feat(context,infra): compression pipeline stages 4-5 — token pruning + relevance + ToMe

- build() is now suspend: pipeline runs all stages in fixed order, each gated by level
- TOKEN_PRUNE (level 3): TokenPruner interface + LLMLingua-2 sidecar (sidecars/llmlingua)
  + HttpTokenPruner adapter (fails open if sidecar down); prunes freeform, preserves
  protected spans, skips tier-0 turns when TIER_SPLIT on
- TOME_MERGE (level 8): ToMeMerger collapses near-duplicate freeform turns (Jaccard)
- Stage 5 selection: RelevanceScorer + EmbeddingRelevanceScorer (cosine over Embedder);
  query-conditioned reorder so least-relevant freeform drops first under budget
- [orchestration] compression_level + token_pruner_url config, wired in Main
- suspend ripple fixed across builder callers/stubs

core/context/src/test/kotlin/CompressionPipelineStagesTest.kt

@@ -81,11 +81,11 @@ class CompressionPipelineStagesTest {
     }
 
     @Test
-    fun `builder pins a fact sheet at level 7 and omits it below`() {
+    fun `builder pins a fact sheet at level 7 and omits it below`() = kotlinx.coroutines.runBlocking {
         val entries = listOf(entry("chat", ContextLayer.L1, EntryRole.USER, "deploy to 10.0.0.1"))
         val compressor = com.correx.core.context.compression.DefaultContextCompressor()
 
-        fun buildAt(level: Int) = com.correx.core.context.builder.DefaultContextPackBuilder(
+        suspend fun buildAt(level: Int) = com.correx.core.context.builder.DefaultContextPackBuilder(
             compressor, CompressionPolicy(level)
         ).build(
             com.correx.core.events.types.ContextPackId("p"),
@@ -98,6 +98,43 @@ class CompressionPipelineStagesTest {
         val atSeven = buildAt(7).layers.values.flatten()
         assertTrue(atSeven.any { it.sourceType == "factSheet" && it.content.contains("10.0.0.1") })
         assertFalse(buildAt(6).layers.values.flatten().any { it.sourceType == "factSheet" })
+        Unit
+    }
+
+    @Test
+    fun `tome merges near-duplicate freeform turns keeping the newest`() {
+        val merger = com.correx.core.context.compression.ToMeMerger(threshold = 0.8)
+        val a = entry("chat", ContextLayer.L2, EntryRole.USER, "the deploy failed on host web").copy(ordinal = 0)
+        val b = entry("chat", ContextLayer.L2, EntryRole.USER, "the deploy failed on host web again").copy(ordinal = 1)
+        val c = entry("chat", ContextLayer.L2, EntryRole.USER, "totally unrelated content here").copy(ordinal = 2)
+        val out = merger.merge(listOf(a, b, c))
+        // a and b are near-duplicates -> only the newest (b) survives; c is distinct
+        assertEquals(listOf(1, 2), out.map { it.ordinal })
+    }
+
+    @Test
+    fun `token pruning at level 3 rewrites freeform preserving protected spans`() = kotlinx.coroutines.runBlocking {
+        val pruner = object : com.correx.core.context.compression.TokenPruner {
+            override suspend fun prune(content: String, protectedSpans: List<String>, targetRatio: Double): String =
+                "PRUNED " + protectedSpans.joinToString(" ")
+        }
+        val builder = com.correx.core.context.builder.DefaultContextPackBuilder(
+            com.correx.core.context.compression.DefaultContextCompressor(),
+            CompressionPolicy(3),
+            tokenPruner = pruner,
+        )
+        val entries = listOf(entry("chat", ContextLayer.L1, EntryRole.USER, "please redeploy 10.0.0.1 now"))
+        val pack = builder.build(
+            com.correx.core.events.types.ContextPackId("p"),
+            com.correx.core.events.types.SessionId("s"),
+            com.correx.core.events.types.StageId("st"),
+            entries,
+            com.correx.core.context.model.TokenBudget(limit = 4000),
+        )
+        val content = pack.layers.values.flatten().first { it.sourceType == "chat" }.content
+        assertTrue(content.startsWith("PRUNED"), content)
+        assertTrue(content.contains("10.0.0.1"), content)
+        Unit
     }
 
     @Test

core/context/src/test/kotlin/DecisionJournalPinningTest.kt

@@ -20,7 +20,7 @@ class DecisionJournalPinningTest {
     private val packId = ContextPackId("pack-1")
 
     @Test
-    fun `decisionJournal entry survives an under-budget build`() {
+    fun `decisionJournal entry survives an under-budget build`() = kotlinx.coroutines.runBlocking {
         val entries = listOf(
             ContextEntry(
                 id = ContextEntryId("journal-1"),
@@ -37,5 +37,6 @@ class DecisionJournalPinningTest {
             allRetained.any { it.sourceType == "decisionJournal" },
             "decisionJournal entry must be retained even when budget is exceeded"
         )
+        Unit
     }
 }