Keep recorded branch labels stable across re-imports

[starkdmi]

Summary

A project's branch label came from the record when it carried one, and otherwise from the checkout's HEAD at scan time. HEAD at scan time says nothing about where a past session ran, so a re-import made while a repository sat on another branch relabeled its history and moved those sessions' rollups to that branch.

• Codex: session_meta records the branch, but turn_context repeats only the cwd. Each turn_context resolved the project again without a branch, fell back to HEAD, and overwrote the recorded branch, so almost every session took the checkout's branch. A turn in the directory session_meta named now keeps the remote and branch it recorded. A turn in another directory has no recorded branch.
• HEAD fallback removed. A branch is labeled only when the record carries one, and is otherwise left unset. A recency heuristic would still guess, and would make parse output depend on when the scan ran.
• OpenCode records no branch, so its sessions are no longer labeled.
• Claude Code and Grok Build archive scans now pass the branch their records carry, instead of relying on the fallback.
• Parser revisions bumped for Codex, Claude Code, OpenCode, and Grok Build. The first scan after upgrading rescans and re-archives them once, and drops every branch taken from a checkout's HEAD. Rollups for sessions whose branch changes are marked dirty once.

Verification

• cargo fmt -p statsai-adapters, cargo clippy -p statsai-adapters --all-targets -- -D warnings, cargo test -p statsai-adapters

• ./scripts/rust-ci.sh full (passed locally and in the pre-push hook)

• New adapter tests re-scan the same sessions with different branches checked out:

  • A Codex session keeps its session_meta branch and hash across checkouts.
  • A Codex session without a git block, and a turn that moved to another directory, never take the checkout's HEAD.
  • An OpenCode session keeps its branch unset across checkouts.

  The three Codex tests fail with the fix reverted.

• Ran the build against an isolated clone of a real store: a full rescan restored Codex sessions to the branches their records carry, and a second uncached rescan changed no branch label or hash.