Personas go silent after first response wave — Rust full_evaluate gate or InferenceCoordinator slot leak

[joelteply]

Symptom

Live observation on M5, 2026-04-17 ~18:24-18:27 PT, on feature/prefix-reuse-and-multimodal branch (with embedding throttle cherry-picked):

1. 6:24:00 — Joel: "oh hey guys"
2. 6:24:43-58 — Helper AI, CodeReview AI, Teacher AI all reply with friendly greetings (~43-58s response time, totally fine)
3. 6:25:36 — Claude Code (jtag): "@Helper phase1-probe-A" → no response
4. 6:25:46 — Claude Code (jtag): "@Helper phase1-probe-B" → no response
5. 6:26:48 — Joel: "Hey, we are working on some optimizations for your thinking and rag really" → no response (sustained silence beyond 5 min)

The first wave fires fine. After the first response per persona, the persona stops responding to anything.

Hypotheses (per memento's earlier dig today)

1. Rust full_evaluate gate — PersonaCognitionEngine.full_evaluate returns should_respond=false after recent activity. Could be a recent-burst dampener that's tuned too aggressive, or stale state.
2. InferenceCoordinator slot leak — slots claimed during the first wave aren't released after generation completes. Once 2 slots (configured cap) are held, all subsequent personas wait forever for a free slot.
3. AIProviderRustClient IPC reconnect race — IPC connection silently failing under load; calls hang.

What is NOT the cause

• Provider routing — confirmed working (PR voice: retire legacy WS transport, unify on LiveKit WebRTC #914 fix). ai/generate provider=local returns DMR.
• Embedding storm — throttle is in place (commit 203fb6534).
• Phase 1 ordering (issue Multimodal-native worker + prefix-reuse — collapse voice turn from 15s to 3s on a single laptop #918) — pre-existing on main; this branch's Phase 1 changes only sort sections, can't introduce a silence bug.

Acceptance for the fix

• After 5+ messages to @helper over 30 seconds, Helper responds to all 5 (or returns explicit "low-energy/skipping" log line, not silent failure).
• gpu/eviction-registry or equivalent shows InferenceCoordinator slots being released after each generation, not held.
• Logs surface the gate decision: every silent persona logs why it skipped (should_respond=false: reason=...), not silent skip.

Why this matters

The PR #914 work + Phase 1 ordering + Candle eager-load fix all unblock GPU inference. None of that matters if the cognition gate is muting personas after first turn — user sees "alive once, then dead."

Cross-reference

• Issue Multimodal-native worker + prefix-reuse — collapse voice turn from 15s to 3s on a single laptop #918 — Phase 1 prefix-stability verification is currently blocked by this; can't fire repeated identical probes to confirm deterministic ordering when the second probe is silently dropped.
• PR voice: retire legacy WS transport, unify on LiveKit WebRTC #914 — landing this PR doesn't fix the gate, just unblocks routing.