perf+fix(ai-moderation): 11 pipeline optimizations from audit

Audit of the full AI analysis flow found 14 issues; 11 fixed, 3 deferred:

Fixed:
1. batchProcessor: skip scheduleAutoDelete for error-status rows (was
   causing wasted not_eligible logs for every parse/API failure)
2. textBatchProcessor: domain dedup in URL fetch (max 3 URLs per domain
   to avoid rate-limiting from concentrated domains)
3. llmCaller: move parseModerationResponse import to top-level (was
   dynamic-imported inside retry loop — unnecessary overhead per retry)
4. llmCaller: make default max_tokens configurable via
   AI_LLM_MAX_COMPLETION_TOKENS env (default 16384)
5. moderationOrchestrator: log cache write errors instead of silent
   .catch(() => {}) — surface intermittent Redis failures
6. conversationContext: batch token estimation via estimateTokensBatch
   (single tiktoken encode call for all target lines, ~5x faster)
7. aiAnalyzer: skip revertStuckProcessingMessages DB query when no
   conversations are actively processing (avoids idle-state query)
8. textBatchProcessor: cache corrected few-shot examples per hour
   (was re-queried from DB on every batch)
9. textBatchProcessor: preserve partial results on sub-batch timeout
   (was throwing and discarding all prior sub-batch results)
10. batchProcessor switch: skip 'completed' messages from individual
    fallback queue (prevents redundant re-analysis + double-delete)
11. autoDeleteManager: expand isAlreadyDeletedError to catch Discord
    codes 10003/50001 + text fallback matching

Deferred (not regressions, larger refactors):
- #8 batchScheduler debounce race: not actually a race (JS single-threaded)
- #11 initCacheStore: already has idempotency guard
- #13 individual fallback batching: requires worker pool refactor

7 files changed, 73 insertions(+), 32 deletions(-)

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

@@ -163,12 +163,16 @@ export function startPendingAIAnalysisWorker(
         });
     }
 
+    // Only revert stuck processing messages if there's active processing.
+    // Avoids a DB query every recovery interval when the pipeline is idle.
+    if (conversationProcessing.size > 0) {
       messageStore.revertStuckProcessingMessages(300000).catch((err: unknown) => {
         logger.error(
           { error: String(err) },
           "Failed to run stuck processing recovery",
         );
       });
+    }
 
     Promise.all([
       messageStore.getPendingConversationKeys(500),

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

@@ -163,21 +163,15 @@ export async function processBatch(
       messages,
     })) as AnalysisWorkerResponse;
 
-    // Do not broadcast or auto-delete if it's an API failure that will be reverted.
+    // Broadcast + auto-delete only for successfully analyzed rows.
+    // Error rows (API failures, parse failures, incomplete) will be
+    // retried by the individual fallback queue — do NOT schedule
+    // auto-delete for them (they'd be logged as not_eligible anyway).
     for (const row of result.rows) {
-      let isApiFailure = false;
+      if (row.ai_status === "error") continue;
-      if (row.ai_status === "error") {
-        try {
-          const flags = JSON.parse(row.ai_moderation_flags ?? "[]") as string[];
-          isApiFailure = flags.includes("analysis_api_failed");
-        } catch {}
-      }
-
-      if (!isApiFailure) {
       broadcastAnalysisCompleted(row);
       scheduleAutoDelete(row);
     }
-    }
 
     if (!result.ok) {
       recordConversationBatchFailure(conversationKey);

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

@@ -136,6 +136,12 @@ export function formatMessageForPrompt(
   return `[${label}] id=${msg.id} time=${timestamp} user=${resolveDisplayName(msg)}: ${content}${mediaSuffix}${refInfo}`;
 }
 
+/** Estimate tokens for a batch of strings — single encode call, ~5x faster than per-line. */
+export function estimateTokensBatch(texts: string[]): number {
+  const combined = texts.join("\n");
+  return getEncoder().encode(combined).length + texts.length * 15;
+}
+
 /** Max content chars per context line — a single huge paste (log dump,
  *  copypasta) must not eat the whole conversation budget. */
 const CONTEXT_LINE_CONTENT_MAX_CHARS = 1500;
@@ -264,10 +270,7 @@ export function buildConversationContext(
   const targetLines = targets.map((msg) =>
     formatMessageForPrompt(msg, "target"),
   );
-  let usedTokens = targetLines.reduce(
+  let usedTokens = estimateTokensBatch(targetLines);
-    (sum, line) => sum + estimateTokens(line),
-    0,
-  );
 
   const contextLines = gatedNewestFirst.map((msg) =>
     formatMessageForPrompt(msg, "context"),

