fix(infer): reject prompts without generation space

[RerankerGuo]

PR type

• Bug Fix
• New Feature
• Document Updates
• More Models or Datasets Support

PR information

Fixes #6441.

Background

When an encoded prompt consumes all or more than the model context,
set_default_max_tokens() currently calculates a zero or negative
generation limit and stores it in the request configuration.

The invalid value then reaches a backend-specific configuration class,
which raises a secondary error such as:

ValueError: `max_new_tokens` must be greater than 0, but is -5923.

This obscures the actual problem: the prompt itself leaves no room for
generation.

Changes

• Reject requests when the available generation space is zero or
  negative.
• Report the input token count, model context limit, and configured
  offset in the error.
• Preserve the existing behavior that caps valid requests to the
  remaining context.
• Add CPU boundary tests for overlong, exactly-full, and valid prompts.

Verification

• PYTHONPATH=. .venv/bin/python -m unittest tests.infer.test_infer_engine_limits
• PYTHONPATH=. .venv/bin/python tests/run.py --pattern test_infer_engine_limits.py
• .venv/bin/pre-commit run --all-files
• git diff --check

Impact

Valid inference requests are unchanged. Requests that previously
propagated zero or negative generation limits now fail earlier with an
actionable error. The change applies consistently to inference backends
that use set_default_max_tokens().

Experiment results

Before the fix, prompts of length 5 and 4 with max_model_len=4
produced generation limits of -1 and 0 without a direct error. After
the fix, both raise the input-length error, while a valid prompt still
caps the requested generation length from 4 to 2.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.