shell: land the owned message delegate and the speech fixes

Applied 0009 then 0008 (0008's re-synthesize control calls a function 0009
adds). Casey applied and reloaded; verified against a running shell rather
than from a clean load.

0009 Speech.qml
  stop() could not stop. playProc ran `sh -c "mpv ... || ffplay ..."` — a
  compound command, so sh does not exec-replace itself and SIGTERM killed the
  wrapper while the player kept sounding as an orphan. Reproduced directly.
  That is the back-to-back TTS slam: every stop left audio playing and the
  next speak started a second player over it. Shell dropped; the pid held is
  now the pid making noise. ffplay fallback deleted rather than repaired —
  needing a fallback is what forced the wrapper that broke the kill.
  resynthesize() added, bypassing the cache: the button existed for "that came
  out wrong" and, being keyed on the same text, always replayed the identical
  file. synthesizing/playing split out of one `speaking` boolean.

0008 AiChat.qml
  One line: delegate AiMessage -> AgentMessage. The vendor block named 8 tools;
  the registry holds 19. The 11 it could not see were exactly the interiority
  surface — outfit, nickname, subagent, atmosphere, reach, consult, itinerary,
  todo, schedule, halt, intrusive.

  regenerate and edit dropped deliberately, per Casey: text regeneration is not
  possible (Ai.regenerate() already returned advice) and there is no in-place
  edit (the vendor wrote to a local array the server never sees). delete is
  armed and states that it hides locally only. Do not restore them.

Queue entries removed here, per the directory's own discipline.

surfaces/quickshell/modules/ii/sidebarLeft/AiChat.qml

@@ -6,6 +6,7 @@ import qs.modules.common.functions
 import qs.modules.souveraine.subconscious
 import qs.modules.ii.sidebarLeft.aiChat
 import qs.modules.souveraine.island
+import qs.modules.souveraine.agent
 import QtQuick
 import QtQuick.Controls
 import QtQuick.Layouts
@@ -446,7 +447,7 @@ Inline w/ backslash and round brackets \\(e^{i\\pi} + 1 = 0\\)
                         return message?.visibleToUser ?? true;
                     })
                 }
-                delegate: AiMessage {
+                delegate: AgentMessage {
                     required property var modelData
                     required property int index
                     messageIndex: index

surfaces/quickshell/patches/0008-agent-surface-owned-message-delegate.patch

@@ -1,41 +0,0 @@
-From f76c911a2f37352b0846b2ec70a96126ac275ee2 Mon Sep 17 00:00:00 2001
-From: Souveraine <souveraine@wiuf.net>
-Date: Wed, 12 Aug 2026 13:09:27 -0400
-Subject: [PATCH] shell: render the agent surface through the owned message
- delegate
-
-The sidebar's message list drew through ii-base's AiMessage, so every tool
-call was rendered by the vendor snapshot's ToolCallBlock, which knows 8 of
-the registry's 19 tools. The 11 it cannot name are exactly the interiority
-verbs. TASK-72 step 2.
-
-Requires modules/souveraine/agent/ composed (deploy.sh MANIFEST, already in
-tree). One line reverts it.
----
- surfaces/quickshell/modules/ii/sidebarLeft/AiChat.qml | 3 ++-
- 1 file changed, 2 insertions(+), 1 deletion(-)
-
-diff --git a/surfaces/quickshell/modules/ii/sidebarLeft/AiChat.qml b/surfaces/quickshell/modules/ii/sidebarLeft/AiChat.qml
-index bba12a1..0d6e04e 100644
---- a/surfaces/quickshell/modules/ii/sidebarLeft/AiChat.qml
-+++ b/surfaces/quickshell/modules/ii/sidebarLeft/AiChat.qml
-@@ -6,6 +6,7 @@ import qs.modules.common.functions
- import qs.modules.souveraine.subconscious
- import qs.modules.ii.sidebarLeft.aiChat
- import qs.modules.souveraine.island
-+import qs.modules.souveraine.agent
- import QtQuick
- import QtQuick.Controls
- import QtQuick.Layouts
-@@ -446,7 +447,7 @@ Inline w/ backslash and round brackets \\(e^{i\\pi} + 1 = 0\\)
-                         return message?.visibleToUser ?? true;
-                     })
-                 }
--                delegate: AiMessage {
-+                delegate: AgentMessage {
-                     required property var modelData
-                     required property int index
-                     messageIndex: index
--- 
-2.55.0
-