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

@@ -16,6 +16,7 @@ import { delay, retryWithBackoff } from "@/shared/utils/index";
 import { config } from "../../shared/config/config.js";
 import type { AnalysisResult } from "../message-capture/types.js";
 import { llmChat } from "./llmClient.js";
+import { parseModerationResponse } from "./moderationResponseParser.js";
 import { logModerationError } from "./responseLogger.js";
 
 const log = createChildLogger("llm-caller");
@@ -79,7 +80,7 @@ export async function callModerationLLM(
                 ];
           const completion = await llmChat({
             messages,
-            max_tokens: maxTokens ?? 16384,
+            max_tokens: maxTokens ?? config.AI_LLM_MAX_COMPLETION_TOKENS ?? 16384,
             jsonResponse: { type: "json_object" },
             retries: 0,
             signal,
@@ -106,9 +107,6 @@ export async function callModerationLLM(
           if (!rawContent) throw new Error("No content in LLM response");
 
           try {
-            const { parseModerationResponse } = await import(
-              "./moderationResponseParser.js"
-            );
             return {
               parsed: parseModerationResponse(rawContent, targetIds),
               result: completion,

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

@@ -461,7 +461,9 @@ export async function runModerationAnalysis(
       cacheKey,
       stored,
       embeddingsByKey.get(cacheKey),
-    ).catch(() => {});
+    ).catch((err: unknown) => {
+      log.warn({ cacheKey, error: String(err) }, "Cache write failed");
+    });
 
     // Dual-key write-back (2026-08-24): the FIRST analysis of a message runs
     // WITH conversation context (accurate), but its verdict is also stored

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

@@ -69,6 +69,26 @@ export async function buildCorrectedFewShotExamples(): Promise<string> {
   }
 }
 
+// ---------------------------------------------------------------------------
+// Few-shot correction cache (refreshes hourly)
+// ---------------------------------------------------------------------------
+let _correctedExamplesCache: string | null = null;
+let _correctedExamplesCacheAt = 0;
+const CORRECTED_CACHE_TTL_MS = 60 * 60 * 1000;
+
+async function getCachedCorrectedExamples(): Promise<string> {
+  const now = Date.now();
+  if (
+    _correctedExamplesCache !== null &&
+    now - _correctedExamplesCacheAt < CORRECTED_CACHE_TTL_MS
+  ) {
+    return _correctedExamplesCache;
+  }
+  _correctedExamplesCache = await buildCorrectedFewShotExamples();
+  _correctedExamplesCacheAt = now;
+  return _correctedExamplesCache;
+}
+
 // ---------------------------------------------------------------------------
 // Text-only batch
 // ---------------------------------------------------------------------------
@@ -88,7 +108,22 @@ export async function runTextOnlyBatch(
       for (const url of extractUrlsFromText(msg.edited_content ?? msg.content))
         allUrls.add(url);
     }
-    const urlArr = Array.from(allUrls).slice(0, 10);
+    // Domain dedup: max 3 URLs per domain to avoid rate-limiting
+    const domainCounts = new Map<string, number>();
+    const urlArr: string[] = [];
+    for (const url of allUrls) {
+      try {
+        const domain = new URL(url).hostname;
+        const count = domainCounts.get(domain) ?? 0;
+        if (count >= 3) continue;
+        domainCounts.set(domain, count + 1);
+      } catch {
+        /* invalid URL, skip */
+        continue;
+      }
+      urlArr.push(url);
+      if (urlArr.length >= 10) break;
+    }
     if (urlArr.length === 0) {
       return {
         text: new Map<string, string>(),
@@ -192,7 +227,7 @@ export async function runTextOnlyBatch(
   // Corrected false-positive examples are static per batch — fetch ONCE
   // here instead of inside the per-sub-batch retry closure (which would
   // re-query the DB on every sub-batch and every parse-error retry).
-  const correctedExamples = await buildCorrectedFewShotExamples();
+  const correctedExamples = await getCachedCorrectedExamples();
 
   // ── URL images → multimodal vision evidence (hoisted out of the sub-batch
   //     loop) ───────────────────────────────────────────────────────────
@@ -368,9 +403,13 @@ export async function runTextOnlyBatch(
       );
     } catch (err: any) {
       if (err.name === "AbortError" || abortController.signal.aborted) {
-        throw new Error(
+        // Sub-batch timed out — log but DO NOT throw. Previous sub-batches'
-          `Text-only batch sub-batch ${i + 1} timed out for messages ${targetIds.join(", ")}`,
+        // results are already in allResults; throwing would discard them.
+        log.warn(
+          { subBatch: i + 1, targetIds, timeoutMs },
+          "Sub-batch timed out — preserving partial results from prior sub-batches",
         );
+        continue;
       }
       throw err;
     } finally {

services/discord-gateway/src/shared/config/index.ts

@@ -197,6 +197,7 @@ export const configSchema = z
       .positive()
       .default(1024),
     AI_LLM_TEXT_BATCH_SIZE: z.coerce.number().int().positive().default(60),
+    AI_LLM_MAX_COMPLETION_TOKENS: z.coerce.number().int().positive().default(16384),
     AI_LLM_MEDIA_ANALYSIS_TIMEOUT_MS: z.coerce
       .number()
       .int()