feat: support compose configs via a shared podman store module

[lesnik512]

Summary

Add compose configs support (Bucket B). Configs are the non-sensitive sibling of secrets: same shape, but they mount at the container-root path /<name> (not /run/secrets/<name>) and add an inline content: source.

Rather than duplicate the secret machinery, this branch extracts it into a shared, kind-parameterized compose2pod/store.py (a frozen StoreKind), so secrets.py and configs.py shrink to thin kind-specs and emit.py/parsing.py weave both through one union seam over STORE_KINDS = (SECRET, CONFIG). Because the secret kind keeps store prefix "", the extraction is byte-identical for existing documents.

Design: planning/changes/2026-07-10.03-configs.md.

What's supported

• Top-level configs: with exactly one string source: file: (host path, resolved against --project-dir), environment: (host env var), or content: (inline literal).
• content: runs through to_shell(), so ${VAR} inside it expands at script-run time (the tool's deferred-interpolation model) and its var names feed the stderr interpolation note.
• Service configs: short form [name] and full long form {source, target, uid, gid, mode}.
• Delivered via podman's secret store, pod-namespaced <pod>-config-<name> so a config never collides with a same-named secret; created after podman pod create, removed in the EXIT trap.
• Default target is the container-root /<name>; a long-form target: must be absolute (relative targets rejected). external: true rejected.

Architecture

store.py (leaf) owns the machinery parameterized by StoreKind{label, top_key, prefix, sources, default_target, require_absolute_target}. secrets.py/configs.py are the two kind-specs; stores.py aggregates STORE_KINDS. emit.py/parsing.py call the kind-unioning entry points, so adding a kind is a registry change, not new emit logic.

Security

Config/secret names and env-var names flow into generated shell raw, so they are charset-validated with re.fullmatch (^[a-zA-Z0-9][a-zA-Z0-9_.-]*$ / ^[a-zA-Z_][a-zA-Z0-9_]*$) in the shared gate. The new content: surface was adversarially probed: to_shell neutralizes " ` $ \ and never matches $(...), so arbitrary content cannot break out of its double-quoted printf fragment while legitimate ${VAR} stays live. Long-form target is shlex.quoted by the emit renderer. An independent whole-branch review confirmed no injection path.

Testing

just test-ci at 100% coverage (263 tests), just lint-ci clean, just check-planning OK. The ported secret suite (test_store.py) is the byte-identical guard for the refactor; test_configs.py covers file/env/content create lines, /<name> and absolute targets, relative-target rejection, and content var interpolation; test_emit.py proves the unioned trap and that a same-named secret+config produce distinct stores. architecture/supported-subset.md gains a Configs section and its stale post-refactor secret citations were corrected.

Known minor limitations (parity with secrets, non-blocking)

• A comma in a long-form target is interpreted by podman as extra --secret sub-options (arg-level, not shell injection; the compose author controls both inputs). Same as secrets today.
• A non-string target bypasses the absolute-path rule (podman rejects it at runtime). Same target-passthrough behavior as secrets.

Fixing these for configs only would diverge the two kinds; left as a possible future hardening pass covering both.