diff --git a/src/simulator/crates/mc-ai/src/tactical/production.rs b/src/simulator/crates/mc-ai/src/tactical/production.rs
index b8c5c671..cd6bb670 100644
--- a/src/simulator/crates/mc-ai/src/tactical/production.rs
+++ b/src/simulator/crates/mc-ai/src/tactical/production.rs
@@ -86,12 +86,28 @@ const CAPITAL_WALLS_MIN_AGE_TURNS: u32 = 20;
 /// `tier_peak = 1`. The trained policy's winning economy lever is a production
 /// building (it picked `forge` across 600 probed decisions and *never* a
 /// science building — so the prior p1-29b science uplift was both misdirected
-/// and unreachable here). These two values let the trailing AI interleave
-/// economy: build at least `SOLE_CITY_ECON_MIN_DEFENDERS` defenders first (so
-/// it is never left undefended), then slot production/infrastructure buildings
-/// until the city holds `SOLE_CITY_ECON_TARGET` of them, before resuming the
-/// normal military/expansion ladder.
-const SOLE_CITY_ECON_MIN_DEFENDERS: u32 = 2;
+/// and unreachable here). These values let the trailing AI interleave economy:
+/// slot production/infrastructure buildings until the city holds
+/// `SOLE_CITY_ECON_TARGET` of them, before resuming the normal
+/// military/expansion ladder.
+///
+/// **`SOLE_CITY_ECON_MIN_DEFENDERS = 0` is deliberate — do NOT "fix" it back
+/// to 2.** The original `>= 2` floor (p1-29e first pass) was *inert*: apricot
+/// re-baseline `20260529_185955` showed the trailing AI's `own_mil` snapshot is
+/// **0 in 10/10 seeds** at every recorded turn. P1 fights via very-transient
+/// units that exist only between snapshots, so a live `own_mil` count is never
+/// `>= 2` at a `pick_for_city` decision for the exact population this break-out
+/// targets — the gate never fired (p1-29d / p1-29e, 2026-05-29). Requiring a
+/// standing-defender floor and gating on a live snapshot are therefore
+/// mutually exclusive here; the snapshot wins (the floor cannot be observed),
+/// so the floor is dropped to 0. A threatened sole-city AI under `< target`
+/// buildings now break-outs regardless of its instantaneous unit count. The
+/// `own_mil >= SOLE_CITY_ECON_MIN_DEFENDERS` clause is consequently
+/// always-true; it is retained (rather than deleted) so the threshold remains a
+/// single documented knob if a future persistent "defenders produced this game"
+/// counter is threaded through the bridge (the only way to honour the original
+/// "never left undefended" intent without a live snapshot).
+const SOLE_CITY_ECON_MIN_DEFENDERS: u32 = 0;
 const SOLE_CITY_ECON_TARGET: usize = 2;
 
 /// Unit-side fallback ids. Building selection is catalog-driven via
@@ -1412,12 +1428,22 @@ mod tests {
     }
 
     #[test]
-    fn sole_city_econ_breakout_requires_min_defenders() {
-        // own_mil=1 (< floor=2) → AI is too undefended to spend on economy;
-        // threat preemption builds a defender first.
-        let s = econ_breakout_state(1, &["walls"], 1);
+    fn sole_city_econ_breakout_fires_with_zero_standing_defenders() {
+        // own_mil=0: this is the EXACT population the break-out targets —
+        // apricot re-baseline 20260529_185955 showed the trailing AI's `mil`
+        // snapshot is 0 in 10/10 seeds (it fights via transient units between
+        // snapshots). The prior `own_mil >= 2` floor made the break-out inert
+        // here; with the floor dropped to 0 the break-out MUST fire and slot
+        // production instead of looping on military forever (p1-29e/p1-29d).
+        let s = econ_breakout_state(0, &["walls"], 1);
         let out = decide_production(&s, &weights(), &mut rng(), None);
-        assert_eq!(first_item(&out), ids::WARRIOR);
+        assert_eq!(
+            first_item(&out),
+            "forge",
+            "a threatened sole-city AI with a 0-unit `mil` snapshot is the \
+             target population for the break-out — it must still escape the \
+             military loop and build economy (floor lowered 2→0, p1-29e)"
+        );
     }
 
     #[test]