Extract OffloadingService and test the inference hot path (#45)

[ff225]

Closes #45. Third and last step, after #74 (state) and #76 (recorder).

Stacked on #76 — base is fix/sciot-010-inference-cycle-recorder, not main. Review #74 → #76 → this. Independent of #75.

The #45 checklist

• Extract DeviceStateManager — Extract per-device state out of RequestHandler (#11) #74
• Extract InferenceCycleRecorder — Serialize simulation CSV writes behind a recorder (#10) #76
• Extract OffloadingService — this PR
• Unit tests for handle_device_inference_result — this PR
• Make shared state thread-safe — Extract per-device state out of RequestHandler (#11) #74 (per-device state) and Serialize simulation CSV writes behind a recorder (#10) #76 (CSV writers)

OffloadingService

RequestHandler built the algorithm, ran it, unpicked its candidate list and published the decision telemetry inline. OffloadingService owns that now, returning an OffloadingDecision that carries everything telemetry needs to explain the choice: the layer, the reason, the candidates, the switch penalty and the estimated cost.

Deriving the switch penalty and the cost estimate moves onto the decision object — they are properties of the decision, not of the request. The handler is left orchestrating: it asks for a decision and reports it.

Deliberate deviation from the issue. #45 says "OffloadingService (wrapper around OffloadingAlgo + decision logging)". The service owns the algorithm and the decision logging (append_and_publish_decision), but building the telemetry events stays in the handler: those need message_data, and moving them would couple the service to the transport message shape. Happy to move them if you disagree.

The algorithms are untouched. src/server/offloading_algo/ has zero changes in this PR — only the caller moved. Same OffloadingContext, same create_offloading_algorithm, same select_offloading_layer().

Tests for the hot path

#45 called handle_device_inference_result "the most critical component [with] no dedicated tests". tests/unit/test_handle_device_inference_result.py now covers:

• the EMA update for device layers and for edge layers (correctly offset past the offloading point), and that untouched layers keep their seed;
• variance measurements landing in the per-device detector;
• both edge-inference skip conditions (offloading_layer_index == -1, and past ultimo_layer);
• the offloading decision: the layer is returned, forced-local bypasses the algorithm entirely, the reported network speed is passed through, the previous speed is kept when the packet reports none, and the decision sees the freshly smoothed times rather than the pre-update ones;
• background I/O scheduling for every output, with each queued task actually executed and its effect asserted;
• that the debug snapshot is immune to later mutation of the live tables — the race Extract per-device state out of RequestHandler (#11) #74's snapshot exists to prevent;
• that telemetry is skipped entirely when both outputs are disabled.

The handler is built with object.__new__, so none of the heavy __init__ (profiler, config loading, Firestore) runs.

Result

	before	after
request_handler.py	1158 lines	933
handle_device_inference_result	322 lines	279
request_handler.py coverage	48%	61%

287 passed, 1 skipped. device_state.py, inference_recorder.py and offloading_service.py all at 100% coverage. ruff clean.

Being straight about what is left: the handler is meaningfully thinner and its state is now owned elsewhere, but handle_device_inference_result is still 279 lines — the bulk is telemetry event assembly. Every box on #45 is ticked and the god-class state problem is solved, but if you want that method genuinely short, extracting an event-builder is a natural fourth step. Say the word and I will file it.

Note

Two tests moved out of test_request_handler_helpers.py into tests/unit/test_offloading_service.py, since they targeted methods that now live on the service. They gained coverage of branches nothing exercised before: the IndexError metadata fallback, forced-local, and the switch_penalty derivation (including candidates excluded from selection).