[prabod]

superfluid serve org/model:<variant> ran basert pull org/model before serving, even when that variant was installed. basert pull takes no variant, so it fetched the default quant instead, and for a repo with no default quant it failed and serve exited:

superfluid: pulling basecompute/Qwen3.8-27B-BaseQuant-V1:file-qwen3.8-27b-basequant-v1-medium-6583221f with the basert runtime
Error: no pre-converted .base for quant "q4" in this repo; it offers: large, medium, small, xl, xs (or take one by name with --file: …)

The model had been pulled with --pull-file Qwen3.8-27B-BaseQuant-V1-Medium.base, so it was installed as file-…-medium-<hash>. The only way to serve it again was --offline, or --pull-file, which applies to every --model and so breaks a second one (Whisper beside it: no .base named "Qwen3.8-27B-BaseQuant-V1-Medium.base" in this repo). Local hit this as an endless restart loop.

Now a variant named with the id that basert list --json shows as installed is served as --offline serves it, with no pull. A bare id still pulls (its default quant may be newer), as does a variant that isn't installed, and --pull-force always pulls.

Testing

• cargo test -p superfluid-adapter-basert: 22 pass, including the new an_installed_named_variant_is_not_pulled_again
• cargo clippy -p superfluid-adapter-basert --all-targets -- -D warnings is clean
• End to end on an M-series Mac with libbaseRT 0.4.1: superfluid 0.1.4 with SUPERFLUID_WORKER_BASERT set to this branch's worker served basecompute/Qwen3.8-27B-BaseQuant-V1:file-qwen3.8-27b-basequant-v1-medium-6583221f --model basecompute/whisper-tiny online. It listed both models and answered a chat completion; the shipped worker fails on the same command as above.
• Not run: the full workspace cargo test, and Linux/CUDA.