fix: retry transport errors without misclassifying model text

[Fwind43]

Why

Short transport errors such as !!!Error: ConnectionError: empty response could be mistaken for a final answer. Matching more text prefixes would instead misclassify legitimate model-generated text.

What

• Tag internally generated diagnostics with LLMError (a string subclass preserving streaming/display compatibility).
• Propagate transport/parser errors through content-block metadata into MockResponse.error for native and text-tool clients.
• Use error provenance, not response text, for no-tool error retries, history exclusion and mixed-session fallback.
• Suppress tool calls from failed/partial responses before tool parsing.
• Preserve bounded application retries and existing transport backoff; no new dependencies.

Testing

• python -m unittest discover -s tests: 11 tests pass.
• Covers identical error text from the model vs transport, parser diagnostics, native/text clients, history, mixed-session fallback, partial tools and bounded retries.
• py_compile for ga.py and llmcore.py; git diff --check passes.
• Synthetic transport responses only; no live LLM requests.

Supersedes the initial prefix-matching approach in this PR.