surfaces/quickshell/patches/0009-speech-stop-and-resynthesize.patch

@@ -1,136 +0,0 @@
-From bae112271e6a973bd9d178f68f232774751caf0e Mon Sep 17 00:00:00 2001
-From: Souveraine <souveraine@wiuf.net>
-Date: Wed, 12 Aug 2026 13:53:26 -0400
-Subject: [PATCH] shell: make stop actually stop, and re-synthesis actually
- re-synthesize
-MIME-Version: 1.0
-Content-Type: text/plain; charset=UTF-8
-Content-Transfer-Encoding: 8bit
-
-Three defects in the speech path, all of which present to the user as the
-same symptom — audio that piles up and cannot be called back.
-
-stop() could not reach the player. playProc ran ["sh", "-c", "mpv ... ||
-ffplay ..."], and a compound command means sh does not exec-replace itself:
-it forks the player and waits. Setting running = false SIGTERMs sh, sh dies,
-and the player keeps sounding as an orphan. Reproduced directly — the wrapper
-died and the child survived. Fixed by invoking mpv with no shell at all, so
-the pid quickshell holds is the pid making noise. The ffplay fallback is
-dropped rather than repaired: its own invocation was wrong, and needing a
-fallback is what forced the wrapper that broke the kill.
-
---keep-open=no --idle=no added because mpv can finish a stream, print
-(Paused) and never exit, which renders as speaking forever.
-
-resynthesize() added. speak() opens with a cache check keyed on the text, and
-re-synthesis is by definition the same text, so the re-synthesize control was
-guaranteed to hit cache and replay the identical broken audio. Invalidating
-_readyText before delegating is the fix.
-
-synthesizing/playing split out of speaking. Synthesis of a short line measured
-~11s against the service, and the shell showed the same state throughout as it
-shows while actually talking — which is why a press feels unacknowledged and
-gets pressed again. speaking is left untouched so nothing downstream shifts
-meaning.
----
- surfaces/quickshell/services/Speech.qml | 64 +++++++++++++++++++++++--
- 1 file changed, 60 insertions(+), 4 deletions(-)
-
-diff --git a/surfaces/quickshell/services/Speech.qml b/surfaces/quickshell/services/Speech.qml
-index 0655a5f..6e07eb8 100644
---- a/surfaces/quickshell/services/Speech.qml
-+++ b/surfaces/quickshell/services/Speech.qml
-@@ -33,7 +33,22 @@ Singleton {
-     id: root
- 
-     readonly property bool enabled: Config.options?.speech?.tts?.enable ?? false
-+
-+    // `speaking` is the OR that surfaces have always read — kept as-is so
-+    // nothing downstream changes meaning under them.
-     property bool speaking: synthProc.running || playProc.running
-+
-+    // But one boolean cannot distinguish "waiting on the synthesizer" from
-+    // "audio is coming out of the speaker", and those look nothing alike to a
-+    // person: synthesis of a short line measured ~11s against the service,
-+    // and during all of it the shell showed the same state it shows while
-+    // actually talking. That is the whole reason a press feels unacknowledged
-+    // and gets pressed again.
-+    //
-+    // Split, so a surface can show a spinner for one and a level for the
-+    // other, and so a stop button can say which thing it is about to stop.
-+    readonly property bool synthesizing: synthProc.running
-+    readonly property bool playing: playProc.running
-     property string lastError: ""
- 
-     property string _pendingText: ""
-@@ -131,6 +146,24 @@ Singleton {
-         retryTimer.running = false;
