fix: prevent max_tokens state mutation and clean up dead code

[Kailigithub]

Three targeted fixes in the LLM core and browser driver:

1. Prevent max_tokens state mutation (llmcore.py):

   • ClaudeSession.raw_ask and NativeClaudeSession.raw_ask were mutating self.max_tokens to 8192 when None.
   • This caused the default to "stick" after the first call, preventing subsequent calls from using a different value.
   • Fix: use a local variable max_tokens = self.max_tokens or 8192 instead.

2. Replace redundant JSON round-trip with copy.deepcopy() (llmcore.py):

   • ToolClient.chat() used json.loads(json.dumps(tools, ensure_ascii=False)) to deep-copy the tools schema.
   • copy.deepcopy() is cleaner, faster for non-JSON-serializable objects, and clearer in intent.

3. Fix dead code in TMWebDriver (TMWebDriver.py):

   • connected() callback created a string but never printed or returned it.
   • Added the missing print() call to match handle_close() behavior.

Verification:

• Both files pass python3 -m py_compile
• No new lint violations introduced