fix(inference): tolerate a briefly-absent provider instead of killing the session (#299)

A transient provider connection drop (e.g. OOM-killed llama.cpp mid-request)
collapsed into a hard NoEligibleProviderException that escaped runInference's
try/catch (the route() call sat outside it), propagated past the stage retry
loop, and landed in ServerModule's session-level catch — failing the WHOLE
session even though the failure was retryable and the provider recovered
seconds later.

- SessionOrchestrator.runInference: wrap router.route() in try/catch so
  routing failures become InferenceResult.Failed and flow through the normal
  retryable/backoff/exhaustion machinery instead of escaping.
- DefaultInferenceRouter.route: distinguish "capability never configured on
  any provider" (fail fast, waiting can't help) from "configured but
  currently unhealthy" (bounded wait/backoff — default 3 attempts x 2s —
  re-checking health before declaring NoEligibleProvider terminal).

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GeyGFXczJb8RUWGBKmkm6G

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

@@ -2,6 +2,7 @@ package com.correx.core.inference
 
 import com.correx.core.events.types.ProviderId
 import com.correx.core.events.types.StageId
+import kotlinx.coroutines.delay
 import kotlinx.coroutines.sync.Mutex
 import kotlinx.coroutines.sync.withLock
 import kotlin.time.Duration
@@ -21,6 +22,12 @@ class DefaultInferenceRouter(
     private val strategy: RoutingStrategy,
     private val cacheTtl: Duration = 5.seconds,
     private val timeSource: TimeSource = TimeSource.Monotonic,
+    // A provider that briefly drops (crash + qa-stack restart) shouldn't collapse into a hard
+    // NoEligibleProvider abort — give it bounded time to come back before giving up. This only
+    // applies when the capability IS configured on some provider but that provider is currently
+    // unhealthy; a capability nobody ever declared fails immediately (see routeCapabilityCandidates).
+    private val unavailableRetryAttempts: Int = 3,
+    private val unavailableRetryDelay: Duration = 2.seconds,
 ) : InferenceRouter {
 
     private val cache = mutableMapOf<ProviderId, HealthEntry>()
@@ -45,12 +52,39 @@ class DefaultInferenceRouter(
         }
     }
 
+    private suspend fun refreshedHealth(provider: InferenceProvider): ProviderHealth =
+        lockFor(provider.id).withLock {
+            val fresh = provider.healthCheck()
+            mapMutex.withLock { cache[provider.id] = HealthEntry(fresh, timeSource.markNow()) }
+            fresh
+        }
+
     override suspend fun route(stageId: StageId, requiredCapabilities: Set<ModelCapability>): InferenceProvider {
         val candidates = requiredCapabilities
             .flatMap { registry.resolve(it) }
             .distinctBy { it.id }
             .ifEmpty { registry.listAll() }
-        val healthy = candidates.filter { cachedHealth(it) !is ProviderHealth.Unavailable }
+
+        // Nobody was ever configured with this capability set — no amount of waiting fixes that,
+        // fail fast instead of burning the bounded-wait budget below.
+        if (requiredCapabilities.isNotEmpty() &&
+            candidates.none { it.capabilities().map { c -> c.capability }.toSet().containsAll(requiredCapabilities) }
+        ) {
+            throw NoEligibleProviderException(stageId, requiredCapabilities)
+        }
+
+        var healthy = candidates.filter { cachedHealth(it) !is ProviderHealth.Unavailable }
+        var attempt = 0
+        while (healthy.isEmpty() && attempt < unavailableRetryAttempts) {
+            attempt++
+            log.warn(
+                "route: capability {} configured but all candidates unhealthy for stage={}" +
+                    " — waiting {} (attempt {}/{}) before declaring NoEligibleProvider",
+                requiredCapabilities, stageId.value, unavailableRetryDelay, attempt, unavailableRetryAttempts,
+            )
+            delay(unavailableRetryDelay)
+            healthy = candidates.filter { refreshedHealth(it) !is ProviderHealth.Unavailable }
+        }
         val selected = strategy.select(healthy, requiredCapabilities)
         // Post-selection re-check closes the TOCTOU window between initial filter and dispatch.
         when (val postHealth = selected.healthCheck()) {

core/kernel/src/main/kotlin/com/correx/core/kernel/orchestration/SessionOrchestrator.kt

@@ -300,7 +300,15 @@ abstract class SessionOrchestrator(
         // ToolCalling score and routes the stage to the best tool-caller.
         val requiredCapabilities = stageConfig.requiredCapabilities +
             if (withTools && stageConfig.allowedTools.isNotEmpty()) setOf(ModelCapability.ToolCalling) else emptySet()
-        val provider = inferenceRouter.route(stageId, requiredCapabilities, stageConfig.modelId)
+        // Routing itself can fail transiently (provider mid-crash-recovery) — it must be retryable
+        // like any other inference failure, not escape and kill the whole session (see #299).
+        val provider = try {
+            inferenceRouter.route(stageId, requiredCapabilities, stageConfig.modelId)
+        } catch (e: CancellationException) {
+            throw e
+        } catch (e: Exception) {
+            return InferenceResult.Failed(e.message ?: "routing failed")
+        }
         log.debug(
             "[Orchestrator] inference session={} stage={} provider={} timeoutMs={}",
             sessionId.value, stageId.value, provider.id.value, timeoutMs,