Fix engram search latency: pin embed model, cache query embeddings, bound activate BFS

Pins the Ollama embed model resident (keep_alive:-1) to avoid multi-second
cold reloads whenever a larger generation model evicts it under unified-
memory pressure (measured cold reload up to ~2.2s vs ~0.02-0.05s warm).

Adds a direct-mapped query-embedding cache (FNV-1a keyed, full strcmp to
reject collisions) so a repeated query costs zero Ollama round-trips —
directly serves the curiosity loop, which reseeds the same query terms
repeatedly.

Replaces engram_activate's unbounded FIFO frontier BFS with a beam-capped,
level-synchronous BFS (default beam 128, tunable via
ENGRAM_ACTIVATE_BEAM) to bound per-hop hub-node explosion that could
previously reach multi-second/crash territory at depth 2-3.

Excludes an inert engram_prune_telemetry build-enabler stub that was only
needed to link this checkout against a newer integration branch — not part
of the fix.

lang/el-compiler/runtime/el_runtime.c

@@ -7056,10 +7056,16 @@ static float* engram_embed_raw(const char* prefix, const char* text, int* out_di
     char* esc = engram_json_escape(text);
     free(trunc);
     if (!esc || !esc_prefix) { free(esc); free(esc_prefix); return NULL; }
-    size_t blen = strlen(esc) + strlen(esc_prefix) + strlen(model) + 64;
+    size_t blen = strlen(esc) + strlen(esc_prefix) + strlen(model) + 96;
     char* body = malloc(blen);
     if (!body) { free(esc); free(esc_prefix); return NULL; }
-    snprintf(body, blen, "{\"model\":\"%s\",\"prompt\":\"%s%s\"}", model, esc_prefix, esc);
+    /* keep_alive:-1 pins the embed model resident in Ollama indefinitely.
+     * Without it the tiny embed model is evicted whenever a large generation
+     * model loads (unified-memory pressure), so the NEXT search pays a cold
+     * model reload — the dominant search-latency cost (measured cold reload
+     * up to ~2.2s vs ~0.02-0.05s warm). Pinning makes cold reload impossible. */
+    snprintf(body, blen, "{\"model\":\"%s\",\"keep_alive\":-1,\"prompt\":\"%s%s\"}",
+             model, esc_prefix, esc);
     free(esc); free(esc_prefix);
 
     CURL* c = curl_easy_init();
@@ -7099,11 +7105,52 @@ static int engram_semantic_enabled(void) {
     g_emb_state = -1; return 0;
 }
 
+/* ── Query-embedding cache ──────────────────────────────────────────────────
+ * The node embeddings are cached (engram_node_vec) but the QUERY was re-embedded
+ * on every search/activate call — a blocking Ollama round-trip each time. Query
+ * embeddings are deterministic for a given model, so we cache them keyed by an
+ * FNV-1a hash of the query string (with a full strcmp to reject hash
+ * collisions). A repeated query then costs zero network round-trips. This makes
+ * warm search latency independent of Ollama entirely, and directly serves the
+ * curiosity loop, which reseeds the same query terms repeatedly. Direct-mapped,
+ * fixed-size, process-lifetime. */
+#define ENGRAM_QCACHE_SIZE 1024
+typedef struct { char* q; uint64_t hash; float* vec; int dim; } EngramQCacheEntry;
+static EngramQCacheEntry g_qcache[ENGRAM_QCACHE_SIZE];
+
+/* Returns a malloc'd COPY of the cached vector (caller frees), or NULL on miss —
+ * preserving engram_embed_query's "caller frees" contract. */
+static float* engram_qcache_get(const char* q, uint64_t h, int* dim) {
+    EngramQCacheEntry* e = &g_qcache[h & (ENGRAM_QCACHE_SIZE - 1)];
+    if (e->vec && e->hash == h && e->q && strcmp(e->q, q) == 0 && e->dim > 0) {
+        float* copy = malloc((size_t)e->dim * sizeof(float));
+        if (!copy) return NULL;
+        memcpy(copy, e->vec, (size_t)e->dim * sizeof(float));
+        *dim = e->dim; return copy;
+    }
+    return NULL;
+}
+static void engram_qcache_put(const char* q, uint64_t h, const float* vec, int dim) {
+    if (!vec || dim <= 0) return;
+    EngramQCacheEntry* e = &g_qcache[h & (ENGRAM_QCACHE_SIZE - 1)];
+    float* stored = malloc((size_t)dim * sizeof(float));
+    char*  qcopy  = el_strdup(q);
+    if (!stored || !qcopy) { free(stored); free(qcopy); return; }
+    memcpy(stored, vec, (size_t)dim * sizeof(float));
+    free(e->q); free(e->vec);           /* evict prior occupant of this slot */
+    e->q = qcopy; e->hash = h; e->vec = stored; e->dim = dim;
+}
+
 /* Embed the query. Returns malloc'd vec (caller frees), or NULL if semantic off. */
 static float* engram_embed_query(const char* q, int* dim) {
     if (!engram_semantic_enabled()) return NULL;
     if (!q || !*q) return NULL;
-    return engram_embed_raw("search_query: ", q, dim);
+    uint64_t h = engram_fnv1a(q);
+    float* hit = engram_qcache_get(q, h, dim);
+    if (hit) return hit;
+    float* v = engram_embed_raw("search_query: ", q, dim);
+    if (v && *dim > 0) engram_qcache_put(q, h, v, *dim);
+    return v;
 }
 
 /* Cached node embedding. Returns a pointer OWNED BY THE CACHE — do not free. */
@@ -7537,6 +7584,39 @@ static double engram_goal_bias(const EngramNode* n, const char* query) {
     return bias;
 }
 
