[gspark-etri]

Background

Public files contained internal cluster node hostnames and a personal home-directory path (#6). Most were in tests, goldens and docs. One spot used the home path in logic: the RBLN hostPath fallback (compile Job in src/app/compile/manifest.rs and the serving Deployment in src/app/deploy/serving.rs) mounted the node's pip install --user site-packages from a hardcoded home directory and put it on PYTHONPATH.

Changes

• Site parameter instead of a hardcoded path: new LMD_RBLN_HOST_SITE_PACKAGES (read in App::new, like the other LMD_* image settings, and passed through build_compile_manifest and serving::Images). When it is set, the compile Job and serving Deployment mount that directory at the same path and put it first on PYTHONPATH, the same as before. When it is unset, only the system Python locations are mounted and no home directory is assumed. lmd-top can't find a remote node's home directory while it generates a manifest, so this is configuration, not detection. It is documented in README.md, README.ko.md and cluster.example.env, which already holds the untracked site values.
• Aliases: replaced every internal hostname and home-path occurrence (25 files, about 70 lines) with npu-node-1, npu-node-2, control-node, gb10-node-1, gb10-node-2 and /home/<user>. This covers unit tests (agent.rs, app.rs, collect.rs, coverage.rs, pod_health.rs, probe.rs, ui/tests.rs), goldens, docs, CHANGELOG.md, STATUS.md, .research/events.jsonl, assets/npu-lora/backends.py and the epp-score-observer Dockerfile comment.
• Goldens: the golden fixture sets the site value to /home/<user>/.local/lib/python3.10/site-packages, so the four RBLN compile/deploy goldens keep the same structure; only that string changed. A new test, rbln_host_fallback_assumes_no_user_site_when_unset, covers the unset case for both compile and deploy.

Verification

• On the project node (cargo 1.96), from a fresh clone of this branch:
  • cargo clippy --locked --all-targets -- -D warnings: clean (exit 0)
  • cargo test --locked: 448 passed, 1 ignored; 11 passed (integration); exit 0
  • cargo fmt --check: clean (fixed in a follow-up commit after the first CI run)
• GitHub CI (check, epp-plugin): green
• Same output as before with the real site value: I set the golden fixture to the real site-packages path and ran UPDATE_GOLDEN=1 cargo test golden. All four RBLN goldens (compile-rbln-{llama,qwen}, deploy-rbln-{llama,qwen}) came out byte-identical to main.
• A case-insensitive search of the tree for the internal names (taken from a private list) finds nothing. check_sensitive.py: clean.

Compatibility and limits

• Operator action: on a cluster that relies on the RBLN hostPath fallback with a pip install --user vendor stack, add LMD_RBLN_HOST_SITE_PACKAGES=<that site-packages> to cluster.local.env. Without it, the fallback mounts only the system locations. Setups using LMD_COMPILE_IMAGE_RBLN or LMD_RBLN_TOOLCHAIN are unaffected.
• build_compile_manifest gains a host_site parameter and serving::Images gains a rbln_host_site field. Both are internal.
• Not changed: links to the old GitHub repo location (gspark-etri/lmd-top in Cargo.toml, install.sh, plugins/lmd-top.yaml, docs). That is an account name rather than a node, and updating it is a separate change.
• STATUS.md / .research/events.jsonl are written by the research-hub reporter, which records the machine's hostname. Future reports can bring a hostname back until the reporter is changed (hub side).
• Git history is not rewritten.

Related Issue

Closes #6

🤖 Generated with Claude Code