refactor(config): move embedder/l3 to top-level [embedder]/[l3] sections

They are shared infra (the orchestrator uses them for workflow repo-knowledge,
not just the router chat layer), so [router.embedder]/[router.l3] was misleading.
Loader/writer use the new names; legacy [router.*] still read as a fallback.
Also notes the STEERING LLM-reformulation cost as a ponytail follow-up.

core/config/src/main/kotlin/com/correx/core/config/ConfigLoader.kt

@@ -419,8 +419,11 @@ object ConfigLoader {
         val toolsFileWriteSection = sections["tools.file_write"] ?: emptyMap()
         val toolsFileEditSection = sections["tools.file_edit"] ?: emptyMap()
         val routerSection = sections["router"] ?: emptyMap()
-        val routerEmbedderSection = sections["router.embedder"] ?: emptyMap()
-        val routerL3Section = sections["router.l3"] ?: emptyMap()
+        // Embedder + L3 are shared infra (the orchestrator uses them for workflow repo-knowledge,
+        // not just the router chat layer). Prefer the top-level [embedder]/[l3] sections; fall back
+        // to the legacy [router.embedder]/[router.l3] so existing configs keep working.
+        val routerEmbedderSection = sections["embedder"] ?: sections["router.embedder"] ?: emptyMap()
+        val routerL3Section = sections["l3"] ?: sections["router.l3"] ?: emptyMap()
         val routerGenerationSection = sections["router.generation"] ?: emptyMap()
         val routerNarrationSection = sections["router.narration"] ?: emptyMap()
         val modelsSection = sections["models"] ?: emptyMap()

core/config/src/main/kotlin/com/correx/core/config/CorrexConfigWriter.kt

@@ -56,13 +56,13 @@ object CorrexConfigWriter {
         b.kv("retrieval_k", cfg.router.retrievalK)
         b.kv("token_budget", cfg.router.tokenBudget)
 
-        b.section("router.embedder")
+        b.section("embedder")
         b.kv("backend", str(cfg.router.embedder.backend))
         b.kv("dimension", cfg.router.embedder.dimension)
         cfg.router.embedder.url?.let { b.kv("url", str(it)) }
         cfg.router.embedder.modelId?.let { b.kv("model_id", str(it)) }
 
-        b.section("router.l3")
+        b.section("l3")
         b.kv("backend", str(cfg.router.l3.backend))
         cfg.router.l3.persistPath?.let { b.kv("persist_path", str(it)) }
         b.kv("python_executable", str(cfg.router.l3.pythonExecutable))

core/router/src/main/kotlin/com/correx/core/router/RouterFacade.kt

@@ -186,6 +186,11 @@ class DefaultRouterFacade(
             emitIdeasCaptured(sessionId, ideas.ideas)
         }
 
+        // ponytail: STEERING launders the user's text through the router LLM (the `content` above)
+        // before injecting it as a note. For a clear instruction that's an extra inference that can
+        // distort intent; the reformulation only earns its cost when the input is terse/context-
+        // dependent. Upgrade path: inject the raw (validated) input directly and skip the rewrite
+        // unless a heuristic flags the message as too short/ambiguous to stand alone.
         val steeringEmitted = mode == ChatMode.STEERING && rawContent.isNotBlank()
         if (steeringEmitted) {
             val validationError = validateSteering?.invoke(content)

docs/sample-config.toml

@@ -107,15 +107,18 @@ top_p = 0.9
 max_tokens = 1024
 max_per_run = 100
 
-[router.embedder]
+# Embedder + L3 vector memory. Shared infra: the orchestrator uses these for workflow
+# repo-knowledge retrieval, not just the router chat layer — hence top-level, not [router.*].
+# (Legacy [router.embedder] / [router.l3] are still read as a fallback.)
+[embedder]
 backend = "noop"               # or "llamacpp"
 dimension = 1536
 # url = "http://127.0.0.1:11000"
 # model_id = "nomic-embed-text"
 
-[router.l3]
+[l3]
 backend = "in_memory"          # or "turbovec"
-# persist_path = "~/.config/correx/router/l3/index.tq"
+# persist_path = "~/.config/correx/l3/index.tq"
 # python_executable = "python3"
 # script_path = "~/.config/correx/python/turbovec_sidecar.py"
 # dim = 1536