feat(gradient): range-aware file read deduplication

[BYK]

Summary

• Replace path-only file dedup with range-aware coverage checks that consider offset/limit parameters on read tool inputs
• Earlier reads are only collapsed when a later read covers the same or wider line range (full-file covers everything; ranged reads use interval superset check)
• Prevents over-deduplication where a narrow later read would incorrectly collapse an earlier full-file read

New helpers

• ReadRange type — { path, offset, limit } tuple for read tool inputs
• extractReadRange() — parses tool input JSON to extract path + line range
• laterReadCovers(later, earlier) — exported range coverage check with comprehensive edge case handling (full-file, open-ended, bounded intervals)
• rangeLabel() — formats range info for dedup annotations (e.g., "lines 10-49")

Changes to deduplicateToolOutputs()

The simple fileLatest Map (path → latest index) is replaced with a per-file list of ReadRange entries. The second pass checks whether any later read covers the current read's range using laterReadCovers(), instead of just comparing latest message index by path.

Annotation text updated from "earlier version of" to "earlier read of" with optional range info.

Test coverage

• 12 unit tests for laterReadCovers() covering all combinations: full-file ↔ full-file, full-file ↔ ranged, ranged ↔ ranged, non-overlapping, open-ended, different paths
• 10 integration tests for range-aware dedup scenarios: full-file covers ranged, ranged doesn't cover full-file, wider covers narrower, non-overlapping both kept, annotation formatting, read_file tool name compatibility
• All 5 existing dedup tests updated for new annotation text

Closes #117