spec: local eval runner & vendor-neutral adapter layer

[sunilgattupalle]

Summary

• Adds specs/2026-06-16-local-eval-runner-design.md — full design spec for the local eval runner, source adapters, targets, importers, config runner, and CLI
• Establishes the layered vendor-neutral architecture: core stays frozen, vendor adapters (Langfuse, Harness, etc.) live behind optional extras
• Primary use case: edit prompt → harness-evals run my-eval.yaml → diff scores against baseline

Key design decisions

• Four adapter family ABCs (BaseDatasetSource, BasePromptSource, BaseEvalCaseSource, BaseEvalConfigSource) + BaseTarget — all vendor-neutral
• Seven plugin registries including _TARGETS, _METRICS, _BASELINE_STORES so third parties can extend every layer without forking
• ResourceRef + dual-syntax resolve() (URI shorthand or typed block)
• MissingAdapterError at config-load time (not a cryptic ImportError at execution)
• HttpTarget v1: bearer/api_key/basic auth; OAuth/mTLS deferred
• datasets.py → datasets/ package migration must be atomic with back-compat re-exports
• Golden.input serialisation contract: json.dumps for non-string inputs in PromptTarget
• gate_against_baseline() raises BaselineRegressionError; CLI exits non-zero

Reviewed by

• Initial architecture review (Sonnet)
• Independent full review (Opus) — five gaps identified and addressed in second commit

Test plan

• Review specs/2026-06-16-local-eval-runner-design.md
• Verify all adapter families, registries, and design decisions look correct before implementation begins

🤖 Generated with Claude Code

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}