test(e2e): read paired emphasis before review outcome checks

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

tests/runner-e2e/chat-catalog-descriptors.json

@@ -830,7 +830,7 @@
       "../../ui/src/components/task-chat/TaskChatComposer.tsx",
       "chat-scenarios/app-permissions.yaml"
     ],
-    "oracleVersion": "paperclip.chat-catalog.app-review.v5",
+    "oracleVersion": "paperclip.chat-catalog.app-review.v6",
     "expectedRunCount": 6,
     "expectedHumanMessageCount": 3,
     "minimumRunCount": 6,

tests/runner-e2e/chat-catalog-domains/external-app-review-outcome-language.test.ts

@@ -101,3 +101,25 @@ it.each([
 ])("rejects a mixed or nested external exception: %s", reply => {
   expect(reviewOutcomeReply(reply, "rejected")).toBe(false);
 });
+
+it.each(["**not**", "__not__", "*not*", "_not_", "***not***", "___not___"])(
+  "reads paired emphasis around negation as visible prose: %s", negation => {
+    expect(reviewOutcomeReply(`You declined. The update was ${negation} applied.`, "rejected")).toBe(true);
+  },
+);
+it.each([
+  "You declined. I **applied** the update.",
+  "You declined. The update was __applied__.",
+  "You declined. I *saved* the update in the notes app.",
+  "You declined. The update was **not** applied, but I **saved** it to the notes app.",
+  "You declined. Nothing was **saved**, but I **applied** the update.",
+  "You declined. The draft was saved **in the notes app** for your review.",
+])("paired emphasis does not hide external completion: %s", reply => {
+  expect(reviewOutcomeReply(reply, "rejected")).toBe(false);
+});
+it.each([
+  "You declined. The update was not** applied.",
+  String.raw`You declined. The update was \**not** applied.`,
+])("does not strip unmatched or escaped emphasis markers: %s", reply => {
+  expect(reviewOutcomeReply(reply, "rejected")).toBe(false);
+});

tests/runner-e2e/chat-catalog-domains/external-app-review-outcomes.ts

@@ -85,6 +85,7 @@ function claimsExternalCompletion(reply: string) {
 }
 
 export function reviewOutcomeReply(reply: string, kind: ReviewKind) {
+  reply = reply.replace(/(?<![\\\w*_])(\*{1,3}|_{1,3})(?=\S)([^\r\n]*?\S)(?<!\\)\1(?![\w*_])/g, "$2");
   const clauses = reply.split(/[.!?;\n]+|\b(?:but|and)\b/i),
     affirmative = clauses.some(
       (clause) =>

tests/runner-e2e/chat-catalog-domains/external-app-review.test.ts

@@ -362,3 +362,18 @@ it.each(["no human approval", "wrong human", "successful invocation", "successfu
     expect(gradeAppReview(o).passed).toBe(false);
   },
 );
+
+it("accepts the measured whole declined reply with bold negation while still requiring every later phase", () => {
+  const reply = `The launch page update was **not** applied: the approval card for the "Synthetic Launch Notes" write was declined, so the page still reads "Controlled launch page."
+
+The update I attempted contained:
+- Public launch: **November 8, 2030**
+
+Do you want me to retry (you'll need to approve the card) or change the content first?`;
+  expect(reviewOutcomeReply(reply, "rejected")).toBe(true);
+  const o = valid();
+  o.phaseReplies[0]!.body = reply;
+  expect(gradeAppReview(o).passed).toBe(true);
+  o.state.phases = o.state.phases.slice(0, 1);
+  expect(gradeAppReview(o).passed).toBe(false);
+});

tests/runner-e2e/chat-catalog-domains/external-app-review.ts

@@ -360,7 +360,7 @@ export const appReviewExternalCases = [
     gradeAppReview,
     {
       authentication: "human-session",
-      oracleVersion: "paperclip.chat-catalog.app-review.v5",
+      oracleVersion: "paperclip.chat-catalog.app-review.v6",
       queryAttachments: [{ stateKey: "factualContextAttachment", filename: "launch-update.md" }],
       expectedRunCount: 6,
       minimumRunCount: 6,