+
+/* ── Beam cap for engram_activate spreading activation ──────────────────────
+ * Bounds the number of frontier nodes expanded PER HOP. Without it a single
+ * high-degree hub enqueues thousands of successors, each re-scanning the whole
+ * edge list, and dense cycles re-enqueue them repeatedly — so capping DEPTH
+ * does not bound work (measured: depth-2/3 in the multi-second range, depth-3
+ * can crash). With the cap, only the top-BEAM highest-activation nodes at each
+ * level spread further. Every reached node is still recorded and returned, so
+ * recall is preserved — the cap bounds only associative spread, never the
+ * direct seed matches or the reported set. Tunable via ENGRAM_ACTIVATE_BEAM
+ * (default 128); set very high to restore unbounded behaviour. */
+static int64_t engram_activate_beam(void) {
+    static int64_t v = -1;
+    if (v >= 0) return v;
+    const char* s = getenv("ENGRAM_ACTIVATE_BEAM");
+    int64_t d = 128;
+    if (s && *s) { char* e = NULL; long t = strtol(s, &e, 10); if (e != s && t > 0) d = (int64_t)t; }
+    v = d; return v;
+}
+
+/* Partition the k highest-`score` entries of idx[0..n) to the front (order
+ * within the top-k is unspecified). O(k*n) partial selection — k is the small
+ * beam width, so this is cheap relative to a hop's edge scan. */
+static void engram_beam_select(int64_t* idx, int64_t n, int64_t k, const double* score) {
+    if (k >= n) return;
+    for (int64_t i = 0; i < k; i++) {
+        int64_t best = i;
+        for (int64_t j = i + 1; j < n; j++)
+            if (score[idx[j]] > score[idx[best]]) best = j;
+        if (best != i) { int64_t t = idx[i]; idx[i] = idx[best]; idx[best] = t; }
+    }
+}
+
 el_val_t engram_activate(el_val_t query, el_val_t depth) {
     EngramStore* g = engram_get();
     const char* q = EL_CSTR(query);
@@ -7606,53 +7686,65 @@ el_val_t engram_activate(el_val_t query, el_val_t depth) {
         for (int64_t s = 1; s < seed_count; s++)
             seed_epoch = (seed_epoch + seeds[s].created_at) / 2;
     }
-    typedef struct { int64_t idx; int64_t hops; double act; } Frontier;
-    Frontier* fr = malloc((size_t)(g->node_count * (max_depth + 1)) * sizeof(Frontier) + 16 * sizeof(Frontier));
-    if (!fr) {
+    /* ── Beam-capped, level-synchronous BFS ────────────────────────────────
+     * Expand the graph hop-by-hop; at each hop expand only the top-`beam`
+     * nodes by current best background activation (engram_beam_select). This
+     * replaces the old unbounded FIFO frontier, which let a hub enqueue
+     * thousands of successors and dense cycles re-enqueue them without limit
+     * (the breadth explosion). `reached` / `best_bg` / `best_hops` keep the
+     * exact same meaning, so the downstream executive/override passes and the
+     * reported result set are unchanged — only how far weak spread propagates
+     * is bounded. `cur`/`nxt` hold node indices for this/next level; `in_nxt`
+     * dedups a node to at most one entry per level. */
+    const int64_t beam = engram_activate_beam();
+    const double SPREAD_DECAY = 0.7;
+    int64_t* cur    = malloc((size_t)g->node_count * sizeof(int64_t));
+    int64_t* nxt    = malloc((size_t)g->node_count * sizeof(int64_t));
+    int*     in_nxt = calloc((size_t)g->node_count, sizeof(int));
+    if (!cur || !nxt || !in_nxt) {
+        free(cur); free(nxt); free(in_nxt);
         free(best_bg); free(best_hops); free(reached); free(seeds); return out;
     }
-    int64_t fhead = 0, ftail = 0;
-    int64_t fcap = (int64_t)((size_t)(g->node_count * (max_depth + 1)) + 16);
-    for (int64_t s = 0; s < seed_count; s++) {
-        if (ftail >= fcap) break;
-        fr[ftail].idx  = seeds[s].idx;
-        fr[ftail].hops = 0;
-        fr[ftail].act  = seeds[s].act;
-        ftail++;
-    }
-    const double SPREAD_DECAY = 0.7;
-    while (fhead < ftail) {
-        Frontier f = fr[fhead++];
-        if (f.hops >= max_depth) continue;
-        const char* cur_id = g->nodes[f.idx].id;
-        for (int64_t ei = 0; ei < g->edge_count; ei++) {
-            EngramEdge* e = &g->edges[ei];
-            const char* other = NULL;
-            if (e->from_id && strcmp(e->from_id, cur_id) == 0) other = e->to_id;
-            else if (e->to_id && strcmp(e->to_id, cur_id) == 0) other = e->from_id;
-            else continue;
-            int64_t oi = engram_find_node_index(other);
-            if (oi < 0) continue;
-            EngramNode* on = &g->nodes[oi];
-            double tbonus  = engram_temporal_proximity_bonus(on->created_at, seed_epoch);
-            double tdecay  = engram_temporal_decay(on, now_ms);
-            double dampen  = engram_activation_dampen(on);
-            double new_act = f.act * e->weight * SPREAD_DECAY * (1.0 + tbonus)
-                           * tdecay * dampen;
-            int64_t new_hops = f.hops + 1;
-            if (!reached[oi] || new_act > best_bg[oi]) {
-                best_bg[oi]   = new_act;
-                best_hops[oi] = new_hops;
-                reached[oi]   = 1;
-                if (ftail < fcap) {
-                    fr[ftail].idx  = oi;
-                    fr[ftail].hops = new_hops;
-                    fr[ftail].act  = new_act;
-                    ftail++;
+    int64_t cur_n = 0;
+    for (int64_t s = 0; s < seed_count && cur_n < g->node_count; s++)
+        cur[cur_n++] = seeds[s].idx;
+    for (int64_t hop = 0; hop < max_depth && cur_n > 0; hop++) {
+        if (cur_n > beam) { engram_beam_select(cur, cur_n, beam, best_bg); cur_n = beam; }
+        int64_t nxt_n = 0;
+        for (int64_t ci = 0; ci < cur_n; ci++) {
+            int64_t fidx = cur[ci];
+            double  f_act = best_bg[fidx];
+            const char* cur_id = g->nodes[fidx].id;
+            for (int64_t ei = 0; ei < g->edge_count; ei++) {
+                EngramEdge* e = &g->edges[ei];
+                const char* other = NULL;
+                if (e->from_id && strcmp(e->from_id, cur_id) == 0) other = e->to_id;
+                else if (e->to_id && strcmp(e->to_id, cur_id) == 0) other = e->from_id;
+                else continue;
+                int64_t oi = engram_find_node_index(other);
+                if (oi < 0) continue;
+                EngramNode* on = &g->nodes[oi];
+                double tbonus  = engram_temporal_proximity_bonus(on->created_at, seed_epoch);
+                double tdecay  = engram_temporal_decay(on, now_ms);
+                double dampen  = engram_activation_dampen(on);
+                double new_act = f_act * e->weight * SPREAD_DECAY * (1.0 + tbonus)
+                               * tdecay * dampen;
+                if (!reached[oi] || new_act > best_bg[oi]) {
+                    best_bg[oi]   = new_act;
+                    best_hops[oi] = hop + 1;
+                    reached[oi]   = 1;
+                    if (!in_nxt[oi] && nxt_n < g->node_count) {
+                        in_nxt[oi] = 1;
+                        nxt[nxt_n++] = oi;
+                    }
                 }
             }
         }
+        for (int64_t k = 0; k < nxt_n; k++) in_nxt[nxt[k]] = 0;
+        int64_t* tmp = cur; cur = nxt; nxt = tmp;
+        cur_n = nxt_n;
     }
+    free(cur); free(nxt); free(in_nxt);
     /* Persist layer-1 background_activation to node store. */
     for (int64_t i = 0; i < g->node_count; i++) {
         g->nodes[i].background_activation = reached[i] ? best_bg[i] : 0.0;
@@ -7666,7 +7758,7 @@ el_val_t engram_activate(el_val_t query, el_val_t depth) {
      * memory weight cannot be silenced by attentional suppression. */
     double* inhibition = calloc((size_t)g->node_count, sizeof(double));
     if (!inhibition) {
-        free(best_bg); free(best_hops); free(reached); free(seeds); free(fr);
+        free(best_bg); free(best_hops); free(reached); free(seeds);
         return out;
     }
     for (int64_t ei = 0; ei < g->edge_count; ei++) {
@@ -7692,7 +7784,7 @@ el_val_t engram_activate(el_val_t query, el_val_t depth) {
     double* wm_weights = calloc((size_t)g->node_count, sizeof(double));
     if (!wm_weights) {
         free(best_bg); free(best_hops); free(reached); free(seeds);
-        free(fr); free(inhibition); return out;
+        free(inhibition); return out;
     }
     for (int64_t i = 0; i < g->node_count; i++) {
         if (!reached[i] || best_bg[i] <= 0.0) continue;
@@ -7762,7 +7854,7 @@ el_val_t engram_activate(el_val_t query, el_val_t depth) {
     int64_t rcount = 0;
     if (!results) {
         free(best_bg); free(best_hops); free(reached); free(seeds);
-        free(fr); free(inhibition); free(wm_weights); return out;
+        free(inhibition); free(wm_weights); return out;
     }
     for (int64_t i = 0; i < g->node_count; i++) {
         if (!reached[i]) continue;
@@ -7806,7 +7898,7 @@ el_val_t engram_activate(el_val_t query, el_val_t depth) {
         out = el_list_append(out, entry);
     }
     free(best_bg); free(best_hops); free(reached);
-    free(seeds); free(fr); free(inhibition); free(wm_weights); free(results);
+    free(seeds); free(inhibition); free(wm_weights); free(results);
     return out;
 }
 

peripheral/.gitignore

@@ -1,8 +0,0 @@
-# Build + runtime artifacts — never committed.
-bin/
-# Captured media (camera frames, mic audio) and syntheses. Raw streams stay
-# LOCAL and never egress — including into git.
-out/
-# Runtime consent + resume state (local, per-machine).
-.consent.json
-.resume.json

peripheral/README.md

@@ -1,80 +0,0 @@
-# peripheral — Neuron's I/O organ (own-core, local, consent-gated)
-
-The interface made physical. Two afferent senses in, one efferent voice out —
-all reached the way the agentic surface reaches any tool.
-
-```
-MIC    (hear)  afferent   device -> capture -> descriptor -> ingest -> geometry
-CAMERA (see)   afferent   device -> capture -> descriptor -> ingest -> scene-geometry
-SPEAKER(speak) efferent   render WAV -> PLAY ALOUD out the speaker
-```
-
-Closes the conversational loop: **hear (mic) -> understand (engram) -> speak (speaker)**.
-
-## Rails
-- **Own-core.** macOS-native only: AVFoundation (camera/mic), CoreAudio voice-
-  processing (AEC), afplay (speaker), ImageIO/CoreGraphics (frames), hand-rolled
-  DSP (WAV, LPC, formant synthesis). No cloud, no heavy deps.
-- **Local-only.** Raw streams are written to `out/` and never egress. `.gitignore`
-  keeps captured media out of git.
-- **Consent-gated (two locks).** A Neuron-level grant (`grant`/`revoke`) *and* the
-  OS TCC permission. Sensitive senses (camera/mic) fail closed without both.
-- **Disclosed.** Every device touch prints a `[peripheral]` line on stderr.
-
-## Build
-```
-swiftc -O -o bin/periph src/periph.swift \
-  -framework AVFoundation -framework CoreMedia -framework Foundation \
-  -framework CoreGraphics -framework ImageIO -framework CoreImage
-```
-
-## Commands
-```
-periph grant|revoke <camera|mic>          # Neuron-level consent
-periph status
-periph speak <file.wav>                   # SPEAK ALOUD (efferent)
-periph tone <out.wav> [hz] [sec]          # own-core WAV synth
-periph listen <sec> <out.wav>             # MIC capture (afferent), 16k mono
-periph see <out.jpg>                      # CAMERA one frame (afferent)
-periph feat-audio <wav> | feat-image <jpg>          # capture -> compact descriptor
-periph ingest-audio|ingest-image <file> <engramURL> # descriptor -> engram node (geometry)
-periph voiceprint <voice.wav>             # extract F0 + formants F1-F5
-periph imitate <voice.wav> <out.wav>      # speak back in that voice (LPC resynthesis)
-periph hear-imitate <sec> <out.wav>       # MIC -> signature -> imitate -> SPEAK ALOUD
-periph converse <manifest.json> [--authority F] [--barge-at S[:backchannel|:bargein]] [--resume] [--live-mic]
-```
-
-## The afferent metabolism
-A capture is never shipped raw. It becomes a **compact descriptor** — the afferent
-twin of the music instrument-signature:
-- audio -> `[seconds, sr, ch, rms, peak, zcr, centroid, F0]` (~2400-6000x smaller)
-- image -> `[w, h, meanRGB, brightness, 3x3 luminance grid]` (~400000x smaller)
-- voice -> `[F0, F1..F5, bandwidths]` (11 numbers)
-
-That descriptor is what the ingest organ (engram `POST /api/nodes`) turns into an
-embedded node = geometry.
-
-## Voice by imitation
-`voiceprint`/`imitate` are own-core LPC (autocorrelation + Levinson-Durbin, order
-16 @ 16 kHz), formant extraction from the LPC spectral envelope, and source-filter
-resynthesis (glottal impulse train at F0 through the all-pole formant filter). A
-voice is grabbed by ear as ~a dozen numbers and spoken back — **no training, no
-stolen voice.** Measured fidelity on real speech: resynthesized formants match the
-source within 2-3%. The full phoneme->formant path for *novel* sentences is the
-speech faculty's seam (`elp` audio surface profile); this engine provides the
-formant synthesis primitive it renders through.
-
-## Interruptibility (native turn-taking)
-`converse` plays the utterance as an ordered, salience-tagged **meaning-plan**
-while the mic listens (full-duplex, AEC on so it never barges in on its own voice):
-- **barge-in**: user speech -> pause on the spot (sample-accurate), not "finish the buffer."
-- **yield-or-hold**: a decision grounded in the current segment's salience + progress
-  + the interrupter's authority — YIELD (stop) or HOLD ("hang on, let me finish").
-- **backchannel** ("mm-hm"): brief/low -> keep going, resume seamlessly.
-- **resumable**: on yield the remaining plan persists (`.resume.json`); `--resume`
-  picks the thread back up ("as I was saying").
-
-Live full-duplex uses `--live-mic` (OS AEC). Injected `--barge-at` drives the
-decision loop deterministically for testing.
-```
-```

