[Yushu-Li]

The upstream signal-processing entrypoint omits model configuration, queries a diagnostic experiment before creating the experiment it actually runs, passes sampler/evaluator counts to the controller, and orders results by generic score. These calls prevent it from using the existing local walkthrough.

Extend run_entrypoint.py to execute its original main() with byte-identical staged sources. A signal-specific experiment subclass supplies the explicit --model without mutating the upstream configuration. The runner requires the reviewed diagnostic-then-completed lifecycle, so caught setup failures and early returns cannot appear successful. During this main, SAVE_PROGRAM_DIR points to the staged programs/ directory and is restored on success or failure, keeping optional evaluator archives away from inherited destinations.

The isolated local facade adds a client-scoped, in-memory experiment lookup returning only name/state. Generic score ordering refers to the explicitly selected objective. Positive num_samplers/num_evaluators arguments are accepted with a warning that they are ignored: synchronous execution has already finished, and the controller neither reruns evolution nor creates workers. Other cloud query/controller options remain unsupported. Existing seed-only defaults, explicit endpoint/key selection and best-only results are preserved.

Validation:

• Nine new tests cover signal-processing's original entrypoint shape, actual OE seed/worker execution, explicit model precedence without configuration mutation, diagnostic failures/early returns, archive/environment restoration, layout drift, local metadata lifecycle/isolation, score ordering and controller boundaries.
• Full offline suite: 881 tests passed, including the existing local-loopback HTTP worker/resume test. Black/isort, Python 3.10 grammar checks and Frame scans of all six changed Python files passed.
• Executed the original signal-processing, circle-packing and TSP main() at official revision 34cd919193945c6dc81053f9af594031b54f41a8, each seed-only and with one scripted proposal through real OE workers. Original and staged Python source hashes remained unchanged, SDK imports were blocked, temporary aliases were cleaned up, and guarded socket/HTTP/credential-lookup attempts were zero.
• Signal processing returned overall_score=0.49904861783269006 with success_rate=1.0 in both runs; its original evaluator archive confirms the scripted candidate was evaluated. Both earlier examples also retained their reporting flows. Validation used Matplotlib 3.10.9 as documented by the existing walkthrough; no real model or managed-service requests were made.

Continues #468 without closing it. This covers the three reviewed local entrypoints, not arbitrary SDK examples, cloud concurrency semantics or a full archive API. Signal-processing archive filenames can be reused by the upstream saver. Fixed proposals validate integration rather than optimization quality. Trusted code retains normal process privileges; no sandbox, strict cost cap or base dependency changes are introduced. Dedicated upstream-example CI, complete round-trip run coverage and actual managed-service validation remain follow-ups.