fix(config): write config atomically via temp file + rename

[sachiniyer]

Summary

• Fixes [Detail Bug] CLI config reads can fail due to non-atomic config file updates (truncated TOML race) #270 (Detail bug bug_2e491e1e-0d71-42df-9ca5-ee5a975ac314): concurrent CLI invocations could fail with Failed to parse config when one process was mid-rewrite of config.toml and another tried to read it.
• update_config truncated the file in place (set_len(0) + write_all) and load_config read it without any lock, leaving a window where readers observed empty or partially-written TOML.
• Now writes stage to a sibling temp file (in the same directory, so rename is atomic) and rename onto the target. Readers always see one complete version. A separate config.lock serializes writers — we can't lock the config file itself anymore, since rename swaps in a new inode and would leave concurrent writers locking unrelated dead inodes.
• Unix mode is preserved (mirrored from any existing config; otherwise 0600 since the file holds an API token). On Windows std::fs::rename already replaces the target via MoveFileExW.

Test plan

• cargo fmt --check, cargo clippy -- -D warnings, cargo test (all 291 lib tests + 17 integration tests green).
• New concurrent_reads_during_writes_never_observe_partial_toml stress test (4 readers / 1 writer, 16 KiB rewrites). Verified it reliably fails on the pre-fix code with the truncated TOML errors from the bug report (e.g. key with no value, expected '=' partway through api_token) and passes with the fix in well under the 30 s safety cap.
• Existing tests covering comment preservation, key removal, empty-file path, and writer-vs-writer concurrency still pass.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 799a99b422

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Preserve symlinked config files on write

When config.toml is a symlink (common with dotfile managers), renaming the staged file over path replaces the symlink itself with a regular file instead of updating the symlink target. The previous File::create/OpenOptions writes followed the link, so after this change those users silently stop updating their managed config and leave the real target stale; resolve the symlink target or stage/rename at the target path before publishing.

Useful? React with 👍 / 👎.

[cubic-dev-ai]

No issues found across 1 file

[sachiniyer]

Closing — bug #270 is real but the impact is rare (~0.005% per the report) and the regression test was hard to make reliable on CI. Not worth shipping right now; can revisit if it surfaces in practice.