ESP32-S3 v0.6.5: CSI callbacks extremely sparse (3 callbacks in 70s under heavy WiFi traffic), yield stays 0pps

[labuhardilladelloco]

ESP32-S3 N16R8
firmware v0.6.5
edge_tier=0
WPA2 2.4 GHz canal 6 BW20
YouTube 4K + iperf active
Promiscuous mode enabled (MGMT-only)
CSI callbacks extremely rare (#1 at 8s, #2 at 59s, #3 at 70s)
yield=0pps
sendto ENOMEM sometimes
server receives only small UDP health packets
nodes become stale

Expected:
Under YouTube 4K + iperf traffic, CSI callbacks should be frequent enough to keep the node active and produce non-zero yield.

Actual:
Only HEALTH packets appear to be sent regularly. UDP listener receives mostly 48/60-byte packets.

Question:
Is MGMT-only promiscuous mode expected to produce CSI callbacks from data traffic, or should the firmware capture DATA frames too?

[ruvnet]

Root-caused and fixed in #868, hardware-validated on an ESP32-C6 (COM9).

Answer to your question: MGMT-only was intentional (DATA frames once crashed Core 0 in wDev_ProcessFiq, an SPI-flash cache race), but it's the cause of your sparse callbacks. Two things were wrong:

1. The firmware comment promised "probe-request injection at 10 Hz" to keep CSI alive under MGMT-only — that injection never actually existed. So the node depended entirely on ambient beacons, which starve under heavy DATA traffic exactly as you saw (3 cb / 70 s, 0 pps).
2. I implemented the real probe injection — it transmits fine at 10 Hz, but management-frame CSI stays sparse on these chips.

The actual fix: re-admit DATA frames (MGMT+DATA). The original crash is now neutralized by the WiFi RX/TX IRAM optimizations + a 50 Hz early rate gate that were added since the MGMT-only decision, so the interrupt storm no longer panics. (Kconfig CSI_PROMISC_MGMT_ONLY is available as a fallback.)

Measured on COM9, before → after:

	MGMT-only	MGMT+DATA (50 Hz-gated)
Yield	0 pps (~3 cb/20s)	~9 pps avg, peak 19
Presence/motion	dead	sensing restored
Panics / 35s	—	0

Validated on a C6; the fix is in shared csi_collector.c so the S3 (your N16R8) gets it too. Please re-test on your S3 once #868 lands and report back the yield.