[Feature] Benchmark suite with CI integration

[Siddhant-K-code]

Summary

Add a Go benchmark suite that measures dedup pipeline performance across different data sizes and configurations. Publish results in CI.

Motivation

Publishable benchmark numbers are useful for:

• Blog posts and conference talks ("12ms for 50 chunks" backed by reproducible data)
• Detecting performance regressions on PRs
• Comparing algorithm variants (e.g., average vs complete linkage)
• Giving users confidence in production readiness

Benchmarks to add

// pkg/contextlab/
BenchmarkCluster_10Chunks
BenchmarkCluster_50Chunks
BenchmarkCluster_100Chunks
BenchmarkCluster_500Chunks
BenchmarkMMR_10Chunks
BenchmarkMMR_50Chunks
BenchmarkSelector_10Clusters
BenchmarkSelector_50Clusters

// pkg/dedup/
BenchmarkDistanceMatrix_50
BenchmarkDistanceMatrix_200

// pkg/compress/
BenchmarkCompress_ShortText
BenchmarkCompress_LongText

// End-to-end
BenchmarkFullPipeline_50Chunks
BenchmarkFullPipeline_200Chunks

CI integration

• Add benchmark job to ci.yml that runs go test -bench=. -benchmem ./...
• Use benchstat to compare against baseline on PRs
• Store baseline results in testdata/benchmarks/baseline.txt
• Comment on PR with performance diff if regression > 10%

Deliverables

• Benchmark functions in *_test.go files
• Synthetic test data generator (deterministic embeddings)
• make bench target
• CI workflow for benchmark comparison
• BENCHMARKS.md with latest results table

Acceptance Criteria

• go test -bench=. ./... runs all benchmarks
• Results are reproducible (deterministic test data)
• CI detects regressions > 10%
• README links to benchmark results