fix: thinking controls offer unsupported provider-native efforts (#145248)

Related: #144866, #90703

## What Problem This Solves

Fixes thinking controls offering and accepting advanced levels from provider-native labels such as `MAX` or `XHIGH`, even though the actual model request silently uses a lower level.

## User Impact

The model picker, session validation, and execution now agree on which advanced levels a custom provider supports. Lowercase canonical efforts still work, and explicit mappings such as `max: "MAX"` preserve the provider's exact wire value. Unsupported selections fail without changing the session or its runtime/account selection.

## Why This Change Was Made

The shared thinking profile incorrectly applied user-input aliases to provider-native capability labels. It now reads those labels exactly, preserving the [existing explicit-mapping policy](https://github.com/openclaw/openclaw/pull/102993#issuecomment-4931608378). Provider-owned profiles, explicit maps, and null opt-outs are unchanged; no new aliases or transport policy are introduced.

The earlier lowercase capability-admission bug is already fixed on main. This update addresses the remaining inconsistent advertisement rather than restoring the older transport changes.

## Evidence

A real isolated Gateway and secretless loopback HTTP provider exercised `models.list → sessions.patch → chat.send → chat.history`:

| Case | Before | After |
| --- | --- | --- |
| Unmapped native `MAX` / `XHIGH` | Advertised and accepted `max` / `xhigh`; actual wire used `high` | Not advertised; unsupported selection rejected; accepted `high` sends `high` |
| Canonical lowercase `max` | Session and wire `max` | Unchanged |
| Explicit `max → MAX` mapping | Session `max`, wire `MAX` | Unchanged |
| Explicit null opt-out | Selection rejected | Unchanged |

Rejected candidate selections left the complete persisted session row unchanged. A saved synthetic account selected through `/model`, its explicit OpenClaw runtime, and model identity survived both rejection and a subsequent real HTTP turn. A copied pre-fix session now displays its effective supported level instead of the unavailable Maximum choice.

- [Inspected before/after screenshots from the real Gateway UI](https://github.com/openclaw/openclaw/pull/145248#issuecomment-5740680770).
- New regression fails on the original shared profile; 137 focused policy/admission/wire-contract tests pass after the fix.
- Focused formatting and diff checks pass. Full static and matrix checks remain in hosted CI.
- Local provider responses were synthetic; no external provider or live-account behavior is claimed.

Co-authored-by: Ayaan Zaidi <hi@obviy.us>

src/auto-reply/thinking.levels.test.ts

@@ -708,6 +708,35 @@ describe("listThinkingLevels", () => {
     ).toBe(true);
   });
 
+  it("does not treat provider-native effort labels as user thinking aliases", () => {
+    const catalog = [
+      {
+        provider: "custom",
+        id: "native-efforts",
+        api: "openai-completions",
+        reasoning: true,
+        compat: { supportedReasoningEfforts: ["high", "XHIGH", "MAX", "extra-high", "auto"] },
+      },
+    ];
+
+    expect(listThinkingLevels("custom", "native-efforts", catalog, "openclaw")).toEqual([
+      "off",
+      "minimal",
+      "low",
+      "medium",
+      "high",
+    ]);
+    expect(
+      isThinkingLevelSupported({
+        provider: "custom",
+        model: "native-efforts",
+        level: "max",
+        catalog,
+        agentRuntime: "openclaw",
+      }),
+    ).toBe(false);
+  });
+
   it("uses advanced catalog efforts and derives OpenClaw Ultra from Max", () => {
     const catalog = [
       {

src/auto-reply/thinking.ts

@@ -240,8 +240,7 @@ function appendCatalogAdvancedThinkingLevels(
     }
   }
   let supportsMax = profile.levels.some(({ id }) => id === "max");
-  for (const effort of compat?.supportedReasoningEfforts ?? []) {
-    const level = normalizeThinkLevel(effort);
+  for (const level of compat?.supportedReasoningEfforts ?? []) {
     if (
       level === "ultra" ||
       ((level === "adaptive" || level === "xhigh" || level === "max") &&