diff --git a/src/main/daemon/AGENTS.md b/src/main/daemon/AGENTS.md
index 050a355fdfd..2c5f555be0a 100644
--- a/src/main/daemon/AGENTS.md
+++ b/src/main/daemon/AGENTS.md
@@ -89,6 +89,19 @@ hands → probe once more → `rename` in one syscall → verify we kept it.
     read only afterwards. Pre-existing, and narrowed by this change rather than widened: the
     window now opens only after the daemon has itself reported zero sessions.
 
+  Known limits of the process-table evidence, none of which can license a kill on their own —
+  each only fails to *raise* a verdict, so the cost is a hold not taken:
+
+  - A PTY whose session leader has exited leaves its still-running child reparented outside the
+    daemon's descendant tree. The walk cannot see it, so a daemon with real work can read as
+    childless.
+  - On Windows the evidence abstains entirely. A daemon that closed its listener but is still
+    draining sessions therefore has no protection from the endpoint-dead path.
+  - The self-spawned-probe exclusion matches an exact argv (`sh -c exit 0`). A hosted session
+    leader whose executable basename is `sh` and whose command is exactly that would be
+    discarded. Contrived — `exit 0` returns immediately — but it is executable identity the
+    match cannot establish.
+
   **If you ever raise the classification budget, gate the replace path on headroom first.**
   The budget serves two verdicts with opposite time-costs: reaching "don't kill" slowly is free,
   because the daemon survives however long it took, while reaching `empty` slowly is not — the
diff --git a/src/main/daemon/daemon-live-pty-evidence.test.ts b/src/main/daemon/daemon-live-pty-evidence.test.ts
index 9a2ce8df9bb..d1d12f9bee2 100644
--- a/src/main/daemon/daemon-live-pty-evidence.test.ts
+++ b/src/main/daemon/daemon-live-pty-evidence.test.ts
@@ -271,6 +271,29 @@ describe('inspectDaemonPtyOwnership on win32', () => {
 })
 
 describe('inspectDaemonPtyOwnership sampling', () => {
+  it('will not confirm emptiness from the cached table the first read already used', async () => {
+    // Two agreeing samples are only worth more than one if they are two observations. When the
+    // fresh read is slow — the busy host this evidence exists for — both attempts fell through
+    // to the same TTL-cached snapshot, so a login(1) wrapper photographed before its shell
+    // appeared could be 'confirmed' empty by a second look at the same photograph.
+    const readCachedPosixProcessTable = vi.fn<() => Promise<ProcessTableRow[]>>(async () => [
+      daemonRow
+    ])
+
+    await expect(
+      inspectDaemonPtyOwnership(DAEMON_PID, {
+        platform: 'darwin',
+        posixDeadlineMs: 5,
+        // Never settles inside the deadline, so every attempt reaches for the cache.
+        readPosixProcessTable: () => new Promise<ProcessTableRow[]>(() => {}),
+        readCachedPosixProcessTable
+      })
+    ).resolves.toBe('unknown')
+
+    // Only the first sample may be served from the cache; the confirming one must not be.
+    expect(readCachedPosixProcessTable).toHaveBeenCalledTimes(1)
+  })
+
   it('will not let a blind confirming read turn emptiness into a verdict', async () => {
     // Emptiness authorizes a kill, so it needs corroboration. A read that saw nothing at all
     // corroborates nothing — treating it as agreement is the step this module refuses.
diff --git a/src/main/daemon/daemon-live-pty-evidence.ts b/src/main/daemon/daemon-live-pty-evidence.ts
index 205d503bb1b..e77e47c2710 100644
--- a/src/main/daemon/daemon-live-pty-evidence.ts
+++ b/src/main/daemon/daemon-live-pty-evidence.ts
@@ -149,7 +149,15 @@ function collectLivePtyDescendants(rows: ProcessTableRow[], rootPid: number): Pr
 /** POSIX only; Windows abstains before this is reached. */
 async function probeOnce(
   daemonPid: number,
-  deps: DaemonPtyOwnershipDeps
+  deps: DaemonPtyOwnershipDeps,
+  /**
+   * The cached table may stand in for a slow read when the answer we are protecting is
+   * 'owns-live-ptys', but never when confirming emptiness: a confirmation drawn from the same
+   * TTL-cached snapshot as the sample it confirms is one observation counted twice, and the
+   * window it is meant to exclude — a login(1) wrapper whose shell has not appeared yet — is
+   * shorter than the cache. Denied there, a slow read answers 'unknown', which holds.
+   */
+  allowCachedFallback = true
 ): Promise<DaemonPtyOwnership> {
   const deadlineMs = deps.posixDeadlineMs ?? POSIX_OWNERSHIP_PROBE_DEADLINE_MS
   const deadline = Date.now() + deadlineMs
@@ -166,11 +174,13 @@ async function probeOnce(
     // few hundred milliseconds old still shows whether this daemon has children, and going
     // blind here gets them killed. It shares the attempt's budget rather than doubling it,
     // so an attempt costs what the launch budget was told it costs.
-    (await withDeadline(
-      (deps.readCachedPosixProcessTable ?? getProcessTableSnapshot)(),
-      remaining(),
-      null
-    ))
+    (allowCachedFallback
+      ? await withDeadline(
+          (deps.readCachedPosixProcessTable ?? getProcessTableSnapshot)(),
+          remaining(),
+          null
+        )
+      : null)
   // Why: a walk that never saw the root reports zero descendants for a process it
   // never examined. Only a root we actually observed can prove emptiness — and a read
   // that blew its deadline saw nothing at all.
@@ -206,7 +216,8 @@ export async function inspectDaemonPtyOwnership(
   for (let attempt = 0; attempt < PTY_OWNERSHIP_PROBE_ATTEMPTS; attempt++) {
     let sample: DaemonPtyOwnership
     try {
-      sample = await probeOnce(daemonPid, deps)
+      // The confirming read must be its own observation, so it is denied the cached table.
+      sample = await probeOnce(daemonPid, deps, !emptyAwaitingConfirmation)
     } catch {
       sample = 'unknown'
     }