M2 audit: excise managed connectors and xAI media-gen tools

Managed connectors (grok.com MCP admin) removed root-and-branch:
- The managed-MCP fetch/injection pipeline is gone, including the whole
  kigi-shell-session-support crate (managed-config fetch client, gateway
  tool catalog + dispatch, header injection, refresh task), reactive
  managed re-auth, mcp_doctor's grok.com-source discovery, and the
  [managed_mcps] config surface.
- TUI: the 'Managed by grok.com' section, connectors URL/deep-link,
  Action::OpenManagedConnectors, and session_team_id are gone. Local MCP
  management (list/toggle/add/remove/auth/tools) is fully intact.
- Kept as LOCAL policy: managed-settings.json MCP allow/deny enforcement,
  the multi-source local MCP merge, folder-trust gating. PluginOrigin
  Project/User labels kept (they tag locally discovered plugin dirs).

imagine/media-gen tools (xAI image/video generation) removed:
- image_gen, image_edit, video_gen, image_to_video, reference_to_video
  implementations, registrations, ToolKind/ToolInput/Output variants
  (serde-safe), config plumbing end to end, ZDR video machinery,
  /imagine + /imagine-video commands and guidance text, the bundled
  imagine skill (added to legacy cleanup so user installs delete it),
  and the media-gen render path.
- Kept: image INPUT (paste/attach, [Image #N] meta, pdf/image fetch,
  clipboard wrap), generic media-ref rendering, and the generic tool
  401-retry machinery (tests renamed, assertions unweakened).
- deploy_app stays: it is a permanently-disabled local stub deploying
  nowhere.

121 files changed, 8 deleted. Gates: workspace check/clippy 0/0, fmt,
deny ok; suites green (tools 2554, shell 4862, tui 6608, workspace
1042). Remaining grok.com strings live only in the auth-method ids and
changelog archives (§9/M3 sweep).

crates/codegen/kigi-tui/src/app/agent.rs

@@ -100,7 +100,7 @@ impl QueuedPrompt {
     /// `true` for plain rows (no `wire_blocks`) and for raw skill slash rows
     /// (`/find-session args` — a single Text block equal to `text`, expanded
     /// shell-side at delivery), so interjecting `text` loses nothing. `false`
-    /// when the payload was expanded client-side (`/imagine`, `/loop`):
+    /// when the payload was expanded client-side (`/loop`):
     /// interjecting those by `text` would drop the expansion, and by payload
     /// would render the raw instruction.
     pub fn wire_matches_display(&self) -> bool {
@@ -1272,7 +1272,7 @@ mod tests {
     }
     /// `wire_matches_display` splits interjectable rows (no payload, or a raw
     /// skill slash payload equal to the display text) from client-expanded
-    /// payloads (`/imagine`, `/loop`) that must run as their own turn.
+    /// payloads (`/loop`) that must run as their own turn.
     #[test]
     fn wire_matches_display_classifies_payload_shapes() {
         let text_block = |t: &str| acp::ContentBlock::Text(acp::TextContent::new(t.to_string()));
@@ -1285,7 +1285,7 @@ mod tests {
         assert!(raw_skill.wire_matches_display(), "raw slash payload");
         let expanded = QueuedPrompt {
             wire_blocks: Some(vec![text_block("<skill>body</skill>")]),
-            ..QueuedPrompt::plain(3, "/imagine cat", QueueEntryKind::Prompt)
+            ..QueuedPrompt::plain(3, "/loop cat", QueueEntryKind::Prompt)
         };
         assert!(!expanded.wire_matches_display(), "expanded payload");
         let multi_block = QueuedPrompt {