diff --git a/.env.example b/.env.example
index d8fbcaece..59271288d 100644
--- a/.env.example
+++ b/.env.example
@@ -54,7 +54,11 @@ OPENAI_STRONG_ANSWER_MODEL=gpt-5.6-sol
 OPENAI_QUERY_CLASSIFIER_MODEL=gpt-5.6-luna
 OPENAI_SUMMARY_MODEL=gpt-5.6-terra
 OPENAI_INDEXING_MODEL=gpt-5.6-terra
-OPENAI_MAX_OUTPUT_TOKENS=4000
+# Answer-generation output ceiling (billed per token actually used, not a spend
+# commitment). Must stay well above reasoning headroom: the old 4000 starved the
+# strong route's reasoning budget and discarded the answer after 60-90s (GEN-C1).
+# Matches the env.ts default.
+OPENAI_MAX_OUTPUT_TOKENS=16000
 OPENAI_QUERY_CACHE_SIZE=200
 # Max inputs per embeddings request (OpenAI caps a request at 2048 inputs / ~300k tokens).
 # embedTexts splits a full-corpus re-embed into batches of this size.
@@ -76,7 +80,10 @@ OPENAI_PROMPT_CACHE_TTL=30m
 #OPENAI_SAFETY_IDENTIFIER_SECRET=
 OPENAI_STORE_RESPONSES=false
 OPENAI_FAST_REASONING_EFFORT=low
-OPENAI_STRONG_REASONING_EFFORT=high
+# "high" overruns OPENAI_ANSWER_TIMEOUT_MS and starves the safety-critical
+# medication_dose_risk/table_threshold classes; "medium" is ample for answers
+# grounded in retrieved sources and matches the env.ts default.
+OPENAI_STRONG_REASONING_EFFORT=medium
 OPENAI_SUMMARY_REASONING_EFFORT=medium
 OPENAI_VISION_REASONING_EFFORT=low
 OPENAI_TEXT_VERBOSITY=low
@@ -104,6 +111,12 @@ RAG_AWAIT_QUERY_LOGS=false
 # Design-exploration mockup routes (/mockups/*) 404 in production builds unless
 # explicitly opted in. Always reachable in dev/test.
 #NEXT_PUBLIC_MOCKUPS_ENABLED=false
+# Allow unauthenticated visitors to upload documents — gates an anonymous write
+# path, so leave unset (off) unless you intend a public-intake workspace. When
+# enabled, anonymous uploads are owned by PUBLIC_WORKSPACE_OWNER_ID so that
+# owner-scoped reads still apply to them.
+#NEXT_PUBLIC_PUBLIC_UPLOADS_ENABLED=false
+#PUBLIC_WORKSPACE_OWNER_ID=
 # Optional canonical public origin for generated metadata, for example
 # https://clinical-kb.example.org. Railway deployments otherwise use the
 # provider-supplied RAILWAY_PUBLIC_DOMAIN; request hosts are only used in dev.
diff --git a/playwright.config.ts b/playwright.config.ts
index 27f0c2c5b..38b088b92 100644
--- a/playwright.config.ts
+++ b/playwright.config.ts
@@ -25,6 +25,10 @@ export default defineConfig({
   // hydration, sub-pixel tap targets) without masking a real regression, which
   // still fails 3× before it is reported. Local stays at 0 so flakes surface.
   retries: process.env.CI ? 2 : 0,
+  // Fail the run if a stray `test.only` is committed: otherwise it silently
+  // narrows CI to that one test (and skips the whole release matrix) while the
+  // required check still reports green.
+  forbidOnly: !!process.env.CI,
   expect: {
     timeout: 10_000,
   },
diff --git a/src/app/api/documents/route.ts b/src/app/api/documents/route.ts
index 9a4e24ccb..cb3350718 100644
--- a/src/app/api/documents/route.ts
+++ b/src/app/api/documents/route.ts
@@ -193,11 +193,14 @@ export async function GET(request: Request) {
 
     const { data, error, count } = await query;
 
-    if (error) throw new Error(error.message);
+    // An `offset` past the end of the result set makes PostgREST return PGRST103
+    // ("Requested range not satisfiable"). That is an empty page, not a server
+    // error, so return an empty page instead of throwing a 500.
+    if (error && error.code !== "PGRST103") throw new Error(error.message);
     // An authenticated caller reads PUBLIC (owner_id IS NULL) documents alongside their own via
     // withOwnerReadScope. Redact operator-internal storage fields on the rows they do not own so a
     // shared public document never exposes its owner's storage_path/content_hash/etc. (S1/D1).
-    const rawDocuments = (data ?? []) as unknown as DocumentListRow[];
+    const rawDocuments = (error ? [] : (data ?? [])) as unknown as DocumentListRow[];
     const ownedDocumentIds = new Set(
       rawDocuments.filter((document) => callerOwnsDocumentRow(document, access.ownerId)).map((document) => document.id),
     );
diff --git a/src/lib/rag.ts b/src/lib/rag.ts
index 1c913ed74..a2942d577 100644
--- a/src/lib/rag.ts
+++ b/src/lib/rag.ts
@@ -146,6 +146,7 @@ import {
   queryTextForStorage,
 } from "@/lib/query-privacy";
 import { normalizeSourceMetadata } from "@/lib/source-metadata";
+import { safeErrorLogDetails } from "@/lib/privacy";
 import { isClinicalImageEvidence, normalizeImageBbox } from "@/lib/image-filtering";
 import {
   SOURCE_BACKED_REVIEW_FALLBACK_REASON,
@@ -3042,7 +3043,7 @@ export function parseAnswerJson(raw: string, results: SearchResult[], query?: st
     // GEN-C2 / GEN-H2: numeric faithfulness gate.
     return enrichGroundedReviewCitations(applyNumericVerification(answer), results);
   } catch (error) {
-    console.warn("Failed to parse answer payload, falling back to safe text:", error);
+    console.warn("Failed to parse answer payload, falling back to safe text:", safeErrorLogDetails(error));
     return safeFallbackAnswer(raw, results, query);
   }
 }