peripheral/src/periph.swift

@@ -1,939 +0,0 @@
-// periph.swift — Neuron's PERIPHERAL I/O organ (own-core, LOCAL, CONSENT-GATED).
-//
-// The interface made physical:
-//   MIC    (hear)  = afferent  : device -> capture -> [ingest -> geometry]
-//   CAMERA (see)   = afferent  : device -> capture -> [ingest -> scene-geometry]
-//   SPEAKER(speak) = efferent  : [render WAV] -> PLAY ALOUD out the speaker
-//
-// Rails: own-core (AVFoundation / CoreAudio / afplay — all ship with macOS),
-//        no cloud, no heavy deps, raw streams stay LOCAL and never egress,
-//        every device access is CONSENT-GATED and DISCLOSED.
-//
-// Full-duplex CONVERSE mode implements native interruptibility: while the
-// speaker plays the utterance (a persistent, segmented meaning-plan), the mic
-// listens; on user speech it interrupts instantly, then DECIDES yield-or-hold
-// grounded in the salience of what it is mid-saying, and can RESUME the thread.
-//
-// Build:  swiftc -O -o peripheral/bin/periph peripheral/src/periph.swift \
-//               -framework AVFoundation -framework CoreMedia -framework Foundation
-
-import Foundation
-import AVFoundation
-import CoreMedia
-import CoreGraphics
-import ImageIO
-import CoreImage
-
-// ----------------------------------------------------------------------------
-// Disclosure — every peripheral touch is announced on stderr. Nothing is silent.
-// ----------------------------------------------------------------------------
-func disclose(_ msg: String) {
-    FileHandle.standardError.write("  [peripheral] \(msg)\n".data(using: .utf8)!)
-}
-func emit(_ obj: [String: Any]) {  // machine-readable event on stdout (JSON line)
-    if let d = try? JSONSerialization.data(withJSONObject: obj),
-       let s = String(data: d, encoding: .utf8) {
-        print(s)
-    }
-}
-func die(_ msg: String) -> Never {
-    disclose("ERROR: \(msg)")
-    emit(["ok": false, "error": msg])
-    exit(1)
-}
-
-// ----------------------------------------------------------------------------
-// Consent store — Neuron's OWN gate, on top of the OS (TCC) gate. Two locks on
-// the sensitive senses. Persisted locally next to the binary's organ dir.
-// ----------------------------------------------------------------------------
-struct Consent {
-    static let path: String = {
-        let dir = ProcessInfo.processInfo.environment["PERIPH_HOME"]
-            ?? FileManager.default.currentDirectoryPath + "/peripheral"
-        return dir + "/.consent.json"
-    }()
-
-    static func load() -> [String: Bool] {
-        guard let d = FileManager.default.contents(atPath: path),
-              let o = try? JSONSerialization.jsonObject(with: d) as? [String: Bool]
-        else { return ["camera": false, "mic": false] }
-        return o
-    }
-    static func save(_ g: [String: Bool]) {
-        let d = try! JSONSerialization.data(withJSONObject: g, options: [.prettyPrinted])
-        try? d.write(to: URL(fileURLWithPath: path))
-    }
-    // Neuron-level gate. Sensitive senses (camera/mic) require an explicit grant.
-    static func require(_ device: String) {
-        let g = load()
-        if g[device] != true {
-            die("CONSENT DENIED for '\(device)'. The user has not granted this sense. " +
-                "Run:  periph grant \(device)   (raw streams stay local, never egress).")
-        }
-        disclose("consent OK (Neuron-level) for '\(device)' — local only, never egresses.")
-    }
-}
-
-// ----------------------------------------------------------------------------
-// OS (TCC) permission — the second lock. AVFoundation prompts the user the first
-// time; if denied, we fail cleanly rather than hang.
-// ----------------------------------------------------------------------------
-func requireOSAccess(_ media: AVMediaType, _ label: String) {
-    let status = AVCaptureDevice.authorizationStatus(for: media)
-    switch status {
-    case .authorized:
-        disclose("consent OK (OS/TCC) for \(label).")
-        return
-    case .notDetermined:
-        disclose("requesting OS permission for \(label) (first use) — user must grant...")
-        let sem = DispatchSemaphore(value: 0)
-        var ok = false
-        AVCaptureDevice.requestAccess(for: media) { granted in ok = granted; sem.signal() }
-        _ = sem.wait(timeout: .now() + 30)
-        if !ok { die("OS permission for \(label) was not granted.") }
-        disclose("consent OK (OS/TCC) for \(label).")
-    case .denied, .restricted:
-        die("OS permission for \(label) is DENIED in System Settings > Privacy. " +
-            "Grant it to the controlling terminal/app, then retry.")
-    @unknown default:
-        die("unknown OS permission state for \(label).")
-    }
-}
-
-// ----------------------------------------------------------------------------
-// Own-core WAV writer (16-bit PCM). No library — proves we own the medium.
-// ----------------------------------------------------------------------------
-func writeWav(_ url: URL, samples: [Int16], sampleRate: Int, channels: Int = 1) {
-    var data = Data()
-    func u32(_ v: UInt32) { var x = v.littleEndian; data.append(Data(bytes: &x, count: 4)) }
-    func u16(_ v: UInt16) { var x = v.littleEndian; data.append(Data(bytes: &x, count: 2)) }
-    let bytesPerSample = 2
-    let dataBytes = samples.count * bytesPerSample
-    let byteRate = sampleRate * channels * bytesPerSample
-    data.append("RIFF".data(using: .ascii)!); u32(UInt32(36 + dataBytes))
-    data.append("WAVE".data(using: .ascii)!)
-    data.append("fmt ".data(using: .ascii)!); u32(16); u16(1); u16(UInt16(channels))
-    u32(UInt32(sampleRate)); u32(UInt32(byteRate))
-    u16(UInt16(channels * bytesPerSample)); u16(16)
-    data.append("data".data(using: .ascii)!); u32(UInt32(dataBytes))
-    for s in samples { var x = s.littleEndian; data.append(Data(bytes: &x, count: 2)) }
-    try? data.write(to: url)
-}
-
-// Read a WAV's basic geometry (own-core header parse). Walks chunks to find
-// 'fmt ' and 'data' — robust to JUNK/FLLR padding chunks (AVAudioRecorder emits them).
-func wavInfo(_ path: String) -> (sampleRate: Int, channels: Int, bits: Int, frames: Int)? {
-    guard let d = FileManager.default.contents(atPath: path), d.count > 44 else { return nil }
-    func rd16(_ o: Int) -> Int { Int(d[o]) | (Int(d[o+1]) << 8) }
-    func rd32(_ o: Int) -> Int { Int(d[o]) | (Int(d[o+1])<<8) | (Int(d[o+2])<<16) | (Int(d[o+3])<<24) }
-    var channels = 0, sampleRate = 0, bits = 0, dataSize = 0
-    var o = 12
-    while o + 8 <= d.count {
-        let id = String(bytes: d[o..<o+4], encoding: .ascii) ?? ""
-        let sz = rd32(o+4)
-        if id == "fmt " && o + 24 <= d.count {
-            channels = rd16(o+10); sampleRate = rd32(o+12); bits = rd16(o+22)
-        } else if id == "data" {
-            dataSize = min(sz, d.count - (o+8))
-        }
-        o += 8 + sz + (sz & 1)
-    }
-    let frames = (channels > 0 && bits > 0) ? dataSize / (channels * bits/8) : 0
-    return (sampleRate, channels, bits, frames)
-}
-
-// ----------------------------------------------------------------------------
-// SPEAKER (efferent) — play a WAV ALOUD. Own-core: afplay ships with macOS.
-// ----------------------------------------------------------------------------
-func speak(_ wavPath: String) {
-    guard FileManager.default.fileExists(atPath: wavPath) else { die("no such file: \(wavPath)") }
-    disclose("SPEAKER: playing '\(wavPath)' ALOUD out the local speaker (efferent).")
-    let p = Process()
-    p.executableURL = URL(fileURLWithPath: "/usr/bin/afplay")
-    p.arguments = [wavPath]
-    try? p.run(); p.waitUntilExit()
-    let ok = p.terminationStatus == 0
-    disclose(ok ? "SPEAKER: done — Neuron spoke aloud." : "SPEAKER: afplay failed.")
-    if let i = wavInfo(wavPath) {
-        emit(["ok": ok, "op": "speak", "file": wavPath, "played_aloud": ok,
-              "sample_rate": i.sampleRate, "channels": i.channels,
-              "seconds": Double(i.frames)/Double(max(i.sampleRate,1))])
-    } else {
-        emit(["ok": ok, "op": "speak", "file": wavPath, "played_aloud": ok])
-    }
-}
-
-// ----------------------------------------------------------------------------
