))).toBe(true)
+  })
+  test('foreign plist → refused-foreign, untouched', () => {
+    const env = mkEnv()
+    mkdirSync(env.IAPEER_LAUNCHAGENTS_DIR!, { recursive: true })
+    const path = shadowPlistPath(env)
+    writeFileSync(path, '<?xml version="1.0"?>\n<plist><dict/></plist>\n')
+    expect(uninstallShadowJob(env).action).toBe('refused-foreign')
+    expect(existsSync(path)).toBe(true)
+  })
+})
+
+describe('exports wire to the running job', () => {
+  test('SHADOW_PERSONALITY/LABEL match the live burn-in job naming', () => {
+    expect(SHADOW_PERSONALITY).toBe('shadow-fidelity')
+    expect(SHADOW_PLIST_LABEL).toBe('com.iapeer.shadow-fidelity')
+  })
+})

package/src/shadow/install.ts

@@ -0,0 +1,256 @@
+// Code-managed install / teardown / update-recycle for the read-only tmux→pty fidelity
+// BURN-IN job (`iapeer shadow`). The OBSERVER itself is src/shadow/index.ts (it loads
+// @xterm/headless); THIS module is PURE launchd plumbing — it imports NO @xterm — so the
+// deploy path (update/rollback) and the CLI can install/recycle the job without pulling the
+// terminal emulator into their import graph (the daemon hot path stays @xterm-free).
+//
+// The job is a foundation-OWNED, always-on launchd LaunchAgent (com.iapeer.shadow-fidelity)
+// running the INSTALLED binary `iapeer shadow`. It is NOT a registry peer — it receives no
+// IAP and has no peer profile — so recycleFoundationOwnedInfraJobs' registry SCAN cannot see
+// it; cycleShadowJob recycles it by LABEL instead. It carries the same ownership sentinel as
+// every foundation plist, so the collision guard (install) and the recycle guard both
+// recognize it as ours. KeepAlive owns its lifecycle; stop it with `iapeer shadow-uninstall`
+// (bootout + rm), NEVER the STOP sentinel file (KeepAlive would just respawn straight past it).
+
+import { existsSync, mkdirSync, readFileSync, rmSync, unlinkSync, writeFileSync } from 'node:fs'
+import { homedir } from 'node:os'
+import { join } from 'node:path'
+import { spawnSync } from 'node:child_process'
+import { IAPEER_DIR } from '../core/constants.ts'
+import { IapError } from '../core/errors.ts'
+import { iapeerBinPath } from '../install/index.ts'
+import { pluginLogsDir } from '../storage/index.ts'
+import {
+  cycleLaunchdJob,
+  isFoundationOwnedPlist,
+  launchAgentsDir,
+  launchctlBootstrap,
+  launchdLabel,
+  renderLaunchdPlist,
+  type LaunchdJobCycleResult,
+  type LaunchdPlistSpec,
+} from '../launch/launchd.ts'
+
+/** The burn-in job's personality stem → label com.iapeer.shadow-fidelity (the running job). */
+export const SHADOW_PERSONALITY = 'shadow-fidelity'
+/** com.iapeer.shadow-fidelity — keyed through launchdLabel so the scheme has ONE source. */
+export const SHADOW_PLIST_LABEL = launchdLabel(SHADOW_PERSONALITY)
+
+export function shadowPlistPath(env: NodeJS.ProcessEnv = process.env): string {
+  return join(launchAgentsDir(env), `${SHADOW_PLIST_LABEL}.plist`)
+}
+
+/** The STOP sentinel the observer polls (under <eventLogDir>=~/.iapeer/logs/iapeer). Under
+ *  KeepAlive a STALE one would respawn the observer straight into an immediate exit → a
+ *  throttle-storm, so installShadowJob clears it before (re)loading. */
+function shadowStopFile(env: NodeJS.ProcessEnv = process.env): string {
+  return join(pluginLogsDir('iapeer', { env }), 'shadow-fidelity.STOP')
+}
+
+function inSandbox(env: NodeJS.ProcessEnv): boolean {
+  return env.IAPEER_TEST_SANDBOX === '1' || process.env.IAPEER_TEST_SANDBOX === '1'
+}
+
+/** Resolve the gui-domain uid, or null when `id -u` is non-numeric (never assume 0). */
+function uidOrNull(): string | null {
