fix(edge): gate multi-person count on presence — no phantom persons (#803, #496)

[markt-heximal]

Addresses #803 and #496.

Problem

update_multi_person_vitals() (edge_processing.c) sets pv->active = true unconditionally for n_persons = s_top_k_count / 2 (= 4 at the default EDGE_TOP_K=8) every frame. So the reported n_persons is a fixed function of the subcarrier count — never 0, regardless of how many people are actually present. An empty room reports 4; users see a constant phantom count.

Root cause behind:

• The ESP32-S3 device cannot correctly identify the exact number of people nearby. #803 — "consistently shows 1 person" / "is people-counting even implemented?"
• demo跑起来了，但最基本的人是否在，以及正常的人数显示都无法做到 #496 — "shows 2 people, but I'm alone in the room"

The firmware already computes presence (s_presence_detected, set earlier in the same process_frame() pass) — but the multi-person path never consults it.

Fix

Gate person activation on s_presence_detected: when no presence is detected, mark all persons inactive and return → n_persons = 0 in an empty room; persons activate only once presence crosses the (adaptive) threshold.

if (!s_presence_detected) {
    for (uint8_t p = 0; p < EDGE_MAX_PERSONS; p++) s_persons[p].active = false;
    return;
}

Scope / honest limitations

This bounds the count to presence (0 when empty, ≥1 when someone is there). It does not attempt true multi-person separation — when present, the count is still the subcarrier-derived top_k/2. Real counting/separation from a single link is a larger DSP problem; this PR just removes the "always N phantom people" behavior everyone is hitting.

Testing

• Builds clean — ESP-IDF v5.4, idf.py set-target esp32s3 && idf.py build.
• The gating signal was verified on hardware: presence_score (= motion_energy) / s_presence_detected read ~0 at idle and rise with movement, so the gate yields 0 when empty and activates on presence. The pre-fix constant count (4 at top_k=8) was observed directly on an ESP32-S3.

🤖 Generated with Claude Code

[ruvnet]

Independently verified — this fix is correct, and the call ordering it relies on checks out against current main (edge_processing.c):

• process_frame() runs update_top_k() (L745) → sets s_presence_detected = (s_presence_score > threshold) (L817) → calls update_multi_person_vitals() (L851).
• This PR's gate sits at the top of update_multi_person_vitals (L478), i.e. it executes at L851 — after L817 — so s_presence_detected is the freshly-updated value for this frame. The comment's "updated earlier in the same process_frame() pass" claim is accurate.

Before this, update_multi_person_vitals activated s_top_k_count / 2 persons whenever s_top_k_count >= 2 (essentially always), so an empty room reported a fixed top_k/2 phantom count (the #803/#496/#894 "4–10 people when nobody's there"). Gating activation on presence is the right minimal fix; the existing for (p = n_persons; p < EDGE_MAX_PERSONS) active = false tail already handles the partial case, and this extends it to the zero-presence case.

Worth noting this is the edge/firmware half of the over-count. The server heuristic path has a complementary cap on main — field_bridge.rs bounds occupancy to MAX_SINGLE_LINK_OCCUPANCY = 3 (#907). Together they close both layers: the ESP32 no longer emits phantom persons, and the server can't amplify a single link past 3.

LGTM to merge. (Happy to run an ESP-IDF build to confirm it compiles if that's useful before merge — it only touches existing symbols, so it should be clean.)