fix(daemon): skip side-effect errors when the repo was deleted mid-operation

[posthog]

Problem

• The team gets error-tracking noise. Each temp-repo teardown race opens a new single-event issue that somebody must triage. No user attribution or data is affected.
• After a commit, the daemon runs git diff between the commit and its parent. When the repo is deleted first, git exits 1 with Could not access <oid>.
• is_missing_working_dir_error matches only exit-128 stderr text and NotFound IO errors. So this shape gets to both log sites (command side effect failed, async side-effect error).
• Known sources: the autter doctor attribution self-check deletes its repo without waiting for daemon side effects, and developer tests delete temp repos right after they commit.

Changes

• Classify by repo state, not by error text. The new is_vanished_repo_error treats any GitCliError or IoError as benign when the command's worktree, its git dir HEAD, or its objects dir no longer exists. The existing is_missing_working_dir_error text match is unchanged and stays as a fallback.
• repository_is_gone resolves the git dir with the existing git_dir_for_worktree, so a subdirectory of a live repo resolves to that repo and still reports real errors.
• Doctor self-check waits for side effects. FamilyStatus gets inflight_effects (with #[serde(default)], so older daemons read as 0). The self-check polls it until the value is 0, within its current deadline, before remove_dir_all.

Error shape	Before	After
exit 128 No such file or directory / not a git repository	skipped	skipped
IoError NotFound	skipped	skipped
exit 1 Could not access <oid>, repo deleted	reported	skipped
exit 128 not a tree object / bad object, objects deleted	reported	skipped
any git error, repo still on disk	reported	reported

Note

Draft #64 adds more exit-128 strings to is_missing_working_dir_error. When the objects dir is gone, this PR also covers those shapes. This PR does not edit that function, so both can merge. Draft #66 changes the two log sites, which this PR does not touch.

Testing

• cargo test --lib -- missing_working_dir repository_is_gone: 6 passed (2 new tests: exit-1 Could not access with a live, objects-deleted, and fully deleted repo; and a subdirectory of a live repo).
• cargo test --lib daemon: 139 passed. cargo test --lib diagnostics: 9 passed.
• cargo clippy --all-targets -- -D warnings and cargo fmt --check: clean.
• doctor_validates_checkpoint_round_trip_end_to_end fails in the sandbox with git commit ... (status=1). It fails the same way on unchanged main, so the cause is the sandbox environment. CI must confirm it.

Agent context

• I did not match only paths under the self-check root, because developer temp repos would still make noise.
• I did not add a list of stderr strings, because each new git error shape would need one more string.
• Not in scope: the checkpoint side effect failed log site has no vanished-repo guard, and inflight_effects does not count commands still queued in the family sequencer. Both are possible follow-ups.

---

Created with PostHog Desktop from this inbox report.

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}