-// MIC (afferent) — capture N seconds -> 16k mono 16-bit WAV (formant-ready).
-// ----------------------------------------------------------------------------
-func listen(seconds: Double, out: String) {
-    Consent.require("mic")
-    requireOSAccess(.audio, "microphone")
-    disclose("MIC: capturing \(seconds)s -> '\(out)' (16 kHz mono, LOCAL, never egresses).")
-
-    let url = URL(fileURLWithPath: out)
-    let settings: [String: Any] = [
-        AVFormatIDKey: kAudioFormatLinearPCM,
-        AVSampleRateKey: 16000.0,
-        AVNumberOfChannelsKey: 1,
-        AVLinearPCMBitDepthKey: 16,
-        AVLinearPCMIsFloatKey: false,
-        AVLinearPCMIsBigEndianKey: false,
-    ]
-    guard let rec = try? AVAudioRecorder(url: url, settings: settings) else {
-        die("could not open the microphone recorder.")
-    }
-    rec.record()
-    Thread.sleep(forTimeInterval: seconds)
-    rec.stop()
-    // let the file flush
-    Thread.sleep(forTimeInterval: 0.1)
-    if let i = wavInfo(out) {
-        disclose("MIC: captured \(i.frames) frames @ \(i.sampleRate)Hz — ready to hand to the ingest organ.")
-        emit(["ok": true, "op": "listen", "file": out, "sample_rate": i.sampleRate,
-              "channels": i.channels, "frames": i.frames,
-              "seconds": Double(i.frames)/Double(max(i.sampleRate,1)),
-              "next": "ingest -> phonetic/voice geometry"])
-    } else {
-        die("mic capture produced no readable WAV.")
-    }
-}
-
-// ----------------------------------------------------------------------------
-// CAMERA (afferent) — capture ONE frame -> JPEG on disk.
-// ----------------------------------------------------------------------------
-// Grab one video frame via AVCaptureVideoDataOutput (CLI-safe; no KVO/photo classes).
-final class FrameGrabber: NSObject, AVCaptureVideoDataOutputSampleBufferDelegate {
-    let sem = DispatchSemaphore(value: 0)
-    var cgImage: CGImage?
-    var seen = 0
-    let cictx = CIContext(options: nil)
-    func captureOutput(_ output: AVCaptureOutput, didOutput sampleBuffer: CMSampleBuffer,
-                       from connection: AVCaptureConnection) {
-        seen += 1
-        if cgImage != nil || seen < 5 { return }        // let exposure settle a few frames
-        guard let pb = CMSampleBufferGetImageBuffer(sampleBuffer) else { return }
-        let ci = CIImage(cvPixelBuffer: pb)
-        cgImage = cictx.createCGImage(ci, from: ci.extent)
-        sem.signal()
-    }
-}
-func see(out: String) {
-    Consent.require("camera")
-    requireOSAccess(.video, "camera")
-    disclose("CAMERA: capturing one frame -> '\(out)' (LOCAL, never egresses).")
-
-    let session = AVCaptureSession()
-    session.sessionPreset = .photo
-    guard let device = AVCaptureDevice.default(for: .video),
-          let input = try? AVCaptureDeviceInput(device: device),
-          session.canAddInput(input) else { die("no camera device available.") }
-    session.addInput(input)
-    let output = AVCaptureVideoDataOutput()
-    output.alwaysDiscardsLateVideoFrames = true
-    let grabber = FrameGrabber()
-    output.setSampleBufferDelegate(grabber, queue: DispatchQueue(label: "periph.cam"))
-    guard session.canAddOutput(output) else { die("cannot add video output.") }
-    session.addOutput(output)
-    session.startRunning()
-
-    if grabber.sem.wait(timeout: .now() + 10) == .timedOut { session.stopRunning(); die("camera capture timed out.") }
-    session.stopRunning()
-
-    guard let cg = grabber.cgImage,
-          let dst = CGImageDestinationCreateWithURL(URL(fileURLWithPath: out) as CFURL,
-                                                    "public.jpeg" as CFString, 1, nil)
-    else { die("camera returned no frame.") }
-    CGImageDestinationAddImage(dst, cg, nil)
-    guard CGImageDestinationFinalize(dst) else { die("could not write JPEG.") }
-    let bytes = ((try? FileManager.default.attributesOfItem(atPath: out))?[.size] as? Int) ?? 0
-    disclose("CAMERA: wrote \(cg.width)x\(cg.height) frame (\(bytes) bytes) — ready for scene-geometry ingest.")
-    emit(["ok": true, "op": "see", "file": out, "width": cg.width, "height": cg.height,
-          "bytes": bytes, "next": "ingest -> scene-geometry"])
-}
-
-// ============================================================================
-// FEAT — the afferent METABOLISM: a raw capture becomes a COMPACT descriptor
-// (a few dozen numbers), the mirror of the efferent signature. This is what
-// gets handed to the ingest organ as geometry — NOT the raw stream. Own-core.
-// ============================================================================
-
-// Read all 16-bit PCM samples from a WAV (own-core).
-func readWavSamples(_ path: String) -> (samples: [Double], sr: Int, ch: Int)? {
-    guard let d = FileManager.default.contents(atPath: path), d.count > 44 else { return nil }
-    func rd16(_ o: Int) -> Int { Int(d[o]) | (Int(d[o+1]) << 8) }
-    func rd32(_ o: Int) -> Int { Int(d[o]) | (Int(d[o+1])<<8) | (Int(d[o+2])<<16) | (Int(d[o+3])<<24) }
-    var ch = 0, sr = 0, bits = 0
-    var o = 12
-    while o + 8 <= d.count {
-        let id = String(bytes: d[o..<o+4], encoding: .ascii) ?? ""
-        let sz = rd32(o+4)
-        if id == "fmt " && o + 24 <= d.count { ch = rd16(o+10); sr = rd32(o+12); bits = rd16(o+22) }
-        if id == "data" {
-            guard bits == 16, ch > 0 else { return nil }
-            var samples = [Double](); let start = o + 8
-            let end = min(start + sz, d.count - 1)
-            var i = start
-            while i + 1 < end {
-                var v = Int(rd16(i)); if v >= 32768 { v -= 65536 }
-                samples.append(Double(v) / 32768.0)
-                i += 2 * ch                       // take channel 0 if stereo
-            }
-            return (samples, sr, ch)
-        }
-        o += 8 + sz + (sz & 1)
-    }
-    return nil
-}
-
-// Audio descriptor = compact sound/voice signature (energy, ZCR, centroid, F0).
-// The seed for phonetic geometry + the hear->imitate voice-signature.
-func computeAudio(_ path: String) -> (content: String, vector: [Double], extra: [String: Any]) {
-    guard let (s, sr, ch) = readWavSamples(path), !s.isEmpty else { die("cannot read PCM from \(path)") }
-    let n = s.count
-    let seconds = Double(n) / Double(sr)
-    var sumsq = 0.0, peak = 0.0, zc = 0.0
-    for i in 0..<n {
-        sumsq += s[i]*s[i]; peak = max(peak, abs(s[i]))
-        if i > 0 && (s[i-1] < 0) != (s[i] < 0) { zc += 1 }
-    }
-    let rms = (sumsq / Double(n)).squareRoot()
-    let zcr = zc / Double(n) * Double(sr)          // ~2*dominant freq for tonal
-    // Spectral centroid via a coarse DFT on a mid window (own-core).
-    let W = min(2048, n); let off = max(0, (n - W)/2)
-    var num = 0.0, den = 0.0
-    let bins = 64
-    for k in 1..<bins {
-        let f = Double(k) * Double(sr) / Double(2*bins)
-        var re = 0.0, im = 0.0
-        for j in 0..<W {
-            let ang = -2*Double.pi*Double(k)*Double(j)/Double(2*bins)
-            re += s[off+j]*cos(ang); im += s[off+j]*sin(ang)
-        }
-        let mag = (re*re+im*im).squareRoot()
-        num += f*mag; den += mag
-    }
-    let centroid = den > 0 ? num/den : 0
-    // F0 via autocorrelation (voice pitch) over plausible speech range 70-400 Hz.
-    var bestLag = 0; var bestCorr = 0.0
-    let lagMin = sr/400, lagMax = min(sr/70, n-1)
-    if lagMax > lagMin {
-        for lag in lagMin...lagMax {
-            var c = 0.0
-            var i = 0; while i + lag < min(n, off+W) { c += s[off+i]*s[off+i+lag]; i += 1 }
-            if c > bestCorr { bestCorr = c; bestLag = lag }
-        }
-    }
-    let f0 = bestLag > 0 ? Double(sr)/Double(bestLag) : 0
-    let vector: [Double] = [seconds, Double(sr), Double(ch), rms, peak, zcr, centroid, f0]
-    let content = String(format:
-        "Heard sound (afferent, mic): %.2fs at %dHz. RMS energy %.3f, peak %.3f, " +
-        "zero-crossing rate %.0fHz, spectral centroid %.0fHz, estimated voice pitch F0 %.0fHz. " +
-        "Compact voice/sound signature (%d numbers) — phonetic geometry + hear-to-imitate seed.",
