feat(context): role means message type, layer means pinning (#312/#313)

The system block is now for content that does not change during a run.
Anything the run mutates renders as a user message — both because a
mutating system prefix defeats prompt caching and because models
under-weight system-folded content against the trailing user turn.

PromptRenderer: role alone decides the message type; the old
`layer == L0 ||` clause is gone. That clause was silently overriding
role on four packs (InferenceSummarizer, SemanticReviewerImpl,
CapabilityGapReflectorImpl, Talkie session-naming) which are L0+USER
prompts — they were rendering as a system-only request with no user
turn at all.

Re-roled SYSTEM -> USER, all mutable within a run:
  recoveryTicket, remainingDelta, groundingFeedback, rejectionFeedback
  (trailing slot), plus verifiedBaseline, promotedConcept, claimedTask,
  clarificationAnswer, locked steeringNote, factSheet (inline, still
  L0-pinned).
Left SYSTEM (immutable): systemPrompt, operatingGuidance,
schemaInstruction, projectProfile, operatorProfile, agentInstructions,
successfulPlanShape.

Trailing-slot scarcity guard: repairMandateSourceTypes joined ALL
matches, so a recovery stage on a retry with an unmet delta would stack
three competing mandates. Now a precedence list emits exactly one
(recoveryTicket > retryFeedback > groundingFeedback > rejectionFeedback)
with remainingDelta appended as the completion signal.

ContextClassifier keys STATIC on layer alone — L0 means pinned/never
pruned regardless of role, so the re-roled L0 entries don't fall through
to FREEFORM and get token-pruned.

Also fixes a stale ContextFeedbackTest assertion (expected a CAS hash
the producer deliberately stopped emitting) and a stale initialIntent
doc comment claiming L0/SYSTEM where the code says L1/USER.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

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

@@ -26,7 +26,23 @@ object PromptRenderer {
     // than the original stage task. Instead they render as the FINAL user message, right after the
     // tool evidence, where a weak local model attends strongest and reads it as the next action.
     // Add a sourceType here (and set the entry's role to USER) to route it to that trailing slot.
-    private val repairMandateSourceTypes = setOf("retryFeedback")
+    //
+    // #312: highest precedence FIRST — at most ONE mandate renders per turn. The trailing slot works
+    // because it is scarce and authoritative; a recovery stage on a retry with an unmet delta would
+    // otherwise stack three competing "do this next" blocks and the channel becomes noise again.
+    private val repairMandatePrecedence = listOf(
+        "recoveryTicket",
+        "retryFeedback",
+        "groundingFeedback",
+        "rejectionFeedback",
+    )
+
+    // The remaining-delta checklist is not a competing mandate — it is the stage's completion
+    // signal ("what is left to make true") — so it appends after whichever mandate won rather
+    // than displacing it.
+    private const val COMPLETION_SIGNAL = "remainingDelta"
+
+    private val trailingSourceTypes = repairMandatePrecedence.toSet() + COMPLETION_SIGNAL
 
     // Tiebreak only: when entries carry no chronological ordinal (all 0 — e.g. router
     // chat, which assembles its pack directly), fall back to the old layer priority that
@@ -38,33 +54,48 @@ object PromptRenderer {
     }
 
     fun render(contextPack: ContextPack): List<ChatMessage> {
-        // Every SYSTEM-role entry folds into the single leading system message, whatever its
-        // layer (L0 additionally folds regardless of role). Strict chat templates (e.g. Qwen)
-        // reject any system message that is not the first message, so recalled memory, stage
-        // summaries, and retrieval entries must never render as standalone system turns.
+        // Every SYSTEM-role entry folds into the single leading system message, whatever its layer.
+        // Strict chat templates (e.g. Qwen) reject any system message that is not the first message,
+        // so recalled memory, stage summaries, and retrieval entries must never render as standalone
+        // system turns.
+        //
+        // #312: ROLE alone decides the message type — the old `layer == L0 ||` clause is gone. The
+        // system block is for content that does not change during a run (prompts, guidance, profiles);
+        // anything the run mutates (steering, gate verdicts, tickets, baselines, claimed task) is a
+        // user message, both because a mutating system prefix defeats prompt caching and because
+        // models under-weight system-folded content relative to the trailing user turn. Layer keeps
+        // its own job — budget tier and pin/prune eligibility (see ContextClassifier).
         val (systemEntries, conversationEntries) = contextPack.layers.entries
             .flatMap { (layer, entries) -> entries.map { layer to it } }
-            .partition { (layer, entry) -> layer == ContextLayer.L0 || entry.role == EntryRole.SYSTEM }
+            .partition { (_, entry) -> entry.role == EntryRole.SYSTEM }
         val systemContent = systemEntries
             .sortedWith(compareBy({ it.first.ordinal }, { it.second.ordinal }))
             .joinToString("\n\n") { it.second.content }
             .takeIf { it.isNotBlank() }
         // #293: pull repair mandates out of the inline flow — they render once, as the last turn.
         val (repairPairs, inlinePairs) = conversationEntries
-            .partition { it.second.sourceType in repairMandateSourceTypes }
+            .partition { it.second.sourceType in trailingSourceTypes }
         val conversationMessages = inlinePairs
             .sortedWith(compareBy({ it.second.ordinal }, { layerPriority(it.first) }))
             .map { (_, entry) -> entry.toChatMessage() }
-        val repairMandate = repairPairs
-            .sortedBy { it.second.ordinal }
-            .joinToString("\n\n") { it.second.content }
+        // Only the highest-precedence mandate present survives; the rest stay out of the prompt
+        // entirely (their content is still in the transcript/event log — this slot is not their
+        // only carrier). The completion signal appends after it.
+        val mandate = repairMandatePrecedence.firstNotNullOfOrNull { sourceType ->
+            repairPairs.contentOf(sourceType)
+        }
+        val repairMandate = listOfNotNull(mandate, repairPairs.contentOf(COMPLETION_SIGNAL))
+            .joinToString("\n\n")
             .takeIf { it.isNotBlank() }
-        // 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
+        // Repetition anchoring: steering directives render early (they are pinned standing context),
+        // 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
+        // user message at the end never trips strict system-must-be-first templates. Scans every
+        // entry, not just the system fold: since #312 a locked steering note is USER-role too, and
+        // both the locked and unlocked paths deserve the same anchor.
+        val anchor = (systemEntries + conversationEntries)
             .filter { it.second.sourceType == "steeringNote" }
+            .sortedBy { it.second.ordinal }
             .joinToString("\n") { it.second.content }
             .takeIf { it.isNotBlank() }
         val messages = buildList {
@@ -77,6 +108,12 @@ object PromptRenderer {
         return messages.ifEmpty { listOf(ChatMessage("user", "")) }
     }
 
+    private fun List<Pair<ContextLayer, ContextEntry>>.contentOf(sourceType: String): String? =
+        filter { it.second.sourceType == sourceType }
+            .sortedBy { it.second.ordinal }
+            .joinToString("\n\n") { it.second.content }
+            .takeIf { it.isNotBlank() }
+
     private fun ContextEntry.toChatMessage(): ChatMessage = ChatMessage(
         role = when (role) {
             EntryRole.SYSTEM -> "system"