fix: flag negative and out-of-order timestamps in check_latency (#15)

[Dheerajvarma1]

What & why

Fixes #15.

Root Cause

check_latency() computes response latency as
gap = agent_turn.start_s - user_turn.end_s and previously only checked
if gap > budget_s:.

As a result, materially negative gaps caused by excessive overlap or
out-of-order timing data were silently ignored because negative values
could never exceed the latency budget. This allowed malformed timing data
to pass without any finding.

Fix

Added an elif gap < -0.3: check to check_latency() that emits a
high-severity invalid_timing finding pointing to the offending agent
turn.

Overlap Policy

The existing repository behavior allows small conversational overlap.
Therefore, gaps between -0.3s and 0.0s remain valid and are not flagged.

Tests

Added regression tests covering:

• Negative response latency beyond the normal overlap threshold
• Out-of-order/backwards timestamps
• Legitimate ~200ms conversational overlap

Verification

• pytest -v — 22 passed
• ruff check . — all checks passed

[Dheerajvarma1]

I found that upstream PR #17 already addressed issue #15, so I'm closing this PR to avoid duplicating the fix. Thanks!