+  const u = (spawnSync('id', ['-u'], { encoding: 'utf8' }).stdout ?? '').trim()
+  return /^\d+$/.test(u) ? u : null
+}
+
+/** Is com.iapeer.shadow-fidelity currently loaded in the gui domain? */
+function isShadowLoaded(): boolean {
+  const uid = uidOrNull()
+  if (!uid) return false
+  return spawnSync('launchctl', ['print', `gui/${uid}/${SHADOW_PLIST_LABEL}`], { stdio: 'ignore' }).status === 0
+}
+
+/**
+ * Build the burn-in job's launchd plist spec (PURE — render/lint-testable). Runs the
+ * INSTALLED `iapeer shadow` (prod decoupled from the src tree). The launchd-minimal PATH
+ * MUST include /opt/homebrew/bin: the observer shells READ-ONLY tmux (capture-pane /
+ * has-session / display), and tmux lives there — a bare PATH would make every tmux probe
+ * fail (no peers enumerated). Deterministic (curated PATH, not env.PATH) so the rendered
+ * plist is byte-stable across deploys → installShadowPlist's write-if-changed stays a no-op
+ * when nothing actually changed (no Background Task Management notification spam).
+ */
+export function buildShadowPlistSpec(env: NodeJS.ProcessEnv = process.env): LaunchdPlistSpec {
+  const home = env.HOME?.trim() || homedir()
+  const defaultPath = `${home}/.local/bin:${home}/.bun/bin:/opt/homebrew/bin:/opt/homebrew/sbin:/usr/local/bin:/usr/bin:/bin:/usr/sbin:/sbin`
+  const logDir = pluginLogsDir('iapeer', { env })
+  const environment: Record<string, string> = { PATH: env.IAPEER_SHADOW_PATH?.trim() || defaultPath }
+  // Sandbox/test parity (mirror of installDaemonPlist / installAlwaysOnPlist): bake the
+  // non-default root overrides so a SANDBOXED install resolves the SAME isolated tree the
+  // observer reads. In production these are unset → nothing baked, real ~/.iapeer + /tmp used.
+  for (const key of ['IAPEER_ROOT', 'IAPEER_SOCK_DIR', 'IAPEER_LAUNCHAGENTS_DIR'] as const) {
+    if (env[key]?.trim()) environment[key] = env[key]!.trim()
+  }
+  return {
+    label: SHADOW_PLIST_LABEL,
+    programArguments: [iapeerBinPath(env), 'shadow'],
+    workingDirectory: home, // cwd-agnostic (the observer resolves paths from IAPEER_ROOT/cfg)
+    environment,
+    stdoutPath: join(logDir, 'shadow-fidelity.launchd-stdout.log'),
+    stderrPath: join(logDir, 'shadow-fidelity.launchd-stderr.log'),
+  }
+}
+
+export interface InstallShadowPlistResult {
+  path: string
+  /** false when the on-disk plist already matched byte-for-byte → NO write (no BTM spam). */
+  changed: boolean
+}
+
+/**
+ * Write the burn-in plist (file only — does NOT load it). IDEMPOTENT BY CONTENT
+ * (write-if-changed, mirror of installDaemonPlist: a byte-identical rewrite trips the owner's
+ * Background Task Management notification for zero benefit). Collision guard: refuses to
+ * overwrite a com.iapeer.shadow-fidelity.plist that lacks the ownership sentinel (a foreign
+ * manager owns it) — the guard fires FIRST, before any mkdir/write, so a refusal leaves the
+ * filesystem untouched.
+ */
+export function installShadowPlist(env: NodeJS.ProcessEnv = process.env): InstallShadowPlistResult {
+  const path = shadowPlistPath(env)
+  if (existsSync(path) && !isFoundationOwnedPlist(path)) {
+    throw new IapError(
+      `refusing to overwrite ${path}: ${SHADOW_PLIST_LABEL} exists but is not foundation-managed ` +
+        `(no ownership sentinel) — another manager owns it`,
+    )
+  }
+  const rendered = renderLaunchdPlist(buildShadowPlistSpec(env))
+  mkdirSync(launchAgentsDir(env), { recursive: true })
+  // Never mkdir the REAL ~/.iapeer log dir under a sandbox that forgot IAPEER_ROOT (mirror
+  // installAlwaysOnPlist's guard); a real run (no sandbox flag) always makes it.
