docs(inference): document probe budget helpers

Fixes NVIDIA#10260

Signed-off-by: Deepak Jain <deepujain@gmail.com>

Author: deepujain

src/lib/actions/sandbox/inference-invocation-probe.ts

@@ -49,6 +49,7 @@ export const REBUILD_INFERENCE_INVOCATION_TIMEOUT_MS = 100_000;
 export const READINESS_INFERENCE_INVOCATION_TIMEOUT_MS = 30_000;
 const INFERENCE_INVOCATION_MAX_RESPONSE_BYTES = 64 * 1024;
 
+/** Build the protocol-specific request used to verify sandbox inference. */
 function buildProbeRequest(input: SandboxInferenceInvocationInput): {
   endpoint: string;
   headers: string[];

src/lib/inference/health.ts

@@ -106,6 +106,7 @@ function useStatusProbeTiming(argv: string[]): string[] {
   );
 }
 
+/** Set the visible-output budget on a generated chat-completions probe. */
 function capStatusProbeOutput(
   argv: string[],
   maxTokens: number = HEALTH_PROBE_MAX_TOKENS,
@@ -123,6 +124,7 @@ function capStatusProbeOutput(
   return next;
 }
 
+/** Build curl arguments for an authenticated chat-completions health probe. */
 function buildChatCompletionsStatusProbeCurlArgs(
   model: string,
   endpoint: string,

src/lib/inference/max-tokens-field.ts

@@ -29,9 +29,10 @@
  */
 export const MIN_PROBE_REPLY_TOKENS = 16;
 
-// Gemini 2.5 can spend part of the output budget on reasoning before it emits
-// visible content. Match the established compatible-endpoint smoke budget so
-// a healthy route has room to return the requested acknowledgement (#10260).
+/**
+ * Gemini probe budget with room for reasoning before visible content.
+ * Matches the established compatible-endpoint smoke budget (#10260).
+ */
 export const GEMINI_PROBE_REPLY_TOKENS = 256;
 
 // Matched by prefix rather than exact id: Azure OpenAI deployments append