Parent: #1411

Goal

Add a benchmark/test gate proving compression reduces context size without reducing pass rate on representative PDD tasks.

Prototype links

• AST test compression experiment: https://github.com/promptdriven/pdd_cloud/commit/b7b631a159848d9d2e1e09f7f64508266d8c5d1b
• AST PDD proof arm: https://github.com/promptdriven/pdd_cloud/commit/f0c512cc6d9d03c6fba172ae85b3a7e222ba99c8
• Compressed few-shot + contract-source proof: https://github.com/promptdriven/pdd_cloud/commit/f54f2ef6a7a71a10e8dd003a244414cd3f7fe075

Acceptance criteria

• Benchmark compares full tests, AST tests, AST+contracts, full few-shot, and compressed few-shot.
• Reports pass rate, token counts, wall-clock time, output churn, and missing-contract failures.
• Fails CI or benchmark gate if compression loses required contract symbols or regresses pass rate on frozen fixtures.
• Includes at least one fixture that previously failed without contract-source preservation.

Parent epic: #873.

Migrated from gltanaka/pdd#1418 (originally filed 2026-05-08 by gltanaka). gltanaka/pdd is deprecated.