"Reserved for response" ignores `maxOutputTokens` for extension-contributed models (hard-coded to 8192)

Problem

The Copilot context window widget shows "Reserved for response" as ~8192 tokens for all extension-contributed models, regardless of the maxOutputTokens value declared by the provider.

Expected: "Reserved for response" reflects the actual maxOutputTokens reported by the LanguageModelChatProvider (e.g., 32768, 65536, etc.).

Actual: Always ~8192 tokens, making the reservation bar a tiny sliver even for models configured with large output budgets.

Root Cause

In the Copilot Chat extension's internal endpoint wrapper, ExtensionContributedChatEndpoint ignores the provider's declared maxOutputTokens and returns a hard-coded value:

File: extensions/copilot/src/platform/endpoint/vscode-node/extChatEndpoint.ts

get maxOutputTokens(): number {
  // The VS Code API doesn't expose max output tokens, use a reasonable default
  return 8192;
}

This value flows into toolCallingLoop.ts where it's used as the outputBuffer for usage reporting:

stream.usage({
  completionTokens: fetchResult.usage.completion_tokens,
  promptTokens: fetchResult.usage.prompt_tokens,
  outputBuffer: endpoint.maxOutputTokens,  // ← always 8192 for extension models
  promptTokenDetails,
});

Which then renders in chatContextUsageWidget.ts as the "Reserved for response" bar.

Why This Matters

• Extension providers do report maxOutputTokens on LanguageModelChatInformation — but Copilot's internal wrapper never reads it
• For models with large output budgets (32K–128K), the UI misleads users into thinking only ~8K is reserved
• This is particularly misleading for reasoning/thinking models where output can legitimately be very large
• The maxOutputTokens value is used correctly for the actual vLLM max_tokens parameter — only the display is wrong

Context

We've already solved the related issues in our extension (vLLM-2-Copilot):

• ✅ Token usage reporting via LanguageModelDataPart with 'usage' MIME type and snake_case keys (not blocked by isApiUsage())
• ✅ prompt_tokens / completion_tokens / total_tokens displayed correctly in the context widget
• ✅ provideTokenCount works for prompt budgeting

The only remaining gap is the outputBuffer / "Reserved for response" display.

Proposed Fix

Option A (preferred): Read maxOutputTokens from the model information when constructing ExtensionContributedChatEndpoint:

get maxOutputTokens(): number {
  return this.modelInfo.maxOutputTokens ?? 8192;
}

Option B: Expose maxOutputTokens on the LanguageModelChat API surface so Copilot can access it without reaching into provider internals.

Environment

• VS Code: 1.120+
• Extension API: LanguageModelChatProvider (proposed chatProvider, version 5)
• Extension: vLLM-2-Copilot (https://github.tik.uni-stuttgart.de/ac120805/vLLM-2-Copilot)

Screenshot

Shows "Reserved for response" as a tiny bar (~8192) despite maxOutputTokens: 32768 being configured and correctly reported by the provider.

Related: This is a distinct issue from #314722 (usage = 0 for extension providers), which has already been addressed. The token counts (prompt, completion, total) display correctly — only the output reservation bar is affected.