fix(performance): eliminate redundant regex calls in structured output mode

[sidmohan0]

Summary

• Fixes 3-4x performance regression observed in CI benchmarks
• Eliminates redundant regex processing in smart cascade and auto engine modes
• Optimizes Span class imports in multi-chunk processing

Root Cause Analysis

The performance regression was caused by multiple redundant regex calls when using structured output mode:

1. Double Regex Calls: Smart cascade and auto engine modes called annotate() first, then annotate_with_spans() on the same text when structured=True
2. Redundant Imports: Multi-chunk processing imported Span class for every span instead of once per batch
3. Inefficient Logic: The cascade decision logic required dict format but structured mode needed spans, causing duplicate work

Performance Fixes

Before (Problematic Pattern)

# Double processing - called regex engine twice\!
regex_result = self.regex_annotator.annotate(text)
if should_stop:
    _, result = self.regex_annotator.annotate_with_spans(text)  # REDUNDANT\!
    return result.spans

After (Optimized Pattern)

# Single processing - use spans directly
_, result = self.regex_annotator.annotate_with_spans(text)
regex_result = convert_spans_to_dict(result.spans)  # Convert for decision logic
if should_stop:
    return result.spans  # Already have spans\!

Test Plan

• All benchmark tests pass
• Structured output performance improved significantly
• Regex performance maintained at sub-4ms
• Backward compatibility preserved (identical outputs)
• Auto engine fast path optimized
• Smart cascade performance improved
• Multi-chunk processing optimized

🤖 Generated with Claude Code