Fix O4c container-universe transfers and CPU image fallbacks

[oshaughnessy-junior]

GPU container-universe ILE and CALPILOT jobs on rift_O4c put full URLs inside container_image. HTCondor truncates that selector before match-time expansion, which can hold every worker. This ports the basename-plus-URL-transfer fix from #165 and pins MY.TransferInput to avoid a spurious basename input.

CPU ILE and CALPILOT now use a literal fallback image with no GPU-dependent transfer expression. CPU ILE keeps --gpu --force-xpy --vectorized for the NoLoop likelihood path; these executable options are independent of requesting a physical GPU. CPU calibration reweighting resolves a manifest to its fallback instead of trying to launch the YAML as an image, matching #164's calibration fix. GPU families with local/CVMFS entries fail during generation with actionable guidance.

Validation outside GitHub CI

• Initial combined container-manifest, CIP startup, and DAG postprocessing gate: 41 passed. After the independent review correction, the expanded container suite passes 33 tests, including eight condor_submit -dry-run cases covering list/string/empty-string ILE inputs and CPU/GPU ILE/CALPILOT ads.
• Against the unfixed base, the new effective-ad cases give 3 failed, 1 passed.
• A live CPU control using the merged selector was held because the GPU-capability transfer expression could not expand. The fixed CPU fallback and GPU-selector jobs both exited 0. These small container probes validate image deployment, not scientific ILE inside the probe images.
• Separate real BBH ILE smoke evaluations on a GPU and on the CPU NoLoop path both exited 0 with finite result rows and parseable sample XML. These are operational smoke tests, not convergence validation.
• git diff --check passes.

Independent adversarial review found that string-valued transfer_files concatenated the image selector directly onto a filename. The follow-up commit adds the missing comma and tests that the selector is a distinct transfer item. Independent re-review passed at aeb426d7d1df93fa46ec2d07a440c4ede87c4eb4, with ten additional real HTCondor dry-run cases covering CPU/GPU jobs and list, single-file string, multiple-file string, empty-string, and empty-list inputs. No remaining blocker was found in this PR delta.

Scope

This fixes image delivery and CPU fallback selection without changing likelihood code or normalization. Separate adversarial findings remain: shared selectors do not enforce capability upper bounds, and the optional runtime wrapper added in #164 can select by the first physical GPU instead of the assigned GPU. O4c does not implement that wrapper. The deployment checks above do not validate those behaviors.

[interhost-cross-review-canary-v2]

Independent automated review completed at the recorded exact commit. Detailed findings were withheld from public output by the private-context egress policy and require private human declassification.