fix(moderation): enforcement safety net — username-only offense never auto-deleted

Even with the prompt firewall (username vs content), the LLM can still
occasionally mis-apply a content-level zero-tolerance flag (sara /
conflict_instigation) to a message whose ONLY violation is the username
(e.g. 'matikanetanyahu'). The auto-delete eligibility check only recognized
exact offensive_username flags, so such false positives still deleted the
message.

Add a belt-and-suspenders guard in isNicknameOnlyViolation: if the flag set
is entirely username-attributable (offensive_username/sara/conflict_instigation)
AND the analysis text corroborates that the violation is username-only with
clean message content, route to nickname-reset instead of message deletion.

Adds 6 test cases covering the real matikanetanyahu scenario and the
false-positive/negative boundaries.

services/discord-gateway/src/modules/ai-moderation/autoDeleteEligibility.ts

@@ -72,13 +72,58 @@ export function parseModerationFlags(
  * correct enforcement is resetting the nickname to the default username.
  * Any other flag (sara, harassment, vulgar_language, ...) keeps the normal
  * delete path.
+ *
+ * Two detection paths (belt-and-suspenders):
+ *  1. Exact flags — the ONLY flag is `offensive_username`.
+ *  2. Analysis attribution — the flag set is exactly the username-attributable
+ *     set (`offensive_username`, `sara`, `conflict_instigation`) AND the
+ *     analysis text explicitly attributes the violation SOLELY to the
+ *     username while stating the message content itself is clean/ordinary.
+ *     This catches the false-positive case where the LLM wrongly applies a
+ *     content-level zero-tolerance flag (e.g. `sara`/`conflict_instigation`
+ *     for a username like "matikanetanyahu") instead of `offensive_username`.
  */
+const USERNAME_ATTRIBUTABLE_FLAGS = new Set([
+  "offensive_username",
+  "sara",
+  "conflict_instigation",
+]);
+
+/** Analysis text that clearly states the message content itself is clean. */
+const CONTENT_CLEAN_PATTERN =
+  /(?:isi pesan\s*(?:hanya|bersih|tidak (?:melanggar|ada)|membahas|berisi|bukan)|pesan\s*(?:bersih|tidak (?:melanggar|ada))|tidak ada (?:diskusi|konten|indikasi|pelanggaran))/i;
+
+/** Analysis text that clearly attributes the violation to the username. */
+const USERNAME_ATTRIBUTION_PATTERN =
+  /(?:username|nickname|nama pengguna)[\s\S]{0,40}?(?:mengandung|memiliki|berisi|melanggar|ofensif|mengecam|menyerang)/i;
+
 export function isNicknameOnlyViolation(
   message: MessageRecord,
   analysisResult?: AnalysisResult,
 ): boolean {
   const flags = parseModerationFlags(message, analysisResult);
-  return flags.length > 0 && flags.every((f) => f === "offensive_username");
+  if (flags.length === 0) return false;
+
+  // Path 1: exact offensive_username-only.
+  if (flags.every((f) => f === "offensive_username")) return true;
+
+  // Path 2: username-attributable flag set + analysis text corroborates
+  // that the violation is username-only with clean content.
+  const onlyUsernameAttributable = flags.every((f) =>
+    USERNAME_ATTRIBUTABLE_FLAGS.has(f),
+  );
+  if (!onlyUsernameAttributable) return false;
+
+  const analysis =
+    (analysisResult?.analysis ?? typeof message.ai_analysis === "string")
+      ? message.ai_analysis
+      : "";
+  if (!analysis) return false;
+
+  return (
+    USERNAME_ATTRIBUTION_PATTERN.test(analysis) &&
+    CONTENT_CLEAN_PATTERN.test(analysis)
+  );
 }
 
 /**

services/discord-gateway/tests/nicknameEnforcement.test.ts

@@ -11,7 +11,10 @@ import type {
   MessageRecord,
 } from "../src/modules/message-capture/types.js";
 
-function msg(flagsJson: string | null): MessageRecord {
+function msg(
+  flagsJson: string | null,
+  analysis?: string | null,
+): MessageRecord {
   return {
     id: "m1",
     guild_id: "g1",
@@ -34,6 +37,7 @@ function msg(flagsJson: string | null): MessageRecord {
     reference_guild_id: null,
     metadata: null,
     ai_moderation_flags: flagsJson,
+    ai_analysis: analysis ?? null,
   };
 }
 
@@ -74,4 +78,40 @@ describe("isNicknameOnlyViolation", () => {
     expect(isNicknameOnlyViolation(msg(null))).toBe(false);
     expect(isNicknameOnlyViolation(msg("[]"))).toBe(false);
   });
+
+  // ── Safety-net: content-level flags mis-applied to a username-only offense ──
+
+  it("true: partial flags sara/conflict + analysis attributes violation to username with clean content (matikanetanyahu case)", () => {
+    const analysis =
+      "Pengirim menggunakan username 'matikanetanyahu' yang menyerang tokoh politik terkait konflik Israel-Palestina. Meskipun isi pesan membahas alasan hilangnya tugas, keberadaan username tersebut tetap melanggar aturan server.";
+    expect(
+      isNicknameOnlyViolation(msg('["sara","conflict_instigation"]', analysis)),
+    ).toBe(true);
+  });
+
+  it("true: sara-only flag + analysis clearly attributes to username with clean content", () => {
+    const analysis =
+      "Username mengandung referensi politik (Netanyahu), tapi isi pesan hanya obrolan biasa tanpa diskusi politik. Warning ringan untuk username saja.";
+    expect(isNicknameOnlyViolation(msg('["sara"]', analysis))).toBe(true);
+  });
+
+  it("false: username-attributable flags but analysis lacks clean-content signal", () => {
+    const analysis =
+      "Pengirim mengkritik kebijakan luar negeri Israel secara eksplisit di dalam pesan. Pelanggaran diskusi topik terlarang.";
+    expect(
+      isNicknameOnlyViolation(msg('["sara","conflict_instigation"]', analysis)),
+    ).toBe(false);
+  });
+
+  it("false: mixed flags including a real content violation, even with clean analysis", () => {
+    const analysis =
+      "Username ofensif dan isi pesan berisi ancaman kekerasan terarah.";
+    expect(isNicknameOnlyViolation(msg('["sara","violence"]', analysis))).toBe(
+      false,
+    );
+  });
+
+  it("false: username-attributable flags but no analysis text available", () => {
+    expect(isNicknameOnlyViolation(msg('["sara"]', null))).toBe(false);
+  });
 });