fix(specgit): restore delivery state after failed bootstrap

[LeXwDeX]

Closes #530

Why

A failed specgit issue can delete or rewrite .specgit.yaml before returning nonzero. That leaked partial delivery state into the next attempt and obscured the original failure.

What changed

• Snapshot the pre-run delivery record separately from the temporary harness surface.
• Restore and hash-verify .specgit.yaml only when init or the inner issue command fails or is interrupted.
• Preserve successful bindings, original failure status, and diagnostics.
• Fail closed with exit 3 and retain the forensic snapshot when record restoration cannot be verified.
• Added isolated cases for validation, branch, push, and PR failures, success persistence, dirty/untracked records, restore mismatch, and non-rollback of commits.

Evidence

• bash script/specgit-bootstrap.test.sh: latest parent run 74 passed, 0 failed.
• sh -n / bash -n: passed.
• shellcheck -s sh / shellcheck -s bash: passed.
• Commit hooks: root lint and all 29 typecheck tasks passed.
• DAG verification: PASS; diff arbitration: ACCEPT with no required actions.
• The verifier observed an intermittent pre-existing macOS SIGINT fixture failure in case4b; its control worktree reproduced the same failure before this diff. All 36 new fix(specgit): restore delivery state after failed bootstrap #530 assertions passed in every run.

Checklist

• Successful bootstrap keeps the new binding.
• Failed bootstrap restores pre-run record bytes.
• Branches, commits, and remote side effects are not undone.
• specgit finish exits 0.