Single owner for edge timings and variance (#72, #73)

[ff225]

Closes #72. Closes #73.

Stacked on #74 — base is fix/sciot-011-device-state-manager, not main. Review #74 first; this diff is only the second commit.

Two tangled problems in the same ~15 lines, so the fixes land together.

#72 — the edge EMA was applied twice

track_inference_time applied an EMA with a hard-coded alpha = 0.2 to self.inference_times. That dict is device_profiles[device_id]["edge_inference_times"], passed by reference (edge_initialization.py:140,152 → model_manager.py:285). The request handler then applied the configured EMA on top of it.

Two consequences:

• Over-smoothing. Edge times were smoothed twice, device times once, so the two sides of the offloading cost comparison responded to change at different rates — the edge estimate lagged and the split point adapted slower than intended.
• The config was bypassed. Make EMA alpha configurable instead of hard-coded 0.5 #9 made ema_alpha configurable, but the first application ignored it.

The fix turned out simpler than the issue suggested. Edge.run_inference already measures each layer and returns the raw times — the decorator was cronometering the same call a second time. So rather than reconciling the two paths, one is removed: the decorator is now instrumentation only, and DeviceStateManager.update_edge_times is the single place that smooths, with the configured alpha.

#73 — a second variance detector nobody read

Edge.variance_detector was a process-global detector fed from the same decorator. No production code calls should_retest_offloading() on it — every such call is in tests. It burned time on the hot path and logged re-test warnings that triggered no re-test. Removed; the per-device detectors from #74 are now the only ones.

Fallout cleanup

With the decorator no longer writing, ModelManager.inference_times had no writers left, so the plumbing goes: the inference_times/variance_detector params, the edge_inference_times argument threaded through Edge.run_inference and _get_model_manager, save_inference_times (no callers), and a commented-out __init__ referencing attributes that no longer exist.

Two judgement calls worth a look

• time.time() → perf_counter() in run_inference. Now that this is the authoritative measurement of edge cost, a monotonic clock is the right tool — but it is a change nobody asked for.
• scripts/analysis/variance_analysis.py moved to the per-device API. Extract per-device state out of RequestHandler (#11) #74 broke it (it read RequestHandler.variance_detector) and I missed it there. Worth knowing: that script reads the running server's in-memory state from a separate process, so it always found an empty detector. This makes it coherent, it does not make it useful.

Testing

244 passed, 1 skipped (fast suite + HTTP end-to-end). New tests/unit/test_edge_inference_time_ownership.py pins the single-EMA behaviour and the configured alpha. ruff clean on touched files (the 8 findings in variance_analysis.py are pre-existing — confirmed against main).

⚠️ Edge.run_inference is not exercised by the fast suite — its real tests are marked model_artifact. The regression tests here cover signatures and the EMA maths, but the edge inference chain itself has not been run end-to-end. Worth a run with the model artifacts before merging.