-        seconds, sr, rms, peak, zcr, centroid, f0, vector.count)
-    disclose("FEAT(audio): \(vector.count)-number signature vs \(n) raw samples (~\(n/max(vector.count,1))x compression).")
-    return (content, vector, ["f0_hz": f0, "centroid_hz": centroid, "zcr_hz": zcr,
-                             "rms": rms, "seconds": seconds, "raw_samples": n])
-}
-func featAudio(_ path: String) {
-    let r = computeAudio(path)
-    var out: [String: Any] = ["ok": true, "op": "feat-audio", "file": path,
-                              "vector": r.vector, "content": r.content,
-                              "ingest": ["node_type": "Observation", "tier": "Episodic", "content": r.content]]
-    r.extra.forEach { out[$0] = $1 }
-    emit(out)
-}
-
-// Image descriptor = compact scene-geometry (dims, brightness, region grid).
-func computeImage(_ path: String) -> (content: String, vector: [Double], extra: [String: Any]) {
-    guard let src = CGImageSourceCreateWithURL(URL(fileURLWithPath: path) as CFURL, nil),
-          let img = CGImageSourceCreateImageAtIndex(src, 0, nil) else { die("cannot decode image \(path)") }
-    let w = img.width, h = img.height
-    let cs = CGColorSpaceCreateDeviceRGB()
-    let bpr = w * 4
-    var buf = [UInt8](repeating: 0, count: h * bpr)
-    guard let ctx = CGContext(data: &buf, width: w, height: h, bitsPerComponent: 8,
-                              bytesPerRow: bpr, space: cs,
-                              bitmapInfo: CGImageAlphaInfo.premultipliedLast.rawValue) else {
-        die("cannot rasterize image")
-    }
-    ctx.draw(img, in: CGRect(x: 0, y: 0, width: w, height: h))
-    // 3x3 region average luminance + overall average color.
-    var rAvg = 0.0, gAvg = 0.0, bAvg = 0.0
-    var grid = [Double](repeating: 0, count: 9); var gridN = [Int](repeating: 0, count: 9)
-    let step = max(1, (w*h)/40000)             // subsample for speed
-    var count = 0; var idx = 0
-    while idx < w*h {
-        let x = idx % w, y = idx / w
-        let p = y*bpr + x*4
-        let r = Double(buf[p]), g = Double(buf[p+1]), b = Double(buf[p+2])
-        rAvg += r; gAvg += g; bAvg += b; count += 1
-        let cell = (min(2, y*3/h))*3 + min(2, x*3/w)
-        grid[cell] += 0.299*r + 0.587*g + 0.114*b; gridN[cell] += 1
-        idx += step
-    }
-    if count == 0 { die("no pixels sampled") }
-    rAvg /= Double(count); gAvg /= Double(count); bAvg /= Double(count)
-    for i in 0..<9 { grid[i] = gridN[i] > 0 ? grid[i]/Double(gridN[i]) : 0 }
-    let bright = (0.299*rAvg + 0.587*gAvg + 0.114*bAvg)/255.0
-    let vector = [Double(w), Double(h), rAvg/255, gAvg/255, bAvg/255, bright] + grid.map { $0/255 }
-    let content = String(format:
-        "Saw scene (afferent, camera): %dx%d frame. Mean color rgb(%.0f,%.0f,%.0f), " +
-        "brightness %.2f. 3x3 luminance grid [%.0f %.0f %.0f / %.0f %.0f %.0f / %.0f %.0f %.0f]. " +
-        "Compact scene-geometry (%d numbers) vs %d pixel-channels.",
-        w, h, rAvg, gAvg, bAvg, bright,
-        grid[0],grid[1],grid[2],grid[3],grid[4],grid[5],grid[6],grid[7],grid[8],
-        vector.count, w*h*3)
-    disclose("FEAT(image): \(vector.count)-number scene-geometry vs \(w*h*3) pixel-channels (~\(w*h*3/max(vector.count,1))x).")
-    return (content, vector, ["width": w, "height": h, "brightness": bright])
-}
-func featImage(_ path: String) {
-    let r = computeImage(path)
-    var out: [String: Any] = ["ok": true, "op": "feat-image", "file": path,
-                              "vector": r.vector, "content": r.content,
-                              "ingest": ["node_type": "Observation", "tier": "Episodic", "content": r.content]]
-    r.extra.forEach { out[$0] = $1 }
-    emit(out)
-}
-
-// The afferent WIRE — hand a capture's descriptor to the ingest organ (engram),
-// where it becomes an embedded node = GEOMETRY. Own-core URLSession POST.
-// LOCAL only: point at a local engram; raw stream never leaves the machine.
-func postNode(engramURL: String, content: String, label: String, tags: [String]) -> String? {
-    guard let url = URL(string: engramURL + "/api/nodes") else { return nil }
-    let body: [String: Any] = ["content": content, "node_type": "Observation",
-                               "label": label, "tier": "Episodic",
-                               "salience": 0.7, "importance": 0.6, "confidence": 0.9,
-                               "tags": tags]
-    var req = URLRequest(url: url); req.httpMethod = "POST"
-    req.setValue("application/json", forHTTPHeaderField: "Content-Type")
-    req.httpBody = try? JSONSerialization.data(withJSONObject: body)
-    let sem = DispatchSemaphore(value: 0); var out: String?
-    URLSession.shared.dataTask(with: req) { data, _, _ in
-        if let d = data { out = String(data: d, encoding: .utf8) }
-        sem.signal()
-    }.resume()
-    _ = sem.wait(timeout: .now() + 15)
-    return out
-}
-func ingest(_ path: String, kind: String, engramURL: String) {
-    let r = kind == "audio" ? computeAudio(path) : computeImage(path)
-    let label = kind == "audio" ? "heard:mic" : "saw:camera"
-    disclose("INGEST: handing \(kind) descriptor to the ingest organ at \(engramURL) (LOCAL) -> geometry.")
-    guard let resp = postNode(engramURL: engramURL, content: r.content, label: label,
-                              tags: ["peripheral", kind == "audio" ? "afferent-mic" : "afferent-camera"]) else {
-        die("ingest POST failed (no local engram at \(engramURL)?)")
-    }
-    // pull the node id out of the response (own-core, tolerant)
-    var nodeId = ""
-    if let d = resp.data(using: .utf8),
-       let o = try? JSONSerialization.jsonObject(with: d) as? [String: Any] {
-        nodeId = (o["id"] as? String) ?? (o["node_id"] as? String) ?? ""
-    }
-    disclose("INGEST: landed as node \(nodeId.isEmpty ? "(see response)" : nodeId) — the capture is now geometry in the engram.")
-    emit(["ok": !nodeId.isEmpty, "op": "ingest-\(kind)", "file": path,
-          "node_id": nodeId, "engram_response": resp, "content": r.content,
-          "vector": r.vector])
-}
-
-// ============================================================================
-// VOICE BY IMITATION — hear a voice, grab its compact SIGNATURE (pitch +
-// formants F1-F5 via LPC), and speak back in that voice by source-filter
-// resynthesis. Own-core DSP (physics), no training, no stolen voice. The
-// afferent twin of the music instrument-signature: a voice = a few dozen
-// numbers, not a corpus.
-// ============================================================================
-
-func hamming(_ x: [Double]) -> [Double] {
-    let n = x.count; if n < 2 { return x }
-    return (0..<n).map { x[$0] * (0.54 - 0.46*cos(2*Double.pi*Double($0)/Double(n-1))) }
-}
-func autocorr(_ x: [Double], _ p: Int) -> [Double] {
-    var r = [Double](repeating: 0, count: p+1)
-    for lag in 0...p { var s = 0.0; var i = lag; while i < x.count { s += x[i]*x[i-lag]; i += 1 }; r[lag] = s }
-    return r
-}
-// Levinson-Durbin -> LPC coeffs a[0..p] (A(z)=1+sum a[k]z^-k) and residual energy.
-func levinson(_ r: [Double], _ p: Int) -> (a: [Double], err: Double) {
-    var a = [Double](repeating: 0, count: p+1); a[0] = 1
-    var err = r[0]
-    if err <= 0 { return (a, 0) }
-    for i in 1...p {
-        var acc = r[i]
-        if i > 1 { for j in 1..<i { acc += a[j]*r[i-j] } }
-        let k = -acc/err
-        var na = a; na[i] = k
-        if i > 1 { for j in 1..<i { na[j] = a[j] + k*a[i-j] } }
-        a = na; err *= (1 - k*k)
-        if err <= 0 { break }
-    }
-    return (a, err)
-}
-// Formant peaks from the LPC all-pole spectral envelope.
-func formants(_ a: [Double], sr: Int) -> [(f: Double, bw: Double)] {
-    let p = a.count - 1
-    let steps = 512
-    var mag = [Double](repeating: 0, count: steps)
-    for s in 0..<steps {
-        let w = Double.pi * Double(s) / Double(steps)      // 0..pi -> 0..sr/2
-        var re = 0.0, im = 0.0
-        for k in 0...p { re += a[k]*cos(w*Double(k)); im -= a[k]*sin(w*Double(k)) }
-        mag[s] = 1.0 / max((re*re+im*im).squareRoot(), 1e-9)
-    }
-    var peaks: [(f: Double, bw: Double)] = []
-    for s in 1..<(steps-1) where mag[s] > mag[s-1] && mag[s] >= mag[s+1] {
