Validation & Evidence Boundary

Automation matrix

Layer	v6 / v7 evidence
Web	TypeScript, clearance, Vitest and Vite build; v7: 18 files / 141 tests
Pure C++	11 economy golden snapshots plus gestures, saves, power, IMU and audio/OLED policy
Rendering	Eight scenes × 30 frames, incremental equals full; FX/reduced/particle contracts
Visual	v6 initial/endgame × three pages, 114,546 tested pixels/page, 100% palette parity
Build	Production ESP-IDF app, partition headroom and asset hash
Device boot	ESP32-S3, 16 MiB Flash, 8 MiB PSRAM, CO5300, CST9217, QMI8658, RTC and NVS
Stability	No WDT, Guru Meditation, assert, brownout, coredump or sustained memory loss

Firmware artifacts

v6 production firmware is 1,342,416 bytes with about 84% app-partition headroom. Current v7 is 1,536,336 bytes with about 82% headroom, SHA-256 ec2229…00e6c. The v7 asset pack is 710,740 bytes, SHA-256 9fbefa…aa93.

Performance

• v6 final: 29.61–29.83 FPS.
• v7 ordinary QA: 29.20–29.33 FPS.
• v7 current complex save: 26.17–26.70 FPS.
• Reduced: 14.89–14.96 FPS.
• Ambient: 2.00 FPS.

Every number is bound to a scene; an empty-page peak is never presented as all-state performance.

Automated versus human evidence

Automation proves controller handshake, IRQ, coordinates, formulas, pixels, sequence timing, save read-back, IMU signs and the PCM software chain.

It cannot fully prove AMOLED glass/color/brightness, comfort of hold thresholds, edge friction, external-speaker pressure/distortion or subjective HIGH parallax. The project never treats a clean serial log as proof that eyes and hands are satisfied.

Reproducible commands

npm run check
bash pocket-star/firmware/tools/run_host_tests.sh
bash pocket-star/firmware/tools/render_host_snapshots.sh
idf.py -C pocket-star/firmware/pocket-star build
./pocket-star/firmware/tools/verify_device.sh /dev/cu.usbmodem1101 16

The final command builds, flashes, verifies, resets and checks device markers. It requires the correct connected port and target.