agent-optimize: require merging disjoint-cluster accepted candidates before finalize

[OsherElhadad]

Summary

• merge_search.py (agent-optimize Phase 4: merge-as-graph-search across round survivors #438) already exists to combine disjoint-cluster survivors via integrate.py/funcmerge.py, but nothing required the optimizer's agent session to actually invoke it before sealing test — so a run's reported "best" could be whichever single-cluster fix happened to score highest, never a genuine combination of accepted candidates that each fixed a different subset of tasks.
• SKILL.md / references/algorithm.md: makes running merge_search.py on 2+ accepted disjoint-cluster candidates (then gating the merge like any candidate) an explicit, required step before any end-of-run measurement, not just an available tool nobody reaches for under time pressure.
• merge_search.check_merge_compliance(): a new function that reads graph.jsonl for accepted candidates with pairwise-disjoint target task ids and checks whether any edit_kind == "merge" node exists anywhere in the graph — the same kind of code-level audit signal round.py's existing agent_optimize_compliance event provides for the screen ladder. It never blocks or forces a decision — this is an agent-driven framework where host.py owns no algorithm decisions, so enforcement stays a prompt-level requirement; this is only the observability half.
• measure.py runs the check at finalize time and logs a merge_compliance_warning event to events.jsonl when tripped, so the omission is visible in the dashboard's activity log (every event renders there regardless of kind) exactly like other compliance signals already are.

Test plan

• core/tests/test_merge_compliance_check.py — new unit tests: positive case (2+ accepted disjoint-cluster candidates, no merge attempt anywhere in the graph → flagged), and negative cases (a merge was already attempted; fewer than 2 accepted candidates; accepted clusters overlap; accepted candidates have no recorded target ids at all).
• python -m pytest core/tests/ -q — full suite, 1243 passed, 12 skipped, no regressions.
• python skills/_registry/lint_skills.py skills — no authoring violations; agent-optimize's SKILL.md body stays under the 5000-token budget.

[skillberry-bot]

❌ Automatic Labeling Failed

An error occurred while trying to automatically label this pull request. Please check the workflow logs for details and add labels manually.