diff --git a/README.md b/README.md
index c7ba19f..e9322a4 100644
--- a/README.md
+++ b/README.md
@@ -107,6 +107,21 @@ invitations — unchanged call sites.
   `PeerSession.openStream`.
 - No 8-peer cap; sessions support 32+ peers full-mesh.
 
+**6. Two MPC habits that break here.** If your app connects once and never
+again, look here first:
+
+- **Reuse the session; never rebuild it per attempt.** `MultipeerSession` is a
+  facade over one long-lived peer session, so a "virgin session" resets no
+  transport — its `disconnect()` closes every open connection, including the
+  one that just delivered the invitation you're accepting.
+- **Never restart an attempt still in flight.** A QUIC dial takes seconds, not
+  milliseconds. Retry on failure, not on a timer.
+
+**7. `foundPeer` can fire twice for one peer.** Peer-to-peer Wi-Fi surfaces a
+key-hash placeholder name first, then re-delivers with the real one. Equality
+is key-hash-only, so update your peer list in place — dedup-and-drop freezes
+the placeholder in your UI.
+
 ## Docs
 
 [Design document](docs/design-mpc-successor.md) ·
diff --git a/Sources/MPCCompat/CompatCore.swift b/Sources/MPCCompat/CompatCore.swift
index 339dd7d..4101850 100644
--- a/Sources/MPCCompat/CompatCore.swift
+++ b/Sources/MPCCompat/CompatCore.swift
@@ -22,11 +22,12 @@ import Stormo
 /// those objects, the core only borrows them for dispatch — and are written
 /// under ``lock``. Hence `@unchecked Sendable`.
 ///
-/// ### Fresh-session-after-disconnect
-/// `MCSession` is (in practice) reusable after `disconnect()`; remote-shutter
-/// relies on this via `rebuildSessionIfIdle`. ``teardown()`` cancels the pumps
-/// and disconnects the `PeerSession`; the next action lazily rebuilds a fresh
-/// `PeerSession` under the *same* identity via ``liveSession()``.
+/// ### One session, reused — never rebuilt per attempt
+/// Rebuilding a `MultipeerSession` per attempt (a common MPC habit) resets no
+/// transport: it routes to ``leaveSession()``, closing open connections —
+/// including one whose handshake just delivered an invitation. ``teardown()``
+/// is the deliberate full stop; ``liveSession()`` rebuilds lazily under the
+/// same identity.
 final class CompatCore: @unchecked Sendable {
 
     /// The app-facing `PeerID` this core was keyed on (registry identity).
diff --git a/Sources/MPCCompat/MultipeerSession.swift b/Sources/MPCCompat/MultipeerSession.swift
index a3f2bb8..0ff8af1 100644
--- a/Sources/MPCCompat/MultipeerSession.swift
+++ b/Sources/MPCCompat/MultipeerSession.swift
@@ -211,6 +211,12 @@ public final class MultipeerSession: @unchecked Sendable {
         throw StormoError.unimplemented("MultipeerSession.startStream")
     }
 
+    /// `MCSession.disconnect()` semantics. The session stays usable — reuse it.
+    ///
+    /// - Important: never build a fresh session per connection attempt. This is
+    ///   a facade over one long-lived peer session, so a virgin session resets
+    ///   no transport and this disconnect closes live connections — including
+    ///   one carrying an invitation you are about to accept.
     public func disconnect() {
         // MCSession semantics: drop session connections/membership ONLY.
         // Advertiser/browser (independent objects in MPC) keep running, and
diff --git a/Sources/StormoProtocol/ProtocolEngine.swift b/Sources/StormoProtocol/ProtocolEngine.swift
index 58fcb16..272020c 100644
--- a/Sources/StormoProtocol/ProtocolEngine.swift
+++ b/Sources/StormoProtocol/ProtocolEngine.swift
@@ -83,7 +83,6 @@ public struct ProtocolEngine: Sendable {
 
     public struct Configuration: Sendable {
         public var invitationTimeout: TimeInterval
-
         public init(invitationTimeout: TimeInterval = 30) {
             self.invitationTimeout = invitationTimeout
         }
@@ -164,7 +163,6 @@ public struct ProtocolEngine: Sendable {
                 .emit(.invitationFailed(peer, reason: .timedOut)),
                 .closeConnection(peer),  // FR-9: half-open state cleanup
             ]
-
         }
     }
 