-        let f = Double(s) * Double(sr) / 2 / Double(steps)
-        if f > 150 && f < 5200 {
-            // crude bandwidth: width where magnitude falls to peak/sqrt(2)
-            let thr = mag[s]/1.4142
-            var lo = s; while lo > 0 && mag[lo] > thr { lo -= 1 }
-            var hi = s; while hi < steps-1 && mag[hi] > thr { hi += 1 }
-            let bw = Double(hi-lo) * Double(sr) / 2 / Double(steps)
-            peaks.append((f, bw))
-        }
-    }
-    return Array(peaks.prefix(5))
-}
-func pitchOf(_ frame: [Double], sr: Int) -> Double {
-    let n = frame.count
-    let lagMin = sr/400, lagMax = min(sr/70, n-1)
-    if lagMax <= lagMin { return 0 }
-    var r0 = 0.0; for v in frame { r0 += v*v }
-    if r0 < 1e-5 { return 0 }
-    var bestLag = 0; var best = 0.0
-    for lag in lagMin...lagMax { var c = 0.0; var i = lag; while i < n { c += frame[i]*frame[i-lag]; i += 1 }; if c > best { best = c; bestLag = lag } }
-    return (best / r0 > 0.30 && bestLag > 0) ? Double(sr)/Double(bestLag) : 0   // voiced?
-}
-
-let LPC_ORDER = 16
-let FRAME = 400          // 25ms @16k
-let HOP = 160            // 10ms
-
-// Extract Will's voice-signature: averaged F0 + formants over voiced frames.
-func voiceprint(_ path: String) -> (f0: Double, f0lo: Double, f0hi: Double, formants: [(Double,Double)], content: String) {
-    guard let (x, sr, _) = readWavSamples(path), x.count > FRAME else { die("cannot read speech from \(path)") }
-    var f0s: [Double] = []
-    var fbank: [[Double]] = [[],[],[],[],[]]
-    var bbank: [[Double]] = [[],[],[],[],[]]
-    var pos = 0
-    while pos + FRAME <= x.count {
-        let raw = Array(x[pos..<pos+FRAME])
-        let f0 = pitchOf(raw, sr: sr)
-        if f0 > 0 {                                   // voiced frame only
-            f0s.append(f0)
-            let r = autocorr(hamming(raw), LPC_ORDER)
-            if r[0] > 1e-6 {
-                let (a, _) = levinson(r, LPC_ORDER)
-                let fs = formants(a, sr: sr)
-                for (i, fm) in fs.enumerated() where i < 5 { fbank[i].append(fm.f); bbank[i].append(fm.bw) }
-            }
-        }
-        pos += HOP
-    }
-    func med(_ v: [Double]) -> Double { v.isEmpty ? 0 : v.sorted()[v.count/2] }
-    let f0med = med(f0s)
-    let f0lo = f0s.isEmpty ? 0 : f0s.sorted().first!
-    let f0hi = f0s.isEmpty ? 0 : f0s.sorted().last!
-    var forms: [(Double,Double)] = []
-    for i in 0..<5 where !fbank[i].isEmpty { forms.append((med(fbank[i]), med(bbank[i]))) }
-    let fstr = forms.map { String(format:"%.0f", $0.0) }.joined(separator: "/")
-    let content = String(format:
-        "Voice-signature (afferent, heard a voice): pitch F0 %.0fHz (range %.0f-%.0fHz), " +
-        "formants F1-F5 = %@ Hz. Compact voiceprint (%d numbers) — grabbed by ear for imitation, not trained.",
-        f0med, f0lo, f0hi, fstr, 1 + forms.count*2)
-    return (f0med, f0lo, f0hi, forms, content)
-}
-
-// IMITATE: LPC analysis-resynthesis. Reconstruct the heard voice from its
-// per-frame filter model + pitch — the voice rebuilt from its signature.
-func imitate(inPath: String, outPath: String) {
-    guard let (x, sr, _) = readWavSamples(inPath), x.count > FRAME else { die("cannot read speech from \(inPath)") }
-    var out = [Double](repeating: 0, count: x.count)
-    var state = [Double](repeating: 0, count: LPC_ORDER)   // past outputs
-    var phase = 0.0
-    var lastF0 = 0.0
-    var pos = 0
-    while pos + FRAME <= x.count {
-        let raw = Array(x[pos..<pos+FRAME])
-        let r = autocorr(hamming(raw), LPC_ORDER)
-        let f0 = pitchOf(raw, sr: sr)
-        if r[0] < 1e-7 { pos += HOP; continue }
-        let (a, err) = levinson(r, LPC_ORDER)
-        let gain = max(err, 0).squareRoot()
-        let useF0 = f0 > 0 ? f0 : (lastF0 > 0 ? lastF0 : 0)
-        lastF0 = f0
-        for i in 0..<HOP {
-            let idx = pos + i; if idx >= x.count { break }
-            var e = 0.0
-            if useF0 > 0 {                              // voiced: glottal impulse train
-                phase += useF0/Double(sr)
-                if phase >= 1.0 { phase -= 1.0; e = sqrt(Double(sr)/useF0) }   // energy-normalized impulse
-            } else {                                    // unvoiced: noise
-                e = Double.random(in: -1...1)
-            }
-            var y = gain * e
-            for k in 1...LPC_ORDER { y -= a[k]*state[k-1] }
-            for k in stride(from: LPC_ORDER-1, through: 1, by: -1) { state[k] = state[k-1] }
-            state[0] = y
-            out[idx] = y
-        }
-        pos += HOP
-    }
-    // normalize to peak 0.9
-    let peak = out.map { abs($0) }.max() ?? 1
-    let scale = peak > 1e-9 ? 0.9/peak : 1
-    let samples = out.map { Int16(max(-32767, min(32767, $0*scale*32767))) }
-    writeWav(URL(fileURLWithPath: outPath), samples: samples, sampleRate: sr)
-    let vp = voiceprint(inPath)
-    disclose(String(format: "IMITATE: rebuilt the voice from its signature (F0 %.0fHz, formants %@) -> %@",
-                    vp.f0, vp.formants.map{String(format:"%.0f",$0.0)}.joined(separator:"/"), outPath))
-    emit(["ok": true, "op": "imitate", "in": inPath, "out": outPath,
-          "f0_hz": vp.f0, "f0_range": [vp.f0lo, vp.f0hi],
-          "formants_hz": vp.formants.map { $0.0 },
-          "method": "LPC analysis-resynthesis (own-core, no training, no stolen voice)"])
-}
-
-// ============================================================================
-// CONVERSE (full-duplex) — the interruptible conversational loop.
-//   The utterance is a persistent, ordered meaning-plan of SEGMENTS, each with
-//   a salience. The speaker plays them; the mic listens concurrently. On user
-//   speech: pause INSTANTLY, classify (backchannel vs barge-in), then DECIDE
-//   yield-or-hold from the salience of the current segment + the social read.
-//   Yielded utterances persist their remaining plan so Neuron can RESUME.
-// ============================================================================
-struct Segment { let file: String; let salience: Double; let text: String }
-
-enum Decision { case backchannelContinue, hold, yield }
-
-// The yield-or-hold DECISION — grounded, contextual. Not a fixed rule.
-func decide(currentSalience: Double, progress: Double,
-            interrupterAuthority: Double, isBackchannel: Bool) -> Decision {
-    if isBackchannel { return .backchannelContinue }          // "mm-hm" => keep going
-    // Holding the floor is justified when what I'm saying matters AND I'm nearly
-    // done (cheap to finish) AND the interrupter isn't high-priority.
-    let holdScore = currentSalience * 0.6 + progress * 0.4
-    if holdScore >= 0.6 && interrupterAuthority < 0.8 { return .hold }
-    return .yield                                              // default: be polite, let them in
-}
-
-final class Conversation {
-    let engine = AVAudioEngine()
-    let player = AVAudioPlayerNode()
-    var micLive = false
-    // VAD state (shared with the audio tap thread)
-    let lock = NSLock()
-    var micRMS: Float = 0
-    var speechFrames = 0          // consecutive above-threshold frames
-    var onsetHandled = false
-
-    let resumePath: String
-    init(resumePath: String) { self.resumePath = resumePath }
-
-    // Try to bring the mic up as a live VAD. Returns false if unavailable/denied.
-    func startMic() -> Bool {
-        let status = AVCaptureDevice.authorizationStatus(for: .audio)
-        if Consent.load()["mic"] != true || status != .authorized {
-            disclose("CONVERSE: live mic not available (consent/OS) — using injected barge events for the proof.")
-            return false
-        }
-        let input = engine.inputNode
-        // Acoustic echo cancellation: the OS voice-processing unit subtracts our
-        // own speaker output from the mic so Neuron does NOT hear itself and
-        // barge in on its own voice. This is what makes real-room barge-in work.
-        do { try input.setVoiceProcessingEnabled(true); disclose("CONVERSE: AEC on (echo-cancelled mic — won't self-interrupt).") }
-        catch { disclose("CONVERSE: AEC unavailable (\(error)); raising VAD floor instead.") }
