…--as

- --role was never a CLI flag; replace with --persona <file> (the real
  working-style mechanism; see examples/personas/).
- omit --as from all example commands (the agent id auto-generates); surface
  the generated ids via h5i team status and reference them in downstream
  --agent flags. Keep --as in the flag reference, marked optional.
- document team submit's new auto-snapshot + no-op refusal; add the missing
  team persona row. Regenerate docs/manual from MANUAL.md.

The host can't snapshot a *live* box: the team Stop hook keeps a box alive
for the whole round, so it holds the env run lock the entire session, and the
at-exit ingest never runs mid-round. An agent that writes files but never
commits (codex left quick_sort.py/pyproject.toml as untracked `??` files)
therefore staged a no-op the host refused as 'identical to the team base' —
freeze blocked, the round wedged.

Fix: the in-box `team agent submit` now commits the worktree onto the env
branch itself (commit_box_worktree), where the box has a functional checkout
(rw on its own branch + objects). The host ingest then freezes the advanced
tip — exactly the path a well-behaved agent already takes by committing
in-box. Best-effort: any git error degrades to a no-op so submit is never
blocked; an explicit --commit skips it.

- env::commit_box_worktree / commit_worktree_at (testable, path-based).
- main: call it on the in-box no-explicit-commit submit path.
- test: commit_box_worktree_snapshots_untracked_edits.

commit_box_worktree returned Option and used .ok()? throughout, so ANY git
error inside the box (e.g. a too-narrow box_git_plumbing grant) degraded to a
silent no-op — indistinguishable from a clean worktree, and exactly what makes
an agent's work vanish into a 'no changes to review' submission.

Return Result<Option<Oid>> and have the in-box `team agent submit` print a
clear warning on Err (while still staging — a failed snapshot must never block
the submit; the host then freezes the branch tip). Next run self-diagnoses:
either '▢ snapshotted worktree in-box at <oid>' or a concrete error to act on.

Follow-up to the in-box submit snapshot. Two parts:

1. commit_box_worktree no longer calls git2 index.write() before committing —
   the proven-working `h5i capture commit` path never persists the index from
   git2 (the agent's `git add` does), and writing worktrees/<wt>/index via a
   lock+rename is the step most likely to EACCES in a tight box layout while the
   object + ref writes succeed. Land the commit from the in-memory index
   (write_tree + commit), then refresh the index best-effort. Also return Result
   so a real failure is surfaced (warning) instead of silently becoming a no-op.

2. h5i init guidance (AGENTS.md / .claude/h5i.md templates + this repo's copies):
   'In an agent team, always h5i capture commit before h5i team agent submit.'
   This is the reliable mitigation — a committed worktree submits correctly
   regardless of whether the auto-snapshot can run (it's what well-behaved
   agents already do).