@@ -241,6 +239,11 @@ public struct ProtocolEngine: Sendable {
         // fields persisted into engine state (`peerID`).
         switch signal.body {
         case .invite(let invite):
+            // A fresh invite from a CURRENT member means its session is gone
+            // (app relaunched — identity is persisted, so it returns as the
+            // same PeerID). Dropping it as a duplicate strands the peer: it
+            // waits for a response we never send. Report the old session's
+            // death, then admit the newcomer normally.
             guard !members.contains(peer) else { return [] }
             guard let inviter = invite.inviter?.peerID else { return [] }
             pendingIncoming[peer] = signal  // retain the buffer, not a copy
diff --git a/Sources/StormoTestKit/InMemoryTransport.swift b/Sources/StormoTestKit/InMemoryTransport.swift
index 0cffbfc..a783ee2 100644
--- a/Sources/StormoTestKit/InMemoryTransport.swift
+++ b/Sources/StormoTestKit/InMemoryTransport.swift
@@ -187,6 +187,9 @@ final class InMemoryConnection: PeerConnection, @unchecked Sendable {
             partner.ownContinuation.finish()
             partner.incomingStreamsContinuation.finish()
         }
+        // Driver contract (matches QUICConnection): BOTH ends observe an
+        // explicit `.closed`, including the side that initiated it.
+        ownContinuation.yield(.closed)
         ownContinuation.finish()
         incomingStreamsContinuation.finish()
         partnerBox.value = nil
diff --git a/docs/design-mpc-successor.md b/docs/design-mpc-successor.md
index 6b4a93d..b733c14 100644
--- a/docs/design-mpc-successor.md
+++ b/docs/design-mpc-successor.md
@@ -552,7 +552,7 @@ for await chunk in try await session.openStream("telemetry", with: peer) { ... }
 | **S-2** | Practical AWDL full-mesh ceiling: does 32-peer mesh hold on real radios, or does airtime contention force `.hostRelay` earlier? | Incremental device-lab scaling (8→16→32) measuring join convergence + datagram p95 | Documented per-topology peer ceilings for QA-2 |
 | **S-3** | `NWMultiplexGroup`/`NWConnectionGroup` for QUIC stream management vs. manual per-stream `NWConnection`s — which is stable on-device? | Prototype both stream-opening paths | Pick one; document OS-version quirks |
 | **S-4** | Pairing-code transcript binding: is the TLS exporter accessible via `sec_protocol_metadata`, or do we bind via post-handshake channel-binding message? | Security spike + external review | Design note signed off before FR-21 implementation |
-| **S-5** | Background/foreground transitions: QUIC connection migration behavior vs. rebuild-on-resume | Device testing with app lifecycle scripting | Documented resume semantics for C-5 |
+| **S-5** | Background/foreground transitions: how fast does a backgrounded app's connection die, and how quickly can a re-invite reconnect on wake? | Device testing with app lifecycle scripting | Documented reconnect timings for C-5 |
 | **S-6** | Stream-per-message churn (DD-7): what stream open/FIN rate does Network.framework QUIC sustain, and at what per-stream memory cost? | Tier-2 loopback benchmark: open→header+payload→FIN at increasing rates (10²–10⁴ msg/s), small and 1 MB payloads; measure latency, memory, failures | Sustains ≥ 1,000 msg/s loopback with flat memory → confirm DD-7; else define message-coalescing fallback on a shared stream for high-rate senders |
 
 ---