-        let fmt = input.inputFormat(forBus: 0)
-        if fmt.sampleRate == 0 { return false }
-        input.installTap(onBus: 0, bufferSize: 1024, format: fmt) { [weak self] buf, _ in
-            guard let self = self, let ch = buf.floatChannelData?[0] else { return }
-            let n = Int(buf.frameLength)
-            var sum: Float = 0
-            for i in 0..<n { let v = ch[i]; sum += v*v }
-            let rms = n > 0 ? (sum / Float(n)).squareRoot() : 0
-            self.lock.lock(); self.micRMS = rms; self.lock.unlock()
-        }
-        micLive = true
-        disclose("CONVERSE: full-duplex — mic listening WHILE speaking (barge-in armed).")
-        return true
-    }
-
-    func run(_ segs: [Segment], interrupterAuthority: Double,
-             injectBargeAt: Double?, injectKind: String, startIndex: Int, liveMic: Bool) {
-        engine.attach(player)
-        let firstFmt = (try? AVAudioFile(forReading: URL(fileURLWithPath: segs[startIndex].file)))?.processingFormat
-            ?? AVAudioFormat(standardFormatWithSampleRate: 16000, channels: 1)!
-        engine.connect(player, to: engine.mainMixerNode, format: firstFmt)
-        if liveMic { _ = startMic() }
-        else { disclose("CONVERSE: deterministic mode (live mic off) — barge events \(injectBargeAt != nil ? "injected" : "none").") }
-        do { try engine.start() } catch { die("audio engine failed to start: \(error)") }
-        player.play()
-
-        let injectDeadline = injectBargeAt.map { Date().addingTimeInterval($0) }
-        var injectedFired = false
-        var idx = startIndex
-
-        segmentLoop: while idx < segs.count {
-            let seg = segs[idx]
-            guard let f = try? AVAudioFile(forReading: URL(fileURLWithPath: seg.file)) else {
-                disclose("CONVERSE: missing segment '\(seg.file)', skipping."); idx += 1; continue
-            }
-            let dur = Double(f.length) / f.processingFormat.sampleRate
-            disclose(String(format: "CONVERSE: speaking segment %d/%d (salience %.2f) — \"%@\"",
-                            idx+1, segs.count, seg.salience, seg.text))
-            emit(["op": "converse", "event": "speaking", "segment": idx,
-                  "salience": seg.salience, "text": seg.text])
-
-            let done = DispatchSemaphore(value: 0)
-            // .dataPlayedBack: completion fires only after the audio has actually
-            // played OUT the DAC (not merely been consumed) — so the tail is never
-            // clipped and playback always runs the FULL file length.
-            player.scheduleFile(f, at: nil, completionCallbackType: .dataPlayedBack) { _ in done.signal() }
-            player.play()
-
-            // Monitor this segment: poll VAD / injected event until it finishes.
-            let segStart = Date()
-            while done.wait(timeout: .now() + 0.02) == .timedOut {
-                let elapsed = Date().timeIntervalSince(segStart)
-                let progress = min(elapsed / max(dur, 0.001), 1.0)
-
-                // --- detect an onset (live mic OR injected) ---
-                var onset = false
-                if micLive {
-                    lock.lock(); let rms = micRMS; lock.unlock()
-                    if rms > 0.02 { speechFrames += 1 } else { speechFrames = 0 }
-                    if speechFrames >= 3 && !onsetHandled { onset = true }   // ~60ms of voice
-                }
-                if let dl = injectDeadline, !injectedFired, Date() >= dl, !onsetHandled { onset = true; injectedFired = true }
-
-                if onset {
-                    onsetHandled = true
-                    // (1) BARGE-IN: pause INSTANTLY, on the spot.
-                    player.pause()
-                    let tBarge = Date().timeIntervalSince(segStart)
-                    disclose(String(format: "CONVERSE: << user speech at %.2fs into segment %d — PAUSED instantly >>", tBarge, idx+1))
-                    emit(["op": "converse", "event": "barge_in", "segment": idx,
-                          "at_seconds": tBarge, "progress": progress])
-
-                    // (2) classify backchannel vs real barge-in
-                    let isBackchannel = classifyBackchannel(injected: injectDeadline != nil,
-                                                            kind: injectKind)
-                    let d = decide(currentSalience: seg.salience, progress: progress,
-                                   interrupterAuthority: interrupterAuthority,
-                                   isBackchannel: isBackchannel)
-                    switch d {
-                    case .backchannelContinue:
-                        disclose("CONVERSE: read as BACKCHANNEL (\"mm-hm\") — keep going, resume seamlessly.")
-                        emit(["op": "converse", "event": "backchannel_continue", "segment": idx])
-                        onsetHandled = false; speechFrames = 0
-                        player.play()                                   // seamless resume
-                    case .hold:
-                        disclose("CONVERSE: HOLD the floor — \"hang on, let me finish this thought.\" (high salience, nearly done)")
-                        emit(["op": "converse", "event": "hold_floor", "segment": idx,
-                              "salience": seg.salience, "progress": progress])
-                        onsetHandled = false; speechFrames = 0
-                        player.play()                                   // finish the segment, THEN yield
-                        // after this segment completes we yield the remainder
-                        _ = done.wait(timeout: .now() + dur + 1.0)
-                        persistResume(segs: segs, from: idx + 1, reason: "held-then-yield")
-                        finish(); return
-                    case .yield:
-                        disclose("CONVERSE: YIELD — stop, let them in. Remembering where I was (resumable).")
-                        player.stop()
-                        persistResume(segs: segs, from: idx, reason: "yield")
-                        emit(["op": "converse", "event": "yield", "interrupted_segment": idx,
-                              "resume_from": idx])
-                        finish(); return
-                    }
-                }
-            }
-            emit(["op": "converse", "event": "segment_done", "segment": idx])
-            idx += 1
-        }
-        // whole utterance completed uninterrupted
-        clearResume()
-        disclose("CONVERSE: utterance complete (uninterrupted).")
-        emit(["ok": true, "op": "converse", "event": "complete", "segments": segs.count])
-        finish()
-    }
-
-    // A backchannel is brief/low. Injected kind lets us prove both paths headlessly;
-    // the live path would measure post-onset duration & energy.
-    func classifyBackchannel(injected: Bool, kind: String) -> Bool {
-        if injected { return kind == "backchannel" }
-        // live: sample ~250ms after onset; if speech already died away, it was a backchannel
-        Thread.sleep(forTimeInterval: 0.25)
-        lock.lock(); let rms = micRMS; lock.unlock()
-        return rms < 0.015
-    }
-
-    func persistResume(segs: [Segment], from: Int, reason: String) {
-        let remaining = segs[from...].map { ["file": $0.file, "salience": $0.salience, "text": $0.text] as [String: Any] }
-        let state: [String: Any] = ["resume_from": from, "reason": reason,
-                                    "remaining": remaining, "ts": Date().timeIntervalSince1970]
-        if let d = try? JSONSerialization.data(withJSONObject: state, options: [.prettyPrinted]) {
-            try? d.write(to: URL(fileURLWithPath: resumePath))
-        }
-        disclose("CONVERSE: meaning-plan persisted (\(remaining.count) segments remain) — Neuron can resume the thread.")
-    }
-    func clearResume() { try? FileManager.default.removeItem(atPath: resumePath) }
-    func finish() { player.stop(); if micLive { engine.inputNode.removeTap(onBus: 0) }; engine.stop() }
-}
-
-// ----------------------------------------------------------------------------
-// CLI
-// ----------------------------------------------------------------------------
