feat(eval+server): foundations P0b — comparable hash + path layout + daemon port discovery

[7xuanlu]

Summary

Phase 0b of the eval-foundations refactor. Additive only; no behavior change to the HTTP API.

Builds on P0a (PR #178, merged as 46a9703).

What changed

• comparable_env_hash(env: &ReportEnv) -> String in eval/report.rs — sha256[..8] over the comparable subset of ReportEnv fields: fixture_revision, embedder_revision, llm_provider_class, llm_model, mcp_schema_hash (or empty), skill_prompt_hash (or empty), schema_version, schema_db_version (or empty), similarity_fn_name. Excludes layer, variant, task, n_runs, run_id, timestamp, cost/latency fields. Hash contract documented inline — any field-set change must bump default_schema_version().
• encode_baseline_path(root, env) -> PathBuf in eval/report.rs — layered path <root>/<layer_dir>/<task>/<variant>__<hash>.json. Panics with named-field message if env.layer/task/variant is None (precondition for save_full_report callers in P0c).
• Daemon port advertisement in origin-server/src/main.rs:
  • println!("ORIGIN_LISTENING_ON={}", local_addr) after bind, before serve.
  • Optional ORIGIN_PORT_FILE env writes resolved port; hard-fails on write error (no silent swallow).
  • Existing tracing::info!("Listening on http://{}", local_addr) updated to use resolved local_addr instead of bind addr (otherwise it prints 127.0.0.1:0 under ephemeral-port mode).
• Integration tests:
  • crates/origin-core/tests/eval_report_paths.rs — full (layer × task × variant) collision grid (18 distinct paths from 3×2×3) + cross-layer hash parity.
  • crates/origin-server/tests/port_discovery.rs — spawn origin-server via env!("CARGO_BIN_EXE_origin-server"), verify both stdout printline and ORIGIN_PORT_FILE channels work with 30s timeout + cleanup.

Adversarial review (5 findings, 2 fixed pre-merge)

• IMPORTANT: tracing log line was printing :0 under ephemeral mode while new stdout printline showed resolved port — fixed in 845f7d5 (single Listening log line now uses local_addr).
• IMPORTANT: comparable_env_hash hash-input contract was implicit — fixed in 845f7d5 (doc-comment explicitly states "any modification MUST bump default_schema_version()").
• NIT: std::fs::write(ORIGIN_PORT_FILE) not atomic — acceptable for 5-byte port number on local disk; harness retries via read_to_string().ok().
• NIT: inside-function use std::io::Write — stylistic mismatch with rest of main.rs top-level imports.
• P2 RISK: cross-layer comparable-hash parity depends on L2 runner constructing ReportEnv.embedder_revision literal-identical to L1. P2 plan will add a canonical_embedder_revision() helper.

Test plan

• cargo clippy --workspace --all-targets --features origin-core/eval-harness -- -D warnings → clean
• cargo test --workspace --features origin-core/eval-harness → 1160 lib tests pass; 4 new integration tests pass (2 path-layout, 2 port-discovery); pre-existing eval::retrieval::tests::test_multi_turn_eval network failure unchanged
• Manual smoke: ORIGIN_BIND_ADDR=127.0.0.1:0 ./target/debug/origin-server shows single Listening on http://127.0.0.1:<resolved-port> line + matching ORIGIN_LISTENING_ON= stdout
• ORIGIN_PORT_FILE write contains resolved port

Follow-ups for P0c+

• comparable_env_hash will be consumed by save_full_report (P0c) for filename construction
• encode_baseline_path will be the canonical save path used by P1 baseline regeneration
• ORIGIN_LISTENING_ON + ORIGIN_PORT_FILE will be consumed by P2 http_harness.rs::DaemonHandle::spawn

🤖 Generated with Claude Code