feat(context): compression pipeline stage 1 — level policy + deterministic stages

Additive CompressionPolicy (levels 1-9), ProtectedSpanTagger (regex
load-bearing spans), FormatCompressor (json→dotted lines), ContextClassifier
(static/structured/freeform). Stage 1 (FORMAT_COMPRESS) wired into
DefaultContextPackBuilder behind the policy; structured entries only, lossless.
PromptRenderer repetition-anchoring of steering directives.

core/inference/src/main/kotlin/com/correx/core/inference/PromptRenderer.kt

@@ -37,9 +37,18 @@ object PromptRenderer {
         val conversationMessages = conversationEntries
             .sortedWith(compareBy({ it.second.ordinal }, { layerPriority(it.first) }))
             .map { (_, entry) -> entry.toChatMessage() }
+        // Repetition anchoring: steering directives fold into the leading system message, far
+        // from the final query — weak local models forget them (lost-in-the-middle). Restate
+        // them once as a trailing user turn, where models attend strongest. Template-safe: a
+        // user message at the end never trips strict system-must-be-first templates.
+        val anchor = systemEntries
+            .filter { it.second.sourceType == "steeringNote" }
+            .joinToString("\n") { it.second.content }
+            .takeIf { it.isNotBlank() }
         val messages = buildList {
             systemContent?.let { add(ChatMessage("system", it)) }
             addAll(conversationMessages)
+            anchor?.let { add(ChatMessage("user", "Reminder — active steering directive(s):\n$it")) }
         }
         return messages.ifEmpty { listOf(ChatMessage("user", "")) }
     }