improvement: topic-aware section classification in AGENTS.md extraction

[tbitcs]

Summary

During AGENTS.md extraction, specsmith classifies sections by heading keywords only.
Some content blocks land in semantically wrong files because the heading doesn't
reflect the content type.

Examples

• Register maps (hardware reference) -> verification.md (should be architecture.md)
• Phase 2 roadmap (project status) -> rules.md (should be architecture.md or excluded)
• Repository structure (reference layout) -> rules.md (should be architecture.md)
• Windows path-length mitigation (env setup) -> rules.md (should be drift-metrics.md)

Proposed Fix

Use content-level keyword scanning (not just heading names) to route blocks:

• Lines containing register offsets, address maps -> architecture.md
• Lines containing milestone, completion, roadmap -> architecture.md
• Lines containing one-time setup, machine-specific -> drift-metrics.md
• Project-type hints: fpga-rtl projects route hardware descriptions differently than web apps

Related

Part of import quality improvements. P0 items (diff markers, dedup) already fixed.

[tbitcs]

Implemented in 279eec7