v2.6.0 - ESP32-C5 Support, Context-Aware Calibration, and Stricter Validation Targets

[francescopace]

Highlights

• More robust runtime on modern chips (ESP32-C5/C6): WiFi lifecycle handling is hardened, dual-band protocol/bandwidth APIs are used correctly (with safe fallback), C5 is forced to 2.4 GHz, and C5 CSI 114-byte payloads are normalized to HT20 128-byte internal layout
• Safer calibration and detector state transitions: calibration start failures are now handled explicitly, detector buffers are cold-cleared after calibration/channel switches, and NBVI input/band-size validation is hardened
• Stricter quality bar for motion validation: Python and C++ performance targets are now unified to Recall >95% and FP <5% for both MVS and ML, with docs updated accordingly (PERFORMANCE.md, test/README.md)

Reliability and Runtime Fixes

• Threshold handling unified across stacks: validation is aligned to 0.0-10.0 across ESPHome/C++ and Micro-ESPectre/Python (HA number, Serial, MQTT, detector setters); MQTT now propagates detector rejection correctly; factory_reset restores ML threshold to 5.0
• Serial command parsing hardening: T:<value> now uses validated strtof parsing (endptr, finite/range checks)
• Startup fail-fast behavior: setup now marks the component failed when WiFi initialization/handler registration fails
• Auxiliary task stability: DNS task always clears running_ on early exits, avoiding stale "already running" states
• Safety guards in diagnostics utilities: progress-bar width/marker bounds are now clamped to prevent fixed-buffer edge cases

Calibration, ML, and Dataset Pipeline

• NBVI hot-path optimization: reduced allocations, enforced memory-bounded chunked validation reads (avoids std::bad_alloc/abort() on low-heap targets), and replaced O(window) shifts with ring buffer + running statistics
• Context-aware grid-search metadata workflow: micro-espectre/tools/11_refresh_gridsearch_metadata.py introduced and then simplified to a single C++-aligned evaluation path (legacy hardcoded subcarrier override removed)
• Metadata consistency cleanup: gain_locked is now the single source of truth in .npz and dataset_info.json; deprecated use_cv_normalization and label_id metadata removed
• ML pipeline alignment (training + inference): both stacks now use [12, 14, 16, 18, 20, 24, 28, 36, 40, 44, 48, 52]; models were retrained/re-exported with validated seed, and feature extraction was simplified to the selected 12 runtime features

Tooling and Developer Experience

• Micro-ESPectre deploy diagnostics improved: ./me deploy now performs a REPL health-check and reports explicit remediation for ROM boot-loop (invalid header) with ./me flash --erase
• C5 support in me CLI expanded: C5 auto/manual selection, --chip c5, correct esp32c5 target mapping, and C5 firmware artifact selection (ESP32_CSI_C5.bin)
• Flash mapping hardening: per-chip offsets aligned with MicroPython board deploy options (including C5 0x2000)
• Optional BSSID lock in Micro-ESPectre: WIFI_BSSID support added in src/main.py
• ESP-IDF mock alignment: WiFi mock headers updated to modern protocol bitmasks, band-mode enums, and dual-band API signatures
• General cleanup: removed unused BaseDetector amplitude getters and refreshed stale comments/documentation (including C5 tested status in setup/examples, S2 still experimental)
• ESPHome 2026.2.4 validation: project configuration was re-validated after upgrading ESPHome from 2026.2.0 to 2026.2.4 (esphome config examples/espectre-c6-dev.yaml), with full Python and C++ test suites passing

---

---

This discussion was created from the release v2.6.0 - ESP32-C5 Support, Context-Aware Calibration, and Stricter Validation Targets.