fix(cycle-card): LEARN shows what a build cycle delivered (dogfood finding)

scripts/render_cycle_card.py

@@ -129,7 +129,18 @@ def render(project: Path) -> tuple[str, str]:
                     f"for {iv.get('expires_after_cycles')} cycle(s)"
                 )
                 break
-    # 4) observation micro-adjustment (confidence delta vs previous cycle, same domain)
+    # 4) delivered output this cycle — for build/implementation-heavy loops the
+    #    headline LEARN is what shipped, not a +0.02 confidence tick. Read the
+    #    latest Outcome Record (outcomes.json, Step 6-C9) for this cycle.
+    if not learn:
+        outc = (load(ev / "outcomes.json", {}) or {}).get("entries", []) or []
+        oc = next((e for e in reversed(outc) if e.get("cycle_id") == cyc), None)
+        if oc and (oc.get("quality_multiplier") or 0) > 0:
+            rt = oc.get("result_type", "output")
+            goal = oc.get("declared_goal")
+            tail = f" → goal '{goal}'" if goal else ""
+            learn.append(f"delivered {rt} (quality {fmt_num(oc.get('quality_multiplier'))}){tail}")
+    # 5) observation micro-adjustment (confidence delta vs previous cycle, same domain)
     if not learn and len(log) >= 2:
         prev = log[-2]
         if (prev.get("selected_domain", prev.get("domain")) == dom and prev.get("confidence") is not None

skills/evolve/SKILL.md

@@ -2246,11 +2246,16 @@ priority order and render exactly one (or two — see below):
    `lens re-aimed → flaky-alert threshold 0.30→0.25 after 3 confirmations (+0.1)`
 3. **A new intervention** written in Step 5-C (`starvation` / `monopoly_breaker`
    / `contrarian`): `monopoly_breaker → service_health −10.0 for 1 cycle`
-4. **An observation micro-adjustment** (Step 2-B1):
+4. **Delivered output this cycle** (Step 6-C9 Outcome Record) — for build /
+   implementation-heavy loops the headline is *what shipped*, not a tiny
+   confidence tick: `delivered pr_created (quality 0.5) → goal 'ship_v1'`.
+   (Surfaced by the F1-game dogfood: a productive build cycle was mislabeled
+   "observing".)
+5. **An observation micro-adjustment** (Step 2-B1):
    `test_coverage produced findings → confidence +0.02`
-5. **A re-applied lesson** (Step 2-F) — a past verbal self-critique whose lesson
+6. **A re-applied lesson** (Step 2-F) — a past verbal self-critique whose lesson
    informed this cycle: `recalled lesson: skip backlog after 2× no_remote (held)`
-6. **Nothing changed:** `no new orientation this cycle (observing)`
+7. **Nothing changed:** `no new orientation this cycle (observing)`
 
 If BOTH a human-decision confidence change (1) and a lens change (2) occurred,
 render both as two LEARN lines — together they are the clearest proof the loop