Fix hardcoded model lookup and fragile parsing in split_inference (#47)

[ff225]

Decision: fix-in-place, not full pipeline integration

Closes #47.

I read through src/server/communication/http_server.py's split_inference handler plus ModelManager, RequestHandler, and OffloadingAlgo to judge feasibility. Verdict: fix the cited code-quality issues in place; do not fold this endpoint into the main RequestHandler/OffloadingAlgo/ModelManager pipeline in this change.

Why fix-in-place instead of full integration:

• The main pipeline is built around async device registration, per-device offloading-layer state, and telemetry/evaluation recording (RequestHandler.handle_device_inference_result, OffloadingAlgo, CSV/Firestore telemetry). split_inference is a synchronous "decode tensor in, run remaining TFLite submodels, return prediction" request with no device registration or telemetry semantics at all — reshaping it to fit the main pipeline's data model would be a disproportionate rewrite for what's still a validation/prototype endpoint, with real risk of destabilizing the main pipeline for no behavioral benefit.
• Importantly, this endpoint is not dead code: tests/test_accuracy/test_ImageWithBox.py uses it as ground truth to validate that server-side split inference matches local end-to-end inference, across all 5 FOMO model sizes (48/96/192/224/300). That test is gated behind SCIOT_RUN_ACCURACY_TESTS=1 plus model_artifact/external_server markers (so it doesn't run in normal CI), but the model artifacts it needs are checked into the repo and it is a real, working accuracy harness. Removing the endpoint would silently delete that coverage, which the issue does not ask for and which I don't think is the right call given the artifacts already exist and the test already works.

What changed, issue point by issue point

• Hardcoded model folder names (nomi_cartelle = {"48": "FOMO_48_CUT", ...}): removed. The handler now resolves model_dir from settings.yaml's existing model section (fomo_48x48.model_dir, etc.) via a new HttpServer._resolve_model_dir(size) helper — the same config section RequestHandler/ModelManager already read (request_handler.py:655). Single source of truth, no more drift risk between the two hardcoded dicts.
• Fragile size parsing (device_id.split("_")[1].split("x")[0]): replaced with a validating regex (HttpServer._parse_device_size) that returns None on a malformed device_id instead of raising an unhandled IndexError; the handler now returns a clean 400 with a descriptive message in that case.
• Hardcoded skip connections ({25: [16, 24], 43: [34, 42], 52: [43, 51]}): left in place, but now with a comment explaining why: these indices describe the FOMO submodel's fixed network topology as exported, not a deployment/runtime setting — there is nothing in settings.yaml or ModelManager that models network topology today, and inventing that config surface for one endpoint would be scope creep beyond this issue.
• Does not use RequestHandler/OffloadingAlgo/Edge.run_inference/ModelManager: documented above as an explicit, considered decision, not an oversight.
• Not covered by telemetry/evaluation, can diverge from main flow: still true and accepted — this endpoint intentionally stays a self-contained accuracy-validation surface, not part of the telemetry-tracked production flow.

Test coverage

Added tests/integration/test_split_inference_validation.py: fast, artifact-free tests (using the existing tiny_http_server fixture, no TFLite model files or external server needed) covering:

• malformed device_id (no size, trailing underscore, missing second dimension, empty) -> 400 with a clear message
• unknown/unconfigured model size -> 400 with the existing "non trovato" message
• the new HttpServer._parse_device_size helper directly

These run in normal CI (tests/integration/, no external_server/model_artifact markers), unlike the pre-existing artifact-gated accuracy test, so this endpoint now has always-on regression coverage for its request-validation paths.

Testing performed

• uv run pytest tests/integration/test_split_inference_validation.py -q — 7 passed
• uv run pytest tests/integration/test_http_end_to_end.py tests/integration/test_http_protocol_validation.py tests/integration/test_split_inference_validation.py -q — 15 passed
• Fast CI suite (tests/unit, tests/integration/test_client_resilience.py, test_client_lifecycle.py, test_variance_and_local_inference.py, tests/test_mqtt_client, tests/test_offloading_algo) plus the above HTTP suites — 199 passed, 1 skipped
• tests/test_accuracy/test_ImageWithBox.py --collect-only — still collects 5 tests unaffected (device_id format device_{size}x{size} still matches the new parser)