test(terminal): pin the quarantine cap's accepted cost; correct its fresh-shell claims (STA-2373)

The idle-gap test claimed a fresh command is never eaten, but the idle gate
cannot release the first quarantined byte, so a command typed inside the cap is
dropped in full. Pin that real behavior and the reason the cap stays 5s.

A write rejection means the endpoint stopped accepting writes; the daemon
respawns only on a daemon-gone error, so a transient socket drop reconnects to
the same sessions. Stop asserting reattach always lands on a fresh shell.

src/renderer/src/components/terminal-pane/pty-connection.ts

@@ -3638,8 +3638,9 @@ export function connectPanePty(
       // cooldown window while typing. Local panes keep the lenient gate.
       requireAuthoritativeLiveness:
         Boolean(transport.getConnectionId?.()) || isRemoteRuntimePtyId(undeliverablePtyId),
-      // Why only the rejected path: a stalled-pipeline remount reattaches to the
+      // Why only the rejected path: it is the only one whose remount can land on
-      // same shell, so its half-typed line is still on screen and intact.
+      // a fresh shell. A stalled-pipeline remount always reattaches to the same
+      // shell, so its half-typed line is still on screen and intact.
       endpointReplaced: providerRejected
     })
   }

src/renderer/src/components/terminal-pane/terminal-input-quarantine.test.ts

@@ -56,7 +56,7 @@ describe('terminal input quarantine', () => {
     expect(isTerminalInputQuarantined(TAB)).toBe(false)
   })
 
-  it('releases on an idle gap so a fresh command is never eaten', () => {
+  it('releases on an idle gap once a quarantined byte has been seen', () => {
     armTerminalInputQuarantine(TAB, 0)
     // Burst tail, no Enter — the user was still mid-line when the endpoint died.
     expect(shouldDropQuarantinedTerminalInput(TAB, 'c', REATTACH_MS)).toBe(true)
@@ -66,6 +66,22 @@ describe('terminal input quarantine', () => {
     expect(isTerminalInputQuarantined(TAB)).toBe(false)
   })
 
+  // The accepted cost of the cap, pinned so it is not "fixed" by shortening it:
+  // over-quarantine is visible (nothing echoes) and recoverable by retyping,
+  // while under-quarantine executes half a command. A tail takes seconds to
+  // type, so a cap short enough to spare this command would cut that tail
+  // mid-line and deliver its remainder to the fresh shell.
+  it('eats a fresh command typed inside the cap, terminator included', () => {
+    armTerminalInputQuarantine(TAB, 0)
+    // No byte yet, so the idle gate cannot release the first keystroke however
+    // long the user waited; from there normal typing never opens a 700ms gap.
+    let at = 1_500
+    for (const char of 'ls -la\r') {
+      expect(shouldDropQuarantinedTerminalInput(TAB, char, at)).toBe(true)
+      at += 150
+    }
+  })
+
   it('does not let the idle gate fire on the re-attach delay itself', () => {
     armTerminalInputQuarantine(TAB, 0)
     // First quarantined byte arrives well past the idle window, because the

src/renderer/src/components/terminal-pane/terminal-pane-recovery.ts

@@ -38,10 +38,11 @@ type RecoveryRequest = {
    *  registry doesn't own, and treating null as "proceed" would let a
    *  disconnected remote pane churn reconnects on every cooldown window. */
   requireAuthoritativeLiveness?: boolean
-  /** The provider rejected the write because its endpoint died, so re-attach
+  /** The provider rejected the write because its endpoint stopped accepting
-   *  lands on a *fresh* shell. Only then is the in-flight line mangled, and
+   *  writes, so re-attach MAY land on a *fresh* shell (a respawn; a transient
-   *  only then may input be quarantined — a recovery that keeps the same live
+   *  socket drop reconnects to the same sessions). Only this path can mangle the
-   *  shell would have a legitimate command eaten. */
+   *  in-flight line, and only it may quarantine input — a recovery that always
+   *  keeps the same live shell would have a legitimate command eaten. */
   endpointReplaced?: boolean
 }