Bridge actual numeric evaluations into naming fixture proposals

[Exohayvan]

Connect the existing numeric prototype's actual evaluator outputs to the naming fixture without training a new model. Keep synthetic-quadratic-expert as the identity; Numeric is only an internal fixture alias, not Ash/Ember or an assigned production family. This example stays outside SDK runtime orchestration.

Run the trusted existing worker's evaluation operation exactly three times on documented fixed coefficients: baseline, fitted polynomial and regression control. Independently recompute every rational prediction, verify exact task/output shape and source/context bindings, then derive fixture reports and pins. The fitted artifact measures held-out SSE 250 → 0 with zero regressions and yields an eligible fixture proposal. The regression control measures SSE 4 with one regressed case and yields an experimental proposal without a name/version increment. No trainer/solver or learning-journal operation is invoked by the bridge.

Commit exact baseline/candidate artifacts, held-out/regression suite, schema, metric semantics and worker/demo/bridge source revisions. Reject altered task/prediction/artifact/context evidence and source drift before/after execution; reject non-integral SSE instead of rounding it. Retain original outputs, translated reports, policy and pins for inspection. Existing helper pipe/deadline/CPU/cancellation bounds apply. No downloaded code, credentials, network service, model loading or user configuration is involved.

Local arithmetic verification is real; evaluator identity/independence and publication authority are not established. Embedded fixture decisions stay explicitly simulated, with no committed promotion/name reservation/live capability proof. Outer output explicitly denies training, authenticated evidence, independent operators, production family assignment and production promotion. The suite remains tiny, fixed, public and synthetic; zero arithmetic uncertainty is not a generalization claim.

Five focused tests (10 subtests) exercise actual helper outputs, evaluation-only calls, accepted/rejected proposals, bindings/tampering/source drift/cancellation and non-lossy metric conversion. Python 3.11 focused tests pass with ResourceWarning as error; bridge-focused required coverage reaches 100%. The new installed acceptance executes the bridge against the built SDK with Python -I outside the checkout. Existing repository regression tests retain their original tiny synthetic training checks; the bridge does not call those training paths. Native fast gate passed: 1,362 tests / 2,654 subtests, required coverage, lint/format, strict typing and repository policy checks. Wheel/sdist build and installed acceptance passed, including the new evaluation-only bridge. No new E-level complexity. All 24 fresh hosted checks passed on the reconciled exact head, including build/install, coverage, security and repeatability. Mutation not run. Independent review cleared the original head; a normal guarded merge is authorized after fresh checks pass.

Review head: 331cbaf7b0fe7b509242d7822db1f25615820ede; tree dcac70777b09b3d454262135aa668bbccd6f2984. Base: merged PR325 7c8e5324bd7d9b64d7fb183ede4b8c06a7208277.

Reconciliation proof: original reviewed head b9a76ad342392bc822c0311afa192d9d5e041a5c and rebased head have the identical full tree dcac70777b09b3d454262135aa668bbccd6f2984. The five-file feature patch is byte-identical. Fresh native fast gate passed again (1,362 tests / 2,654 subtests).