Eliminate file re-read in offloading decision path

[lbedogni]

Status: Partially Complete. Eliminating file re-read addressed in SCIOT-009 per-device state work. Remaining: remove redundant JSON parsing for offloading history.

[ff225]

Investigated the current state of the offloading decision path — the redundant JSON re-parsing this issue describes no longer exists:

• src/server/offloading_algo/offloading_algo.py:18-19,53-67 — _valid_points_cache (class-level dict) caches the valid-offloading-points JSON per model dir; loaded once, not per request.
• src/server/communication/request_handler.py:906-930 — _layer_sizes_cache (class-level dict) caches per-model layer sizes JSON, with an inline comment noting layer sizes never change at runtime so they're read once per model.
• src/server/communication/request_handler.py:232-239 — num_layers is computed once at first __init__, not per request.
• iter_offloading_decisions (telemetry/offloading_decisions.py:288) is only used for dashboard/reporting reads, not the hot decision path.

This looks like it was already addressed by the per-device state work referenced in the issue's own status note (blocked by SCIOT-009). Closing as resolved — reopen if there's a specific remaining re-read case I missed.