feat(continual-learning): activate negative-feedback loop (RSI steps 1-2)

Step 1 — policy.ts shouldIntervene: suppress on a prior dismissal alone.
The former `&& !priorOutcome.wasRealCollision` term was dead code (outcomes
record wasRealCollision hardcoded true), so the 85 real dismissals in Atlas
were never used. Now a prior accepted===false suppresses the next identical
nudge. Spec: continual-learning/policy.md:41, spec.md:163.

Step 2 — podman.ts handle: only escalate severity to 'critical' (the spoken
alert trigger) when the recalled prior was an accepted *real* collision,
instead of blanket-escalating every recall. Surfaces learned routing in
preferredAction; stops dismissed/false priors over-escalating to voice.
Spec: continual-learning/policy.md:62-63, plan.md:66.

Documentation-first: adds PLAN.md section 8 "P0.5 - RSI negative-feedback
activation" with both rungs + follow-ups. No schema change. backend typecheck
passes. Independent of the MongoDB-cleanup handoff (Codex).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

backend/src/agent/podman.ts

@@ -100,7 +100,14 @@ export class PodMan {
     if (this.activeConflicts.has(key)) return; // single-shot: already voiced, still unresolved
 
     const prior = await recallSimilar(collision); // Loop A: exact/vector recall raises confidence
-    if (prior) collision.severity = 'critical';
+    // Only escalate to critical (which triggers the spoken alert) when the
+    // recalled prior was an *accepted real* collision. Blanket-escalating every
+    // recall — including dismissed/false-positive priors — masked the learned
+    // routing in preferredAction and made recalled noise scream "CRITICAL".
+    // (RSI Step 2 — continual-learning/policy.md:62-63, plan.md:66)
+    if (prior?.priorOutcome?.accepted && prior?.priorOutcome?.wasRealCollision) {
+      collision.severity = 'critical';
+    }
     if (!shouldIntervene(collision, prior)) return; // Loop B: policy gate
 
     this.activeConflicts.add(key); // claim now we're alerting; re-armed in onScreenFrame on resolution

backend/src/memory/policy.ts

@@ -12,7 +12,12 @@ export function shouldIntervene(collision: Collision, prior: RecalledCollision |
   if (collision.severity === 'info') return false;
 
   const priorOutcome = prior?.priorOutcome;
-  if (priorOutcome && !priorOutcome.accepted && !priorOutcome.wasRealCollision) return false;
+  // Suppress when the identical prior was dismissed (accepted === false). The
+  // former `&& !priorOutcome.wasRealCollision` term was dead code: outcomes are
+  // recorded with wasRealCollision hardcoded true, so the gate never fired and
+  // the 85 real dismissals in Atlas were ignored. Dismissals are the negative
+  // signal per continual-learning/policy.md:41 + spec.md:163. (RSI Step 1)
+  if (priorOutcome && !priorOutcome.accepted) return false;
 
   const cooldown = cooldownMs();
   const last = lastNudgeByPod.get(collision.podId) ?? 0;