[SajjadPSavoji]

Deployments with a gateway that rejects big requests had to patch the code to get a smaller per-request size. This adds three optional env vars: SKILLSPECTOR_MAX_BATCH_INPUT_TOKENS, SKILLSPECTOR_MAX_PROMPT_BYTES and SKILLSPECTOR_TP4_MAX_BATCH_INPUT_TOKENS. Each one only lowers a limit, so the model budget and the CLI provider byte cap still win when they're smaller. Unset or blank keeps today's behavior and bad values log a warning without echoing the value.

The byte cap goes through the existing chunking in get_batches and TP4's own chunk planner. Prompts are also checked right before sending in the sync and async paths. A line or fixed prompt that can't fit is marked SIZE_LIMIT for that batch, nothing is truncated and other batches keep running. README has a short section on units and precedence.

Thanks @SajjadPSavoji for the suggested design in the thread.

Tests: new TestConfigurableRequestLimits in tests/nodes/test_llm_analyzer_base.py plus the TP4 token ceiling, byte ceiling and SIZE_LIMIT partial tests in tests/test_mcp_tool_poisoning.py. They fail before the change and pass after. I also ran the rest of test_llm_analyzer_base.py, test_mcp_tool_poisoning.py, test_meta_analyzer.py and test_meta_analyzer_fallback.py, and ruff check on the touched files.

Fixes #1003