+  const logDir = pluginLogsDir('iapeer', { env })
+  const realRoot = join(homedir(), IAPEER_DIR)
+  if (!(inSandbox(env) && logDir.startsWith(`${realRoot}/`))) {
+    mkdirSync(logDir, { recursive: true, mode: 0o700 })
+  }
+  let existing: string | null = null
+  try {
+    existing = readFileSync(path, 'utf8')
+  } catch {
+    existing = null
+  }
+  if (existing === rendered) return { path, changed: false }
+  writeFileSync(path, rendered, { mode: 0o644 })
+  return { path, changed: true }
+}
+
+export type ShadowJobAction =
+  | 'loaded' // bootstrapped now (was not loaded)
+  | 'reloaded' // plist changed on an already-loaded job → bootout+bootstrap applied the new one
+  | 'already-loaded' // loaded and the plist did not change → left running untouched
+  | 'refused-foreign' // a non-foundation plist sits at the label → never loaded
+  | 'failed' // launchctl bootstrap/cycle failed (detail carries the rescue recipe)
+  | 'skipped-sandbox' // IAPEER_TEST_SANDBOX=1 → never touch real launchd
+
+export interface InstallShadowJobResult extends InstallShadowPlistResult {
+  action: ShadowJobAction
+  detail?: string
+}
+
+/**
+ * Install the burn-in plist AND ensure the job is live: write-if-changed, clear a stale STOP
+ * sentinel, then load it (or RELOAD via bootout+bootstrap when the plist content changed on
+ * an already-loaded job — bootstrap alone is idempotent and would NOT pick up a changed
+ * plist). Idempotent: a repeat call on an unchanged, loaded job is a no-op ('already-loaded').
+ * Sandbox-safe: under IAPEER_TEST_SANDBOX it writes the (overridden-dir) plist FILE but never
+ * calls launchctl ('skipped-sandbox'). This is the code-managed replacement for a hand-written
+ * burn-in plist — reproducible, ownership-sentinel-guarded, fleet-deployable.
+ */
+export function installShadowJob(env: NodeJS.ProcessEnv = process.env): InstallShadowJobResult {
+  const { path, changed } = installShadowPlist(env)
+  if (inSandbox(env)) return { path, changed, action: 'skipped-sandbox' }
+  // Clear a STALE STOP sentinel BEFORE (re)loading — else KeepAlive respawns the observer
+  // into an instant exit (it polls the STOP file) → 10s-throttled respawn storm.
+  try {
+    unlinkSync(shadowStopFile(env))
+  } catch {
+    /* none present */
+  }
+  if (isShadowLoaded()) {
+    if (!changed) return { path, changed, action: 'already-loaded' }
+    const c = cycleLaunchdJob(SHADOW_PLIST_LABEL, path, env)
+    return c.state === 'restarted'
+      ? { path, changed, action: 'reloaded' }
+      : { path, changed, action: 'failed', detail: c.detail }
+  }
+  const b = launchctlBootstrap(SHADOW_PERSONALITY, path, env)
+  const action: ShadowJobAction =
+    b.state === 'loaded'
+      ? 'loaded'
+      : b.state === 'already-loaded'
+        ? 'already-loaded'
+        : b.state === 'refused-foreign'
+          ? 'refused-foreign'
+          : b.state === 'skipped-sandbox'
+            ? 'skipped-sandbox'
+            : 'failed'
+  return { path, changed, action, detail: b.detail }
+}
+
+/**
+ * Recycle the burn-in job by LABEL after a binary swap (deploy path). The plist runs the
+ * just-replaced `iapeer shadow` off the same binary, so the SAME LWCR managed-launch-constraint
+ * bomb that bit the daemon (EX_CONFIG(78) on a KeepAlive respawn into the stale constraint)
+ * applies here too — only a full bootout→bootstrap clears it. It is NOT a registry peer, so
+ * recycleFoundationOwnedInfraJobs' registry scan misses it; this cycles it directly. Only
+ * touches a LOADED, foundation-owned job: absent / foreign / not-loaded → null (no-op).
