fix: reject out-of-order voice timing

[mikemikimike]

Summary / Problem

Fixes #15.

check_latency only reported gaps above the response budget. A corrupt call
whose agent turn started before the caller turn ended therefore produced no
finding and could look like a fast response.

Changes

• Report negative response gaps as a high-severity invalid_timing finding.
• Point the finding at the agent turn and include the overlap duration in the
  diagnostic message.
• Preserve the existing slow_response behavior for non-negative gaps.
• Add a regression test covering the finding, severity, turn index, and
  message, and update the documented test count.

Tests

• RED verification: the new regression test failed against the base commit
  with zero invalid_timing findings.
• pytest -q: 20 passed
• Focused regression test: 1 passed
• Ruff check: passed
• Ruff format check for modified files: passed
• CLI smoke tests: the clean fixture exits 0 and the misheard fixture exits 1
  with --strict
• Wheel build: passed with python -m build --wheel

Compatibility / Known limitations

• Existing positive latency gaps and severity thresholds are unchanged.
• Negative gaps are treated as invalid timing data and fail strict evaluation
  because the new finding is high severity.
• This check only evaluates adjacent user/agent pairs exposed by
  Interaction.pairs(); it does not infer timing across unrelated turns.

Issue

#15