feat: wire mid-turn steering for interjections

Replace the legacy stop-and-respawn interjection model with mid-run
steering injection. User messages typed during an agent run are now
injected at turn boundaries as LoopMessage::User wrapped with
MID_TURN_STEER_WRAPPER, letting the model treat them as guidance
within the same run instead of spawning a new run.

Changes:
- provider: spawn_runner gains steering_queue parameter, wires it
  into LoopSpawnConfig (replaces dead 'stays None' comment)
- ui: interjection_queue converted from VecDeque to
  Arc<Mutex<VecDeque<String>>>, shared with spawn_runner at all
  call sites
- ui: removed interject signal send from mid-run push path —
  steering handles injection without stopping the run
- ui: updated comments and status messages to reflect 'inject at
  turn boundary' behavior
- acp: passes None (no UI, no steering needed)

Infrastructure built in prior commits (steering.rs factory, loop
polling at run.rs:238/588, LoopSpawnConfig.steering_queue field)
is now wired end-to-end.

Author: Yogthos

PLAN_LEARNING.md

@@ -34,6 +34,10 @@ Hermes has four layers that work together. We port all of them, adapted:
 └─────────────────────────────────────────────────┘
 ```
 
+VERY IMPORTANT: when making a plan you MUST examine current features such as compression and memory,
+and make sure they are consolidated with the plan, do not create dead code or duplicated functionality,
+old functionality must be either integrated or replaced
+
 ### Per-Project Storage Layout
 
 ```

src/extras/acp/mod.rs

@@ -197,7 +197,7 @@ async fn run_prompt(
     )
     .await;
 
-    let runner = agent.spawn_runner(prompt_text.to_string(), vec![]);
+    let runner = agent.spawn_runner(prompt_text.to_string(), vec![], None);
     let mut rx = runner.event_rx;
 
     // F5: correlate rig tool-call ids with ACP ids so parallel

src/provider/mod.rs

@@ -511,7 +511,7 @@ impl AnyAgent {
         // event channel collecting text.
         let runner = self
             .clone()
-            .spawn_runner(effective_prompt.clone(), Vec::new());
+            .spawn_runner(effective_prompt.clone(), Vec::new(), None);
         let task = runner.task;
         let mut event_rx = runner.event_rx;
 
@@ -651,7 +651,12 @@ impl AnyAgent {
         }
     }
 
-    pub fn spawn_runner(self, prompt: String, history: Vec<Message>) -> AgentRunner {
+    pub fn spawn_runner(
+        self,
+        prompt: String,
+        history: Vec<Message>,
+        steering_queue: Option<std::sync::Arc<std::sync::Mutex<std::collections::VecDeque<String>>>>,
+    ) -> AgentRunner {
         use crate::agent::agent_loop::{
             LoopSpawnConfig, loop_tool_to_rig_definition, retrying_stream_fn,
             rig_history_system_prompt, rig_history_to_loop_messages, spawn_loop_runner,
@@ -703,16 +708,11 @@ impl AnyAgent {
         } else {
             Some(self.model_name.clone())
         };
+        cfg.steering_queue = steering_queue;
         #[cfg(feature = "plugin")]
         {
             cfg.plugin_mgr = crate::plugin::hook::global();
         }
-        // steering_queue stays None for h-6 — UI's existing
-        // interjection_queue isn't shared yet. interject_tx →
-        // signal.cancel() handles the legacy "stop the run"
-        // case via the LoopRunner::into_agent_runner adapter.
-        // A follow-up UI commit will share the queue so the
-        // model observes interjections mid-run.
 
         let loop_runner = spawn_loop_runner(cfg);
         loop_runner.into_agent_runner()