fix(filters): always drop .netrc/_netrc credential files from lineage

[christophergeyer]

What

Always filter .netrc / _netrc from provenance — reads and writes — regardless of filter config. These hold machine login tokens (wandb, huggingface, curl/requests) and are not reproducibility-relevant.

Why

A nanochat e2e run on Lambda (verifying #133 + #134 against dev) surfaced ~/.netrc as a tracked artifact: wandb reads it at startup, so the tracer captured it. Only the path+hash reach GLaaS (not contents), but a credential file appearing in lineage is undesirable and confusing. It belongs in the same always-dropped bucket as .git metadata and roar-internal files — not gated behind filters.* config (a user clearing ignore_paths must not re-expose it).

Change

New credential_files filter category (parallel to roar_internal / git_metadata):

@staticmethod
def _is_credential_file(path: str) -> bool:
    basename = path.rsplit("/", 1)[-1]
    return basename in (".netrc", "_netrc")

Applied in both categorize_read (inputs) and the written-files loop (outputs). Surfaced in run-report verbosity output like the other categories.

Tests

• test_file_filter.py: .netrc and _netrc dropped from reads + writes.
• test_verbosity.py: credential-files count rendered in the report.

Scope

Deliberately minimal — just the netrc family (the observed leak). Follow-ups could add .aws/credentials, .ssh/id_*, .config/gcloud/, etc., but those merit their own discussion (some are already covered by system-path filters; SSH keys are trickier since pubkeys are sometimes legitimately tracked).

🤖 Generated with Claude Code