Phase EJ Step 7: Flags-on smoke gate (real SC2)

[aberson]

Build step 7 of 8 — Phase EJ (umbrella #281) — Type: operator

Build doc: documentation/plans/evolve-judging-plan.md §7 Step EJ.7

What to build (operator smoke — no /build-step dispatch)

Verify the six flags compose end-to-end on real infrastructure before an overnight soak.

1. Register 2 frozen anchors (e.g. v10, v13) in data/baselines.json via the EL baselines CLI.
2. Run one short flags-on evolve:

   uv run python scripts/evolve.py --pool-size 3 --games-per-eval 3 --hours 0.75 \
     --priors-exclude-promoted --screen-null-diff --regression-rule one-sided \
     --panel-floor --refresh-dedup --budget-fit --fitness-mode both
   

3. Inspect data/evolve_results.jsonl for the new row types where triggered; confirm no crash, no orphaned bots/cand_* dirs, and pointer integrity.
4. Then run one 15-minute defaults-off smoke and confirm behavior matches a pre-EJ run (no new row types, no flag effects).

Existing context

• data/baselines.json is ABSENT until this step registers anchors (EL.6 smoke cleaned up its seeds). This step is the first time the panel floor can actually fire.
• Real-SC2 required (SC2 must be running). Follow evolve pre-launch hygiene (.claude/rules/evolve.md): git diff --staged --stat before launch — EVO_AUTO=1 commits sweep the entire index.
• This is a Type: operator step — it runs the artifact built by EJ.1–EJ.6; it authors no new code.

Done when

Flags-on run completes without crash; every triggered EJ mechanism left its audit row; defaults-off run shows zero behavior change; verdicts recorded per checklist item.

Flags (recommended)

n/a — Type: operator (no /build-step dispatch; /build-phase halts here for operator execution)

Depends on

#287 (all six flags must be built + documented first)

Parallel-safe with

none — strictly sequential (validates the composed output of EJ.1–EJ.6; hard Depends on: EJ.6)

Produces

Smoke evidence appended to the run log under documentation/soak-test-runs/; pass/fail verdict per checklist item

Operator workflow notes

• Operator step: /build-phase halts here by design (halt-contract class: operator step). Resume in a fresh session after the smoke completes.
• Requires real SC2 running + manual anchor registration.

---

Synced from documentation/plans/evolve-judging-plan.md by /repo-sync at c6099d9