+ * Sandbox-safe (cycleLaunchdJob returns skipped-sandbox under the test flag).
+ */
+export function cycleShadowJob(env: NodeJS.ProcessEnv = process.env): LaunchdJobCycleResult | null {
+  const path = shadowPlistPath(env)
+  if (!existsSync(path) || !isFoundationOwnedPlist(path)) return null
+  return cycleLaunchdJob(SHADOW_PLIST_LABEL, path, env)
+}
+
+export type ShadowUninstallAction =
+  | 'booted-out' // job was loaded → booted out, plist removed
+  | 'removed-not-loaded' // plist removed, job was not loaded
+  | 'absent' // nothing to do (no plist)
+  | 'refused-foreign' // the plist is not foundation-owned → left untouched
+  | 'skipped-sandbox'
+  | 'failed'
+
+export interface UninstallShadowJobResult {
+  path: string
+  action: ShadowUninstallAction
+  detail?: string
+}
+
+/**
+ * Stop and remove the burn-in job — the CORRECT way to halt it (the STOP sentinel would be
+ * respawned past by KeepAlive): bootout the loaded job, then delete the plist. Refuses a
+ * non-foundation plist (ownership guard). Sandbox-safe. Reversibility for the install verb.
+ */
+export function uninstallShadowJob(env: NodeJS.ProcessEnv = process.env): UninstallShadowJobResult {
+  const path = shadowPlistPath(env)
+  if (!existsSync(path)) return { path, action: 'absent' }
+  if (!isFoundationOwnedPlist(path)) {
+    return { path, action: 'refused-foreign', detail: `${path} is not foundation-owned — not booting it out` }
+  }
+  if (inSandbox(env)) return { path, action: 'skipped-sandbox' }
+  let action: ShadowUninstallAction = 'removed-not-loaded'
+  if (isShadowLoaded()) {
+    const uid = uidOrNull()
+    const r = uid
+      ? spawnSync('launchctl', ['bootout', `gui/${uid}/${SHADOW_PLIST_LABEL}`], { encoding: 'utf8' })
+      : { status: 1, stderr: 'cannot resolve uid' }
+    if (r.status !== 0 && isShadowLoaded()) {
+      return { path, action: 'failed', detail: `launchctl bootout failed: ${(r.stderr ?? '').trim() || `exit ${r.status}`}` }
+    }
+    action = 'booted-out'
+  }
+  // Also clear any STOP sentinel so a later reinstall starts clean.
+  try {
+    unlinkSync(shadowStopFile(env))
+  } catch {
+    /* none */
+  }
+  try {
+    rmSync(path)
+  } catch {
+    /* best-effort */
+  }
+  return { path, action }
+}

package/src/update/index.ts

@@ -41,6 +41,7 @@ import {
   type DaemonRestartResult,
   type LaunchdJobCycleResult,
 } from '../launch/launchd.ts'
+import { cycleShadowJob, SHADOW_PERSONALITY, SHADOW_PLIST_LABEL } from '../shadow/install.ts'
 import { defaultDaemonSocketPath } from '../daemon/index.ts'
 
 /** The npm package the foundation publishes / updates from. */
@@ -242,6 +243,14 @@ export function recycleFoundationOwnedInfraJobs(env: NodeJS.ProcessEnv = process
     const r = cycleLaunchdJob(label, plist, env)
     out.push({ personality: peer.personality, runtime, label, state: r.state, detail: r.detail })
   }
+  // The tmux→pty fidelity BURN-IN job (com.iapeer.shadow-fidelity) is foundation-owned and
+  // runs the just-replaced `iapeer shadow` off the same binary, so it carries the same LWCR
+  // constraint — but it is NOT a registry peer, so the scan above cannot reach it. Cycle it
+  // by label too (no-op when absent / not-loaded / foreign).
+  const shadow = cycleShadowJob(env)
+  if (shadow) {
+    out.push({ personality: SHADOW_PERSONALITY, runtime: 'foundation', label: SHADOW_PLIST_LABEL, state: shadow.state, detail: shadow.detail })
+  }
   return out
 }