vLLM priority scheduling for agents

Thread request priority through the API call chain to vLLM's
priority scheduler. Lower value = higher priority, with preemption.

Priority is set per-agent in the .agent header:
- interactive (runner): 0 (default, highest)
- surface-observe: 1 (near-realtime, watches conversation)
- all other agents: 10 (batch, default if not specified)

Requires vLLM started with --scheduling-policy priority.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>

src/subconscious/defs.rs

@@ -46,6 +46,9 @@ pub struct AgentDef {
     pub chunk_size: Option<usize>,
     pub chunk_overlap: Option<usize>,
     pub temperature: Option<f32>,
+    /// vLLM scheduling priority (lower = higher priority).
+    /// 0 = interactive, 1 = near-realtime, 10 = batch (default).
+    pub priority: i32,
     /// Bail check command — run between steps with pid file path as $1,
     /// cwd = state dir. Non-zero exit = stop the pipeline.
     pub bail: Option<String>,
@@ -75,6 +78,9 @@ struct AgentHeader {
     /// LLM temperature override
     #[serde(default)]
     temperature: Option<f32>,
+    /// vLLM scheduling priority (lower = higher priority, default 10 = batch)
+    #[serde(default = "default_priority")]
+    priority: i32,
     /// Bail check command — run between steps with pid file path as $1,
     /// cwd = state dir. Non-zero exit = stop the pipeline.
     #[serde(default)]
@@ -82,6 +88,7 @@ struct AgentHeader {
 }
 
 fn default_model() -> String { "sonnet".into() }
+fn default_priority() -> i32 { 10 }
 
 /// Parse an agent file: first line is JSON config, rest is the prompt(s).
 /// Multiple prompts are separated by `=== PROMPT [phase:name] ===` lines.
@@ -149,6 +156,7 @@ fn parse_agent_file(content: &str) -> Option<AgentDef> {
         chunk_size: header.chunk_size,
         chunk_overlap: header.chunk_overlap,
         temperature: header.temperature,
+        priority: header.priority,
         bail: header.bail,
     })
 }