-     }
- 
-+    // resynthesize — request fresh audio for text we may already have cached.
-+    //
-+    // speak() opens with a cache check keyed on the text itself, and
-+    // re-synthesis is by definition the *same text* — so calling speak() to
-+    // "try again" is guaranteed to hit the cache and replay the identical
-+    // broken audio. The one control that exists for "that came out wrong"
-+    // could not do the only thing it is for.
-+    //
-+    // Invalidating _readyText before delegating is the whole fix: it forces
-+    // speak() down the synthesis path rather than the playback path.
-+    function resynthesize(text) {
-+        const t = String(text ?? "").trim();
-+        if (t.length === 0 || !root.enabled) return;
-+        stop();
-+        root._readyText = "";
-+        root.speak(t);
-+    }
-+
-     // The active agent's own voice, if she has one. The shell picks no voice
-     // of its own: souveraine owns the who→voice mapping, and that mapping is
-     // per-agent now — [_souveraine].voice_id rides on the public agent list.
-@@ -278,15 +311,38 @@ Singleton {
-     }
- 
-     // ── playback ─────────────────────────────────────────────────────────
-+    //
-+    // No `sh -c` wrapper, deliberately. The previous form was
-+    //
-+    //     ["sh", "-c", "mpv ... || ffplay ..."]
-+    //
-+    // and a compound command means sh does NOT exec-replace itself: it forks
-+    // the player as a child and waits. So `playProc.running = false` sends
-+    // SIGTERM to *sh*, sh dies, and the player keeps making noise as an
-+    // orphan. Reproduced directly: killing the wrapper left the child alive.
-+    //
-+    // That is why stop() never stopped anything, and why two speak() calls in
-+    // a row played over each other instead of replacing one another.
-+    //
-+    // One player, invoked directly, so the pid quickshell holds is the pid
-+    // making sound. The ffplay fallback is dropped rather than fixed: its
-+    // invocation was already wrong (raw input needs -i) and a fallback is
-+    // exactly what forced the shell wrapper that broke the kill. mpv is
-+    // present on both the laptop and the phone; if it is ever missing, the
-+    // honest outcome is a named error, not silent audio nobody can stop.
-+    //
-+    // --keep-open=no --idle=no is not cosmetic. mpv can reach the end of a
-+    // stream, print (Paused), and never exit — which, since `speaking` is
-+    // derived from playProc.running, renders as speaking forever.
-     Process {
-         id: playProc
--        command: ["sh", "-c",
--            `mpv --no-video --really-quiet '${root._outFile}' 2>/dev/null ` +
--            `|| ffplay -nodisp -autoexit -loglevel quiet '${root._outFile}'`]
-+        command: ["mpv", "--no-video", "--really-quiet",
-+                  "--keep-open=no", "--idle=no", root._outFile]
-         onExited: (exitCode) => {
-             // Exit 15 = SIGTERM from stop(); not a real failure — suppress.
-+            // This now actually reaches mpv rather than a shell wrapper.
-             if (exitCode !== 0 && exitCode !== 15) {
--                root.lastError = `audio playback failed (exit ${exitCode}) — mpv/ffplay present?`;
-+                root.lastError = `audio playback failed (exit ${exitCode}) — is mpv installed?`;
-                 console.log("[Speech]", root.lastError);
-             }
-         }
--- 
-2.55.0
-

