[multi-lora] 1/n - 2, ci: verify import targets in the source-resolution harness step

[yushengsu-thu]

Extends tests/ci/verify_source_resolution.py — the pre-test step every CPU and GPU CI job already runs — with import-integrity checks: statically resolve every miles-internal import site (module-level, function-local, and relative) across miles/ and examples/ against the source tree, walk the optional multi-lora namespaces in full when present (absent packages skip cleanly on main), and import the update_weight/ lazy-import targets. Failures raise RuntimeError with the offending file:line and import target; missing non-miles optional deps stay tolerated.

Running this in the harness step instead of a fast test covers every job (CPU and GPU) before pytest even starts, which is broader than the fast suite. Born from the mixin.py dead lazy-import incident, where a moved module was only referenced from a function-local import and nothing failed until runtime.

Extracted from #2273; the stack rebases after this merges.

[claude]

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.

_{Tip: disable this comment in your organization's Code Review settings.}

[guapisolo]

I think this PR should be narrowed to the third motivation only: protecting update_weight function-local/lazy imports from going stale.

The current change mixes three concerns: repo-wide AST linting, optional-namespace import smoke tests, and update_weight lazy-import integrity, then runs them in every CPU/CUDA job. The mixin.py incident only justifies the last one. The broader resolver is also incomplete (from package import submodule checks only the package), while namespace imports add side effects and environment coupling.

Suggested minimal cut:

• statically inspect only actor.py and update_weight/**/*.py for function-local Miles imports;
• resolve targets without importing backends, so Mooncake/RDMA remains lazy;
• preserve file:line and add a negative fixture proving a missing mixin.py fails;
• run the check once in one CPU preflight job.

I would remove the repo-wide miles/ + examples/ scan, OPTIONAL_PACKAGES walk, and all-job fan-out. A general repository import-integrity linter can be a separate PR with explicitly defined import semantics and its own focused tests.