Feat/assessment benchmark

[tebanieo]

Closes #82

Assessment Quality Benchmark

Adds a staged benchmark that scores pipeline output against human-reviewed reference answers, so we can tell good runs from bad and catch quality regressions. v1 implements the engine-assignment stage end-to-end.

What's included

• Harness: case loader + manifest, acceptable-set scoring, assignment stage adapter, outcome-aware reporting with a trust banner, CLI (python -m benchmarks.runner.run --stage assignment)
• Assignment stage (deterministic-regression mode that is rules based, where a score drop = a code regression)
• Seed dataset: 2 human-reviewed cases + benchmarks/README.md
• ADR-024 (Proposed) recording the design

Design highlights

• Dataset is the durable asset; runner is thin/replaceable (SWE-bench model)
• Two modes: deterministic-regression (assignment) vs. LLM-quality (triage/schema/synthesis, later)
• Acceptable-set-per-query scoring (avoids penalizing defensible choices)
• Gold-input isolation (no error amplification across stages)
• Infra failures (throttles) quarantined, never scored as wrong; incomplete runs flagged

Testing

13 benchmark tests + full unit suite pass. src/ untouched (purely additive).

Out of scope (v1)

LLM-quality stages, performance/cost benchmark, synthetic case generation, CI wiring.