fix: keep quarantined cleanup out of automatic recovery projection

Also record the successful two-turn grounding qualification and semantic review.

Co-Authored-By: Paperclip <noreply@paperclip.ing>

server/src/services/execution-projection.test.ts

@@ -44,6 +44,9 @@ describe("execution truth projection", () => {
       permittedActions: ["inspect_run", "inspect_recovery"],
     });
     expect(projectExecution(stopped, coordinator({ phase: "retryable_failure" }), [], undefined, now).phase).toBe("recovery_needed");
+    expect(projectExecution({ ...stopped, finishedAt: now }, coordinator({
+      phase: "terminal_failure", failureCode: "native_provider_terminal_failed",
+    }), [], undefined, now)).toMatchObject({ phase: "recovery_needed", recoveryOwner: "board" });
     expect(projectExecution({ ...stopped, errorCode: "provider_frame_too_large" }, undefined, [], undefined, now).phase).toBe("failed");
     expect(projectExecution(stopped, undefined, [], { status: "resolved", cause: "native_session_cleanup_quarantined", nextAction: "Verify prior execution",
       evidence: { executionReconciliation: { providerStopped: true }, continuationRunId: "successor" } }, now))

server/src/services/execution-projection.ts

@@ -260,6 +260,7 @@ export function projectExecution(
   }
   if (coordinator?.phase === "terminal_failure" || recoveryAction || cleanupQuarantined) {
     if (
+      !cleanupQuarantined &&
       coordinator?.failureCode === "native_provider_terminal_failed" &&
       !detail.replacementDenied &&
       run.finishedAt &&

tests/runner-e2e/QUALIFICATION-2026-09-21.md

@@ -88,3 +88,30 @@ Two user-facing limitations remain worth deciding:
 
 Production instructions were preserved. These observations must not be hidden by
 changing prompts solely to make the benchmark green.
+
+## Grounded status answers: corrected live proof
+
+[Campaign 35658262695](https://github.com/paperclipai/paperclip/actions/runs/35658262695)
+on `4a26f10be7dd6aeabf2ca7b44b3d6b2ece7817e0`: **2/2 passed**, both
+cleanup passes. Each provider answered both turns, preserved both source tasks,
+and started no execution on either task. The original failed attempts above are
+retained; this is a new campaign with an explicit output contract.
+
+Semantic review of all four retained replies against the task descriptions and
+chronological comments found:
+
+- Both identified venue confirmation as the current blocker, printing as deferred,
+  no task execution, and unknown attendance. Both gave the useful next step of
+  confirming the venue before printing.
+- Both rejected the obsolete budget claim and unsupported printing claim on the
+  follow-up, and distinguished a planned Friday from a guaranteed calendar date.
+  Neither invented a venue, date, or attendance count.
+- Codex's explanations were compact and clear. Claude's second explanation was
+  longer but readable and grounded in named task records. Its phrase “no venue has
+  been confirmed” is slightly stronger than “no confirmation is recorded”; its
+  immediately following quotation and null fact make the evidence limitation clear.
+
+This qualifies these two-turn grounding stories, not general answer quality,
+statistical reliability, multilingual behavior, or arbitrary long conversations.
+The five review dimensions remain factual grounding, stale-premise correction,
+honest uncertainty, useful next step, and clear prose. No production prompt changed.