surfaces/quickshell/services/Speech.qml

@@ -33,7 +33,22 @@ Singleton {
     id: root
 
     readonly property bool enabled: Config.options?.speech?.tts?.enable ?? false
+
+    // `speaking` is the OR that surfaces have always read — kept as-is so
+    // nothing downstream changes meaning under them.
     property bool speaking: synthProc.running || playProc.running
+
+    // But one boolean cannot distinguish "waiting on the synthesizer" from
+    // "audio is coming out of the speaker", and those look nothing alike to a
+    // person: synthesis of a short line measured ~11s against the service,
+    // and during all of it the shell showed the same state it shows while
+    // actually talking. That is the whole reason a press feels unacknowledged
+    // and gets pressed again.
+    //
+    // Split, so a surface can show a spinner for one and a level for the
+    // other, and so a stop button can say which thing it is about to stop.
+    readonly property bool synthesizing: synthProc.running
+    readonly property bool playing: playProc.running
     property string lastError: ""
 
     property string _pendingText: ""
@@ -131,6 +146,24 @@ Singleton {
         retryTimer.running = false;
     }
 
+    // resynthesize — request fresh audio for text we may already have cached.
+    //
+    // speak() opens with a cache check keyed on the text itself, and
+    // re-synthesis is by definition the *same text* — so calling speak() to
+    // "try again" is guaranteed to hit the cache and replay the identical
+    // broken audio. The one control that exists for "that came out wrong"
+    // could not do the only thing it is for.
+    //
+    // Invalidating _readyText before delegating is the whole fix: it forces
+    // speak() down the synthesis path rather than the playback path.
+    function resynthesize(text) {
+        const t = String(text ?? "").trim();
+        if (t.length === 0 || !root.enabled) return;
+        stop();
+        root._readyText = "";
+        root.speak(t);
+    }
+
     // The active agent's own voice, if she has one. The shell picks no voice
     // of its own: souveraine owns the who→voice mapping, and that mapping is
     // per-agent now — [_souveraine].voice_id rides on the public agent list.
@@ -278,15 +311,38 @@ Singleton {
     }
 
     // ── playback ─────────────────────────────────────────────────────────
+    //
+    // No `sh -c` wrapper, deliberately. The previous form was
+    //
+    //     ["sh", "-c", "mpv ... || ffplay ..."]
+    //
+    // and a compound command means sh does NOT exec-replace itself: it forks
+    // the player as a child and waits. So `playProc.running = false` sends
+    // SIGTERM to *sh*, sh dies, and the player keeps making noise as an
+    // orphan. Reproduced directly: killing the wrapper left the child alive.
+    //
+    // That is why stop() never stopped anything, and why two speak() calls in
+    // a row played over each other instead of replacing one another.
+    //
+    // One player, invoked directly, so the pid quickshell holds is the pid
+    // making sound. The ffplay fallback is dropped rather than fixed: its
+    // invocation was already wrong (raw input needs -i) and a fallback is
+    // exactly what forced the shell wrapper that broke the kill. mpv is
+    // present on both the laptop and the phone; if it is ever missing, the
+    // honest outcome is a named error, not silent audio nobody can stop.
+    //
+    // --keep-open=no --idle=no is not cosmetic. mpv can reach the end of a
+    // stream, print (Paused), and never exit — which, since `speaking` is
+    // derived from playProc.running, renders as speaking forever.
     Process {
         id: playProc
-        command: ["sh", "-c",
+        command: ["mpv", "--no-video", "--really-quiet",
-            `mpv --no-video --really-quiet '${root._outFile}' 2>/dev/null ` +
+                  "--keep-open=no", "--idle=no", root._outFile]
-            `|| ffplay -nodisp -autoexit -loglevel quiet '${root._outFile}'`]
         onExited: (exitCode) => {
             // Exit 15 = SIGTERM from stop(); not a real failure — suppress.
+            // This now actually reaches mpv rather than a shell wrapper.
             if (exitCode !== 0 && exitCode !== 15) {
-                root.lastError = `audio playback failed (exit ${exitCode}) — mpv/ffplay present?`;
+                root.lastError = `audio playback failed (exit ${exitCode}) — is mpv installed?`;
                 console.log("[Speech]", root.lastError);
             }
         }