-func loadManifest(_ path: String) -> (segs: [Segment], utterance: String) {
-    guard let d = FileManager.default.contents(atPath: path),
-          let o = try? JSONSerialization.jsonObject(with: d) as? [String: Any],
-          let arr = o["segments"] as? [[String: Any]] else { die("bad manifest: \(path)") }
-    let segs = arr.map { Segment(file: $0["file"] as? String ?? "",
-                                 salience: ($0["salience"] as? NSNumber)?.doubleValue ?? 0.5,
-                                 text: $0["text"] as? String ?? "") }
-    return (segs, o["utterance"] as? String ?? "")
-}
-
-let args = CommandLine.arguments
-guard args.count >= 2 else {
-    print("""
-    periph — Neuron peripheral I/O (own-core, local, consent-gated)
-      grant <camera|mic>            grant a sensitive sense (Neuron-level consent)
-      revoke <camera|mic>           revoke it
-      status                        show consent state
-      speak <file.wav>              SPEAK ALOUD (efferent) via the speaker
-      tone <out.wav> [hz] [sec]     own-core synth a test WAV (no deps)
-      listen <sec> <out.wav>        MIC capture (afferent) 16k mono
-      see <out.jpg>                 CAMERA one frame (afferent)
-      feat-audio <file.wav>         extract compact voice/sound signature (for ingest)
-      feat-image <file.jpg>         extract compact scene-geometry (for ingest)
-      ingest-audio <file.wav> <engramURL>   capture -> descriptor -> engram node (geometry)
-      ingest-image <file.jpg> <engramURL>   capture -> descriptor -> engram node (geometry)
-      voiceprint <voice.wav>        extract voice-signature (F0 + formants F1-F5)
-      imitate <voice.wav> <out.wav> speak back in that voice (LPC analysis-resynthesis)
-      hear-imitate <sec> <out.wav>  MIC -> extract signature -> imitate -> SPEAK ALOUD
-      wav-info <file.wav>           print WAV geometry
-      converse <manifest.json> [--authority F] [--barge-at S[:backchannel|:bargein]] [--resume]
-                                    full-duplex interruptible utterance
-    """)
-    exit(0)
-}
-
-switch args[1] {
-case "grant":
-    guard args.count >= 3 else { die("grant needs a device") }
-    var g = Consent.load(); g[args[2]] = true; Consent.save(g)
-    disclose("granted '\(args[2])' — the user consents; raw stream stays local, never egresses.")
-    emit(["ok": true, "op": "grant", "device": args[2], "consent": g])
-case "revoke":
-    guard args.count >= 3 else { die("revoke needs a device") }
-    var g = Consent.load(); g[args[2]] = false; Consent.save(g)
-    emit(["ok": true, "op": "revoke", "device": args[2], "consent": g])
-case "status":
-    emit(["ok": true, "op": "status", "consent": Consent.load()])
-case "speak":
-    guard args.count >= 3 else { die("speak needs a wav") }
-    speak(args[2])
-case "tone":
-    guard args.count >= 3 else { die("tone needs an out path") }
-    let hz = args.count >= 4 ? Double(args[3]) ?? 220 : 220
-    let sec = args.count >= 5 ? Double(args[4]) ?? 1.0 : 1.0
-    let sr = 16000
-    var s = [Int16](); s.reserveCapacity(Int(Double(sr)*sec))
-    for i in 0..<Int(Double(sr)*sec) {
-        let t = Double(i)/Double(sr)
-        let env = min(1.0, min(t*20, (sec - t)*20))          // gentle attack/release
-        s.append(Int16(env * 0.3 * 32767 * sin(2*Double.pi*hz*t)))
-    }
-    writeWav(URL(fileURLWithPath: args[2]), samples: s, sampleRate: sr)
-    disclose("tone: wrote own-core \(sec)s @ \(hz)Hz WAV to \(args[2]).")
-    emit(["ok": true, "op": "tone", "file": args[2], "hz": hz, "seconds": sec])
-case "listen":
-    guard args.count >= 4 else { die("listen needs <sec> <out.wav>") }
-    listen(seconds: Double(args[2]) ?? 3.0, out: args[3])
-case "see":
-    guard args.count >= 3 else { die("see needs an out path") }
-    see(out: args[2])
-case "feat-audio":
-    guard args.count >= 3 else { die("feat-audio needs a wav") }
-    featAudio(args[2])
-case "feat-image":
-    guard args.count >= 3 else { die("feat-image needs an image") }
-    featImage(args[2])
-case "ingest-audio":
-    guard args.count >= 4 else { die("ingest-audio needs <wav> <engramURL>") }
-    ingest(args[2], kind: "audio", engramURL: args[3])
-case "ingest-image":
-    guard args.count >= 4 else { die("ingest-image needs <image> <engramURL>") }
-    ingest(args[2], kind: "image", engramURL: args[3])
-case "voiceprint":
-    guard args.count >= 3 else { die("voiceprint needs a wav") }
-    let vp = voiceprint(args[2])
-    disclose("VOICEPRINT: \(vp.content)")
-    emit(["ok": true, "op": "voiceprint", "file": args[2], "f0_hz": vp.f0,
-          "f0_range": [vp.f0lo, vp.f0hi], "formants_hz": vp.formants.map { $0.0 },
-          "bandwidths_hz": vp.formants.map { $0.1 }, "content": vp.content,
-          "ingest": ["node_type": "Observation", "tier": "Episodic", "content": vp.content]])
-case "imitate":
-    guard args.count >= 4 else { die("imitate needs <voice.wav> <out.wav>") }
-    imitate(inPath: args[2], outPath: args[3])
-case "hear-imitate":
-    guard args.count >= 4 else { die("hear-imitate needs <sec> <out.wav>") }
-    let secs = Double(args[2]) ?? 4.0
-    let outp = args[3]
-    let capp = outp.replacingOccurrences(of: ".wav", with: "") + ".heard.wav"
-    disclose("HEAR-IMITATE: open the ear, listen \(secs)s, grab the voice, speak it back.")
-    listen(seconds: secs, out: capp)                 // afferent: hear the voice
-    imitate(inPath: capp, outPath: outp)             // extract signature + resynthesize
-    speak(outp)                                      // efferent: speak back ALOUD in that voice
-case "wav-info":
-    guard args.count >= 3, let i = wavInfo(args[2]) else { die("wav-info needs a readable wav") }
-    disclose("WAV \(args[2]): \(i.sampleRate)Hz \(i.channels)ch \(i.bits)bit \(i.frames) frames")
-    emit(["ok": true, "op": "wav-info", "sample_rate": i.sampleRate, "channels": i.channels,
-          "bits": i.bits, "frames": i.frames,
-          "seconds": Double(i.frames)/Double(max(i.sampleRate,1))])
-case "converse":
-    guard args.count >= 3 else { die("converse needs a manifest") }
-    let (segs, utter) = loadManifest(args[2])
-    var authority = 0.5
-    var bargeAt: Double? = nil
-    var bargeKind = "bargein"
-    var resume = false
-    var liveMic = false
-    var i = 3
-    while i < args.count {
-        switch args[i] {
-        case "--authority": if i+1 < args.count { authority = Double(args[i+1]) ?? 0.5; i += 1 }
-        case "--barge-at":
-            if i+1 < args.count {
-                let parts = args[i+1].split(separator: ":")
-                bargeAt = Double(parts[0]) ?? nil
-                if parts.count > 1 { bargeKind = String(parts[1]) }
-                i += 1
-            }
-        case "--resume": resume = true
-        case "--live-mic": liveMic = true
-        default: break
-        }
-        i += 1
-    }
-    let resumePath = (ProcessInfo.processInfo.environment["PERIPH_HOME"]
-        ?? FileManager.default.currentDirectoryPath + "/peripheral") + "/.resume.json"
-    var startIndex = 0
-    var runSegs = segs
-    if resume, let d = FileManager.default.contents(atPath: resumePath),
-       let o = try? JSONSerialization.jsonObject(with: d) as? [String: Any],
-       let rem = o["remaining"] as? [[String: Any]] {
-        runSegs = rem.map { Segment(file: $0["file"] as? String ?? "",
-                                    salience: ($0["salience"] as? NSNumber)?.doubleValue ?? 0.5,
-                                    text: $0["text"] as? String ?? "") }
-        startIndex = 0
-        disclose("CONVERSE: resuming — \"as I was saying...\" (\(runSegs.count) segments left).")
-        emit(["op": "converse", "event": "resume", "remaining": runSegs.count])
-    }
-    if runSegs.isEmpty { die("no segments to speak") }
-    disclose("CONVERSE: utterance = \"\(utter)\" (\(runSegs.count) segments).")
-    let convo = Conversation(resumePath: resumePath)
-    convo.run(runSegs, interrupterAuthority: authority,
-              injectBargeAt: bargeAt, injectKind: bargeKind, startIndex: startIndex, liveMic: liveMic)
-default:
-    die("unknown command: \(args[1])")
-}