fix(cache): warm assert-ai's artifact cache from the baseline so paired McNemar can reach PASS

[tangym]

Summary

The paired regression gate was silently unreachable on any consumer whose behavior YAMLs don't set an explicit suite: field (both assert-ci-banking-demo and azure-doc-qa-eval are in this shape). Two independent misses had to line up:

1. assert-ai defaults suite_id to eval-<current-timestamp> when the config leaves suite: unset (see assert_ai/config.py:219). Every fresh assert-ai run writes into its own results/eval-XXX/ and never sees the prior run's artifacts under results/eval-YYY/.
2. The action downloaded the baseline artifact to a sibling directory (assert-ai-baseline/<slug>/) that assert-ai's cache lookup never inspects — the cache lives under each behavior's artifacts_root, which is assert-ai-artifacts/<slug>/.

Net effect: test_set regenerated on every dispatch (LLM call, non-deterministic), so detect_test_set_drift.py correctly observed a SHA mismatch and reported TestSetChanged. No PR could ever reach PASS.

Observed on assert-ci-banking-demo runs 31795690406 and 31796356987 after #9 landed: both used the fixed action, both found and downloaded the previous baseline, both regenerated the test_set from scratch, both reported TestSetChanged.

Fix

Two coupled changes:

scripts/plan_behaviors.py — pin suite in the frozen config

suite = cfg.get("suite") or slug
cfg["suite"] = suite

Every dispatch of the same behavior now writes into <artifacts_root>/results/<slug>/ deterministically. Consumer configs that already set suite: are preserved as-is.

action.yml — warm the artifact cache from the baseline

New composite step after Download baseline artifact:

- name: Warm the artifact cache from the baseline
  if: ${{ steps.resolve-baseline.outputs.needs-download == 'true' && steps.baseline-run.outputs.found == 'true' }}
  shell: bash
  run: |
    python <<'PY'
    ...
    shutil.copytree(baseline_slug_dir, artifacts_root, dirs_exist_ok=True)
    PY

assert-ai run then hydrates its context from the baseline's latest.json via activate_latest_artifacts, cache-hits on test_set (identical fingerprint = identical SHA), and skips regeneration.

Downstream stages behave correctly

• No-code-change PR: all stages cache-hit → identical scores.jsonl → paired McNemar on identical data → PASS.
• Real target-code PR: inference fingerprint changes because the target callable changed → inference cache misses → fresh inference and judging → paired McNemar compares baseline scores against the new agent's scores case-by-case → PASS/FAIL/WARN as the design intends.

One-time transition

Existing baseline artifacts uploaded under the old results/eval-<timestamp>/ layout are ignored by the new results/<slug>/ layout. The first dispatch of any consumer after upgrading will report FirstRun; the next dispatch establishes the new baseline and reaches PASS. Two-dispatch onboarding hiccup, not a steady-state cost.

Tests

Two regressions on test_plan_behaviors.py:

• test_plan_pins_suite_to_slug_when_config_omits_it — behavior config without suite: produces frozen.suite == slug and manifest.suite == slug.
• test_plan_preserves_user_supplied_suite — behavior config with explicit suite: keeps it verbatim.

52/52 tests pass locally.

What consumers need to do

Just bump the action pin to include this commit. No workflow change needed.