Report an empty model response as itself, not as invalid JSON

[mostafaseyedan]

Problem

When the language model returns an empty stream, callLlmJson reports a JSON parsing problem that did not happen.

parseLlmJson('') throws No JSON structure found in LLM response. That message matches the /JSON|parse/i test used to classify retry failures, so the attempt increments parseFailures, and the user sees:

LLM returned invalid JSON after 3 attempts. Please try again.

Nothing was wrong with the model's JSON — there was no response at all. The suggested remedy is wrong too: the usual causes are an exhausted request quota or a model declining the request, and neither is fixed by retrying.

This is actively misleading in practice. From a real runtime.log, with the Copilot auto model:

schema=skill_triage attempt=1 structured=true  model=auto textLen=0 error=No JSON structure found in LLM response
schema=skill_triage attempt=2 structured=false model=auto textLen=0 error=No JSON structure found in LLM response
schema=skill_triage attempt=3 structured=false model=auto textLen=0 error=No JSON structure found in LLM response

All three attempts failed in 11 ms total with textLen=0. The surfaced error sends the reader looking for a prompt or schema problem, when the model never produced a token. The same failure hit context_file_review a minute later, so it affects every AI feature, not one page.

Fix

Detect the empty stream directly and report it with the model id:

The language model (auto) returned an empty response 3 time(s). This usually means the request quota is exhausted or the model declined the request.

Empty responses are tracked separately from parse failures rather than folded into them, so a genuine malformed-JSON reply still reports as before.

Also adds a runtimeDebug line recording the available model catalog when no preferred family matches, so the log shows which models the fallback selection had to choose from — the model=auto above gave no indication of what else was on offer.

Testing

npm run typecheck and npm run lint clean (0 errors); 107 webview tests pass.