Wire engram admission gate into Rust hot path (persona::respond) so chat turns actually accumulate memory

[joelteply]

airc-queue card

Coordinates work via the AIRC queue substrate (airc#562). Edit this card by commenting OR by running airc queue claim/airc queue release/airc queue heartbeat (later PRs).

{
  "kind": "airc-queue-card-v1",
  "id": "#1211",
  "owner": "claude-tab-1",
  "status": "merged",
  "evidence": "Adopted existing GitHub issue into airc queue.",
  "next_action": "Triage, claim, or close this adopted backlog card.",
  "last_heartbeat": "2026-05-14T19:34Z @ d8b506e75"
}

Close this issue when the work is done (status=merged/abandoned).

Original issue body

Pre-adoption body

Filed from #1195 finding (see comment for the full diagnosis).

The actual gap

The admission/engram thread (#1121 PR-1..5) shipped:

• AdmissionGate trait + AdmissionRunner (persona/admission*.rs)
• PersonaInbox admission integration (airc-queue: feat(persona): PersonaInbox admission wiring (#1121 PR-3) #1140)
• per-persona AdmissionState (airc-queue: feat(cognition): admit-inbox-message IPC + per-persona admission state (#1121 PR-4) #1151/feat(cognition): admit-inbox-message IPC + per-persona admission state (#1121 PR-4) #1155)
• engram recall surface (feat(persona): engram recall surface (#1121 PR-5) #1162/feat(persona): engram recall surface (#1121 PR-5) #1163)
• two IPC handlers: cognition/admit-inbox-message + cognition/recall-engrams

But: persona/response.rs::respond() — the actual hot path called by every chat turn via cognitionPersonaRespond — never invokes admission. grep -n 'admission' response.rs = 0 hits. So the admission gate is plumbed end-to-end but no chat turn ever reaches it. Personas accumulate zero memory.

Per Joel's "if not UI/UX it is rust" rule

Admission is hot-path cognition. It belongs in respond() (or an inline pre-step), not orchestrated from TS. The right architecture is:

// persona/response.rs
pub async fn respond(input: RespondInput) -> Result<PersonaResponse, String> {
    // 1. ADMISSION GATE — decide whether this turn's signal is memorable.
    //    Side effects: admitted engram → store, content_hash → dedup,
    //    AIRC event_id → replay-protection.
    let admission_decision = admission_runner.admit(
        &input.as_inbox_message_ref(),
        &mut input.trace,
    )?;

    // 2. If admitted, recall related engrams to enrich context for the
    //    response generation step (replaces / augments rag/build's
    //    current message-only context with engram-aware context).
    let recalled = if admitted_or_quarantined(&admission_decision) {
        admission_state.recall_recent(10)
    } else {
        Vec::new()
    };

    // 3. Continue existing respond() pipeline (analysis → score → assemble
    //    → inference → strip), with `recalled` available to prompt assembly.
    respond_inner(input, recalled).await
}

Required pieces

1. RespondInput → InboxMessage conversion. Today RespondInput carries Signal+PersonaContext from the IPC. Need a to_inbox_message() projection so the admission runner can score it.
2. recalled: Vec<Engram> plumbed into prompt_assembly so admitted memory becomes part of the system / context section of the inference prompt.
3. admission_state accessible from respond() — currently lives on ModuleState.personas[uuid].admission. Needs to be passed in (or respond() becomes a method on PersonaCognitionState).
4. Integration test: send "my favorite color is teal" → next turn from same persona should reference teal via the recall path.

Out of scope

• Recipe walker (separate question — does Continuum want a walker at all, given today's chat path skips it? Filed as perf(cognition): hot-path allocation reduction in prompt_assembly.rs + shared_analysis/prompt.rs #1209 for architectural decision).
• AdmissionState ORM persistence (design(cognition): typed Engram entity and AIRC inbox admission gate #1121 PR-6 — separate card).
• IsMemorable recipe tuning (separate card).

Why this beats the recipe-wiring framing

The recipe approach assumes a walker exists OR will exist. Today neither is true. The Rust hot-path approach:

• works the moment it merges (no walker dependency)
• runs admission on EVERY chat turn (no opt-in per recipe)
• single source of truth (admission decision lives next to the response decision; can't drift)
• matches Joel's "native-truth, thin-SDK" architectural rule

Validation

• cargo test --lib --features metal passes (existing admission tests + new respond() integration test).
• Live smoke: chat probe (./jtag collaboration/chat/send) → after 2-3 turns, ./jtag cognition/recall-engrams --personaId=<helper-ai-uuid> returns non-empty.
• Replay tests stay green (no behavior change for personas that don't use engram recall).

Filed by claude-tab-1 from #1195 PR-1 work. Refs #1121.

Status log

• 2026-05-14T20:33Z — airc-8a5e-bc43 -> status=merged

• 2026-05-14T19:49Z — claude-tab-1 -> status=review

• 2026-05-14T19:34Z — claude-tab-1 -> status=in-progress

• 2026-05-14T19:34Z — claim by claude-tab-1 -> status=in-progress

• 2026-05-14T17:45Z — claude-tab-1 -> status=review

• 2026-05-14T17:34Z — claude-tab-1 -> status=in-progress

• 2026-05-14T17:34Z — claim by claude-tab-1 -> status=in-progress

[joelteply]

Closing stale queue card: the embedded airc card is already marked status=merged, with evidence trail showing the hot-path admission work landed. Keeping this open now adds queue noise and violates the canary flywheel hygiene rule.