v1.4.1 — fleet-scale reliability + SHOW_SUBNET (+ staged migrate/registry)

Main caught up to dev. This is the batch that made fleet collection/status/shutdown reliable at scale (validated live today: red 48/48, blue 70/70 clean, shutdown one-pass).

Live-validated

• ml_status fleet-scale fix — builds its work list from the device file (no false-OFFLINE), parallel probe+connect, low per-device parallelism (ML_STATUS_PARALLEL, default 8), surfaces unreadable devices as ERROR + a post-collection sweep so Total always reconciles.
• stdin-eat fix — backgrounded adb shell was consuming the loop's device list, so only ~MAX_PARALLEL devices survived a sweep. Fixed across ml_status + ml_deploy (run_parallel/shutdown/push).
• SHOW_SUBNET (per-show; both shows 172.16.40.0/22) + fixed detect_subnet (no longer caps at /24) → no under-scan.
• ml_scan/ml_status reset the adb server at start (ML_NO_KILLSERVER=1 to skip) → no "unknown"/contention wall.

Staged — present but NOT yet live-validated (do not use for real ops until one supervised test)

• ml_show_migrate.sh — RED↔BLUE single-device migration (USB; deploy→confirm→provision→inventory-move; --dry-run).
• apps_script/registry.gs — device→show membership web app (one row per serial; unwired reference — deploy when ready).
• Per-show inventory foundation (SHOW_INVENTORY_FILE → inventory/<show>.csv).

Operators: ./update.sh when not mid-flash. Known deferred: ml_deploy locate is silent on provisioned devices (notification sound disabled by provisioning).