Phase EL Step 7: Multi-hour observation soak

[aberson]

Build step 7 of 7 — Phase EL (umbrella #272)

Build doc: documentation/plans/evolution-lines-plan.md §7 Step EL.7

What to build

Type: wait. Multi-hour autonomous observation soak — the end-to-end validation the unit tests cannot give (the loop runs unattended over wall-clock time, so scheduler drift, registry write races, and extinction thrash are invisible to short runs):

python scripts/evolve.py --lineages 3 --population-cap 3 --hours 6 --fitness-mode both

Observe: lineages diverge on fingerprint; ≥1 extinction event fires (or is justified as not-needed under cap); baseline-fitness trends are sane; no orphan SC2 processes; EVO_AUTO commit hygiene holds (clean [evo-auto] rows). Use a 2-hour check-in pattern (inspect data/evolve_run_state.json + recent commits + log tail) before letting the remainder run unattended.

Existing context

• Resume the dashboard against /api/evolve/lineages (Step 5) to watch lineage divergence live.
• Reading evolve state: data/evolve_run_state.json (generations_promoted, parent_current) + data/evolve_results.jsonl tail (per .claude/rules/evolve.md).
• This is the §15 observation phase: it precedes any future "refactor/generalize" work so there's a known-good multi-lineage baseline.

Files to modify/create

• documentation/soak-test-runs/evolution-lines-soak-<ts>.md (operator-authored soak record — not code)

Done when

≥2 generations per lineage complete; the diversity matrix shows non-trivial separation; extinction logic exercised at least once OR documented as not-triggered-by-design under the cap; zero orphan processes; commit log shows clean [evo-auto] rows. Soak record written. Evaluate against the umbrella kill criterion (lineages converging to identical fingerprints with no exploration gain → pause the phase).

Depends on

Step 6 (#278) — smoke gate must pass first.

Parallel-safe with

none — strictly sequential: this is the terminal observation soak; the smoke gate (Step 6) must be green before committing wall-clock.

Produces

documentation/soak-test-runs/evolution-lines-soak-<ts>.md.

Operator workflow notes (autonomous + UI-bundle + parallel)

• Type: wait — /build-phase halts intentionally so wall-clock waiting doesn't consume context. Resume in a fresh session after the soak.
• Observation-based success criteria — even "the loop died after N minutes for reason X" is a valuable finding worth shipping.

---

Synced from documentation/plans/evolution-lines-plan.md by /repo-sync at 154f7b0