Tighten the comments added for PR #8757

studio/frontend/src/features/chat/lib/per-model-params.ts

@@ -127,14 +127,9 @@ export function getReplayedParams(
   if (!remembered) {
     return current;
   }
-  // Key by key over the remembered set, not a spread of the entry. An entry is
-  // whatever is in the settings row, and this app is not its only possible
-  // writer: the backend's ChatInferenceSettings accepts every persisted key
-  // inside a per-model dict, maxSeqLength included, and the client sanitiser
-  // keeps them. Spreading such an entry would replay a context the backend
-  // never loaded -- the exact thing REMEMBERED_INFERENCE_PARAM_KEYS exists to
-  // prevent on the way in -- and would carry any stale key straight into the
-  // live params.
+  // Key by key, not a spread: the row accepts every persisted key (maxSeqLength
+  // included) from any writer, and spreading one would replay a context the
+  // backend never loaded, plus any stale key, into the live params.
   const replayed = { ...current };
   for (const key of REMEMBERED_INFERENCE_PARAM_KEYS) {
     const value = remembered[key];

studio/frontend/src/features/chat/stores/chat-runtime-store.ts

@@ -2215,10 +2215,8 @@ function getHydratedSettingsState(
   ) {
     // Same fence as the global set above: a key the user moved while this
     // request was in flight is their edit, and the memory does not outrank it.
-    // REMEMBERED, not PERSISTED: the same distinction getReplayedParams draws.
-    // maxSeqLength is persisted but never remembered, and an entry that carries
-    // one anyway -- the row accepts it -- must not replace the context the model
-    // actually loaded with.
+    // REMEMBERED, not PERSISTED, as in getReplayedParams: a maxSeqLength the row
+    // happens to carry must not replace the context the model loaded with.
     const replayed = { ...params };
     for (const key of REMEMBERED_INFERENCE_PARAM_KEYS) {
       const value = remembered[key];

studio/frontend/tests/per-model-params-hydration-max-seq-length.test.ts

@@ -1,8 +1,9 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-// The hydration replay is a second, independent copy of the replay rules, so it
-// needs its own guard for the one key the memory deliberately does not keep.
+// The hydration replay is a second copy of the replay rules, so it needs its own
+// guard. Its own file: per-model-params-hydration.test.ts shares store state
+// across its tests, and an appended case picks up an earlier one's temperature.
 
 import assert from "node:assert/strict";
 import { register } from "node:module";
@@ -22,12 +23,8 @@ const { useChatRuntimeStore } = await import(
 const QWEN = "unsloth/Qwen3.5-9B-GGUF";
 
 test("hydration does not replay a maxSeqLength an entry happens to carry", async () => {
-  // maxSeqLength is persisted but never remembered: the context a model loads
-  // with is owned by its load config, and replaying a second copy would leave
-  // the runtime advertising a context the backend never loaded. This app never
-  // writes one into an entry -- but the settings row accepts one, from an older
-  // client, a hand-written payload or a future key set, so the read side has to
-  // hold the rule too, not just the write side.
+  // maxSeqLength is persisted but never remembered: the context belongs to the
+  // load config. This app writes none into an entry, but the row accepts one.
   settingsHttp.settings = {
     inferenceParamsByModel: {
       [QWEN]: { temperature: 0.2, maxSeqLength: 131072 },

studio/frontend/tests/per-model-params.test.ts

@@ -362,10 +362,8 @@ test("the snapshot drops params the current model never set", () => {
   assert.equal("topK" in picked, false);
 });
 
-// An entry is whatever the settings row holds, and this app is not its only
-// possible writer: the backend's ChatInferenceSettings accepts every persisted
-// key inside a per-model dict, so the read side has to hold the same rules the
-// write side does rather than trusting the shape.
+// The row accepts every persisted key from any writer, so the read side has to
+// hold the write side's rules rather than trust the entry's shape.
 test("a maxSeqLength in a stored entry is not replayed over the loaded context", () => {
   const replayed = getReplayedParams(
     true,
@@ -375,8 +373,7 @@ test("a maxSeqLength in a stored entry is not replayed over the loaded context",
     true,
   );
   assert.equal(replayed.temperature, 0.2, "the remembered value still replays");
-  // The same rule pickRememberedParams follows on the way in: a second copy of
-  // the context would leave the runtime advertising one the backend never loaded.
+  // A second copy of the context would advertise one the backend never loaded.
   assert.equal(replayed.maxSeqLength, 4096);
 });
 
@@ -389,17 +386,13 @@ test("a key that is not an inference param cannot reach the live params", () =>
     true,
   );
   assert.equal(replayed.temperature, 0.2);
-  // params flows on into request bodies and into the settings write, where the
-  // payload is extra="forbid".
+  // params flows on into request bodies and into an extra="forbid" settings write.
   assert.equal("notAParam" in replayed, false);
 });
 
-// Model ids are not sanitised identifiers. A Hub repo is slash-separated, a
-// custom folder is an absolute path -- a drive path on Windows, a UNC path from
-// a share or WSL, a spaced path under ~/Library on macOS -- and an external
-// model is provider-qualified. All of them are opaque keys here, and an id that
-// did not survive the round trip would mean that platform silently cannot
-// remember settings at all.
+// Model ids are opaque keys: a Hub repo, an absolute path on any OS, or a
+// provider-qualified external id. One that did not round trip would mean that
+// platform silently cannot remember settings.
 for (const [label, id] of [
   ["a Windows drive path", "C:\\Users\\Daniel\\models\\Qwen3-8B"],
   ["a UNC share path", "\\\\fileserver\\models\\gemma-3-270m-it"],