CI: provide a non-publishing PR lane for clean-wheel model smoke

This was generated by AI during triage.

Problem

The repository has a real clean-wheel model smoke (run_wheel_model_smoke) that creates a fresh venv, installs the built py312 wheel, builds Qwen3-0.6B, inspects the bundle, and runs inference. However, this detector is only scheduled by .github/workflows/nightly.yml; label-triggered PR CI does not run the wheel-model-smoke stage.

This blocks admission of the approved RedTeam-S9-WHEEL probe. A disposable draft PR cannot verify that a source-tree dependency is rejected by the clean-wheel model path:

• run-ci builds and installs the wheel but only exercises generic import, ELF, version, and help checks before source/selective testing.
• Manually dispatching nightly on a probe branch is unsafe: if the detector escapes, the workflow proceeds to publish the intentionally defective wheel as a prerelease.
• Modifying the probe workflow to suppress publication would add a second causal change and violate detector independence.

Evidence

• .github/workflows/trtmc-ci.yml has no wheel-model-smoke step.
• .github/workflows/nightly.yml runs wheel-model-smoke immediately before artifact upload and prerelease publication.
• .github/scripts/run-trtmc-ci.sh::run_wheel_model_smoke is already a suitable clean-venv detector and uses the model-owned tests/e2e/models/qwen/package_smoke.json configuration.
• Current overlapping PR #377 proposes a different model-scoped premerge architecture and removes global wheel stages; any solution here must be reconciled with that PR rather than duplicated blindly.

Required Outcome

• Provide a non-publishing, PR-head-exact path that can run the clean-wheel model smoke for package-impacting changes or an explicitly requested RedTeam probe.
• Surface its result in a PR check that cannot be satisfied by source-tree selective E2E.
• Install and execute only the wheel in a clean environment, with no source-tree fallback.
• Preserve nightly prerelease publication behavior for clean main runs.
• Add tests proving a wheel that depends on checkout-only files fails this lane.

RedTeam Impact

RedTeam-S9-WHEEL remains Not Admitted until this detector can run safely on a disposable draft PR. No mutation branch or probe PR should be created before then.

Start with .github/workflows/trtmc-ci.yml and nightly.yml, then read run_wheel_model_smoke in .github/scripts/run-trtmc-ci.sh and tests/e2e/models/qwen/package_smoke.json. Reconcile the approach with PR #377 before changing the workflows. Done means a PR-head-exact, non-publishing check runs the wheel in a clean environment, proves checkout-only dependencies fail, and leaves nightly publication unchanged.