feat(eval): foundations P0a — additive types + LatencySummary wiring

[7xuanlu]

Summary

Phase 0a of the eval-foundations refactor. Pure additive type extension; zero behavior change for current callers. Foundation for P0b (paths + port discovery), P0c (cost caps + save guards), P1 (L1 baselines), P2 (L2 live-daemon), P3 (L3 MCP contract tests).

What changed

• EvalLayer enum (L1Db / L2Http / L3Mcp / L4Skill) in new eval/layer.rs. L4Skill reserved.
• ReportEnv extended with 22 additive fields: layer, task, variant, embed_dim, similarity_fn_name, judge_model_id, mcp_schema_hash, skill_prompt_hash, schema_version, schema_db_version, migrations_hash, n_runs, is_single_run, run_id, timestamp_utc, git_sha, warmup_iterations, eval_max_usd_baseline_cap, eval_max_usd_run_cap, eval_max_wall_secs_cap, total_cost_usd, total_wall_secs. All Option<T> or #[serde(default)] — legacy JSON still deserializes.
• LatencySummary wired in eval/runner.rs (was hardcoded None); collected via std::time::Instant per query, summarized post-loop. from_micros() helper added to eval/latency.rs.
• fixture_set_hash() in eval/fixtures.rs for directory-wide TOML hashing (sort + per-file inner sha256 + outer hash, sha256[..16]).
• build_locomo_env() + build_lme_env() helpers populate the new env fields in all locomo + lme runners (including _with_gate for both).
• Cargo feature eval-harness gates tests/eval_harness.rs compilation so unrelated PRs don't break on eval refactors.
• pub const SCHEMA_VERSION: u32 = 51 exposed in db.rs (eval cache invalidation key for P1).
• fs2 workspace dep added (used by P0c/P1 for scenario.db file locks).
• Hand-rolled Default for ReportEnv mirrors serde default = "fn" attributes (#[derive(Default)] doesn't honor those — caught by adversarial review).

Honest deviations from plan (all approved in review)

1. Plan listed origin_version: Option<String> as new — already exists as String. Omitted duplicate.
2. Plan listed llm_provider_kind / llm_model_id as MTEB-aligned new field names — these don't exist on actual ReportEnv. Used legacy names (llm_provider_class, llm_model). Note for P0b: comparable_env_hash must hash the legacy names.
3. Roundtrip back-compat test had to include all 9 pre-existing required fields in legacy JSON (the plan's minimal { "fixture_revision": ... } would've failed required-field deserialization).
4. Both _with_gate runners converted to use build_*_env() helpers — adversarial review caught locomo _with_gate initially using ..Default::default() which silently produced inconsistent env state. Fixed in bed96dd9.

Test plan

• cargo clippy --workspace --all-targets --features origin-core/eval-harness -- -D warnings → clean
• cargo test --workspace --features origin-core/eval-harness → 1152 passed, 1 pre-existing failure (eval::retrieval::tests::test_multi_turn_eval — FastEmbed model network download, baseline carried, unrelated to P0a)
• cargo test --workspace (default features, no eval-harness) → eval_harness.rs gated out, no regression
• Existing baseline JSON files still deserialize via #[serde(default)] on new fields
• ReportEnv::default() returns same values as serde_json::from_str of a legacy-shape JSON — drift-catcher test added in bed96dd9

Follow-ups for P0b+

• fs2 dep added but unused at P0a; P0c/P1 use it for scenario.lock.
• migrations_hash field reads option_env!(\"ORIGIN_MIGRATIONS_HASH\") — None until P1's build.rs sets it.
• LatencySummary truncates sub-ms to 0ms (ms-resolution storage). Acceptable for current callers; revisit if sub-ms percentile needed.

Adversarial review

Full review of integrated impl ran before merge per CLAUDE.md "Code review before merge" rule. Two IMPORTANT findings fixed in bed96dd9:

1. locomo _with_gate runner converted to build_locomo_env
2. #[derive(Default)] replaced with hand-rolled impl Default that mirrors serde defaults

🤖 Generated with Claude Code