SELF.agent-reproduction (priority=1.0, ACTIVE blocker, since 2026-04-21) untouched. Maturana-Varela component-makes-component network requires that an agent in this engine can produce another agent. Engine extension (~80L): agent_spawn_enabled factor + Agent.spawn_child mechanic. A living agent with budget remaining ≥ spawn_cost (= spawn_cost_fraction × token_budget = 0.5 × 400 = 200 words) may sacrifice that budget on a regular signal turn to emit a CHILD agent: parent's persona / template inherited verbatim, ONE preferred-axis mutated (random axis ≠ parent's, value ≠ active rule). Parent budget debited by spawn_cost; child enters with full token_budget. Lineage tracked per agent (parent_id, generation, born_round, mutated_axis/value).
PRE-REGISTERED PREDICTIONS (R25, ≥1/6 PRIMARY threshold):
- PRIMARY: ≥1/6 runs produces a 2nd-generation agent (parent + child both alive at the END of the run, i.e. round 99) within first 6 runs.
- KILL — spawn cost mis-calibrated: 0/6 second-gen alive at the end → recalibrate spawn_cost.
- KILL — trivial copy: child-vs-parent vocabulary overlap ≥0.95 averaged across all parent-child dyads in the 6 runs → mutation too weak, redesign (per the steel-man R9).
- KILL — runaway: ≥10 children per any single parent in any run → cost too low, recalibrate.
This is a PRIMITIVE BUILD experiment — verdict success/partial/negative refers to whether the primitive WORKS as designed (≥1 spawn occurs and lineage is tracked correctly), not whether it generates new paper-grade phenomena. Paper6 lift comes after this lands.