Follow-ups from #885: auto-commit and workspace provider on an unborn HEAD #911

Follow-ups from #885 (merged as a9c4076; closes #835), collected from the adjudicator's passes. One tracking issue per PR; each item becomes its own PR when picked up.

• Flair-delivered auto-commit has no unborn-HEAD gate. _runAutoCommitLegacy runs tps agent commit directly; on a clean unborn tree that switches HEAD to task/<id> and then exits with "No changes staged". Give it the same gate the mail-task path uses.
• Test the onTaskFailure paths on an unborn branch. The unborn diff guard is untested (removing it only changes filesChanged from undefined to [], which no test asserts), and there is no test of the dirty-unborn failure checkpoint that makes the branch's first commit.
• One git binary for GitWorkspaceProvider. HEAD classification uses GIT_BIN (/usr/bin/git, TPS_GIT_BIN override) while the provider's other git calls use git from PATH; on a host without /usr/bin/git and with TPS_GIT_BIN unset, every snapshot throws. Decide one resolution for the whole provider (main's agent.ts and codex-runtime.ts already use GIT_BIN).
• Name the probe in baseline()'s indeterminate-HEAD error. It gives the reason but drops which probe returned it.
• Auto-commit off: unborn vs born asymmetry. A completed task on an unborn branch stays uncommitted, while one on a born branch gets a checkpoint commit. Decide whether that is intended and document or align it.
• Real-Flair harness for the WorkspaceState / OrgEvent writes on this path (today a fake fetch through the real client).

🤖 Generated with Claude Code