[infra] Jetson Orin Nano storage + RAM budget for full TERA stack

[jdev-02]

Tracking issue for the cumulative footprint of every artifact TERA needs at runtime on the Jetson Orin Nano (8GB unified RAM, default 64GB SD card). Surfaced after the Sat 19:00 voice-bakeoff exercise: 4 candidate Piper voices = ~240MB just to compare them. Need a budget + monitor before we run out at demo time.

Working-set inventory (rough estimates -- update as we measure)

component	owner	disk	RAM at peak
Gemma 2B Q4_K_M	Kyle	~1.6 GB	~2.5 GB
Piper voice (1x chosen)	Jon	30-75 MB	~150 MB
Whisper-tiny	Jon	~75 MB	~250 MB
ATAK / WinTAK	Ben	~80 MB	~300 MB
TERA python venv + deps	Jon	~500 MB	~250 MB
OSM extract (hero AOI)	Ben	50-250 MB	~100 MB
DEM tiles (hero AOI)	Ben	100-500 MB	~200 MB
Cesium tile cache	Jon	500 MB - 2 GB	~300 MB
Routing graph (Valhalla)	Ben	100-500 MB	~300 MB
Total		3-5 GB	4-5 GB

Asks

• Kyle: measure actual storage + free-RAM on the Jetson with full stack loaded. Document in /docs/PHASE3_RUNBOOK.md or similar.
• Ben: profile OSM + DEM + Cesium cache size for both AOIs (urban SF, austere). Set a per-AOI cap.
• Jon: voice + Whisper -- pick ONE Piper voice for prod, document in pyproject.toml [voice] extras. The 4 bakeoff voices stay as dev-only artifacts.
• Satriyo: not assigned but flagged -- security pipeline + crypto runtime memory should be in the table above too.

Acceptance

• One source of truth (table above, kept up to date in this issue)
• A 'make budget' or similar Make target that prints current artifact sizes before deployment
• A red-line: if total disk > 16 GB or peak RAM > 6 GB, raise an integrator alert

Source: Jon's Sat 19:18 chat after observing 75MB voice-model downloads stack up.