feat(git): route independent-clone reports onto teamai-reports

[jimpablo]

Summary

Independent git clones currently write members/ sessions/ votes/ stats/ onto the default branch via pushRepoDirectly, so every member needs write access on main. This reuses the existing teamai-reports orphan-branch worktree writer that single-repo mode already uses.

• Git-kind reports checkout is a sibling of the clone (<dirname(localPath)>/reports-wt), so clone reset --hard cannot nest-destroy it.
• kind: self keeps today's nested <knowledgeDir>/reports-wt.
• kind: http is unchanged (no git reports branch).
• Empty-repo init may still push teamai.yaml + gitkeeps to the default branch so the knowledge tree exists; member YAML goes to teamai-reports.
• Leftover report files already on main are not copied, not deleted, and not used as the source of truth after the switch.
• No new CLI command. Learnings stay on the default branch (Write learnings/ to a teamai-learnings branch (direct push, no PR) #485 is separate).

Closes #484

Type of Change

• Bug fix (non-breaking change that fixes an existing issue)
• New feature (non-breaking change that adds functionality)
• Breaking change (fix or feature causing existing behavior to change)
• Documentation only
• Refactor / internal cleanup

Test Plan

• npx tsc --noEmit — clean
• npx vitest run focused suites — 117 tests passed, including:
  • src/__tests__/git-kind-reports.test.ts (real git: write to origin/teamai-reports, leftover clone members/ ignored, empty-repo skeleton on default branch, dedicated-root guard)
  • src/__tests__/self-mode-no-business-reset.test.ts (git-kind no longer reset --hards the clone)
  • src/__tests__/members.test.ts (listMembers reads reports worktree; leftover clone YAML is not listed)
  • src/__tests__/team-push-interventions.test.ts (git-kind auto-report uses commitAndPushReports, not pushRepoDirectly)
• npm run build then built CLI (dist/index.js) against a local bare remote whose update hook rejects default-branch pushes:
  1. teamai init https://git.example.test/acme/team.git --scope user --force — English success; Member registered on the teamai-reports branch
  2. teamai session save --push --force --session-id s1 --scope user — Pushed: sessions/alice/2026-09.md
  3. teamai pull — auto-report wrote stats/alice.yaml
  • git ls-tree origin/teamai-reports contained members/alice.yaml, sessions/alice/2026-09.md, stats/alice.yaml
  • default-branch tree still only had teamai.yaml + knowledge gitkeeps (no new report files)
  • reports worktree was a sibling of the clone (~/.teamai/reports-wt); clone had no members/ stats/ sessions/
• EN/ZH docs/usage-guide.md updated so independent git clones also write reports to teamai-reports (learnings still on the default branch)

Related Issues

Closes #484

Notes for Reviewers

• Writers switched: init member registration, session save --push, pull auto-report, Stop-hook votes.
• Readers switched: members, digest, stats, viz, pull votes/stats index, recall, maintenance/paths, projects members (manifest still from knowledge clone).
• reportUsageToTeam still has a no-config fallback that pushRepoDirectlys the clone; production pull now always passes selfConfig for non-HTTP repos.
• Follow-ups: Write learnings/ to a teamai-learnings branch (direct push, no PR) #485 (learnings/ → teamai-learnings) and Document branch layout and minimum Git permissions after reports/learnings leave main #486 (permission-matrix docs).

[jeff-r2026]

Review: one blocking bug — stale sibling reports worktree causes silent data loss

Nice work overall — tsc/build clean, the focused suites + full test run pass, the no-copy/no-delete semantics are tested both directions, and usesReportsBranch() is applied consistently across every reader/writer. One reproducible issue should be fixed before merge, though.

Where: src/utils/reports-branch.ts ensureReportsWorktree

// Already a valid worktree — nothing to do.
if (await isGitRepo(wt)) {
  return wt;
}

Root cause: for git-kind, reports-wt is now a sibling of the clone, linked to the clone's .git. isGitRepo(wt) only checks that a .git file exists — not that the worktree's backing gitdir (<clone>/.git/worktrees/reports-wt) is still alive. When the clone is removed and re-cloned, that link dangles but isGitRepo still returns true, so this early-returns the dead worktree and skips its own remove+recreate self-heal path just below.

Trigger — a routine flow: teamai init against a different --repo URL removes and re-clones localPath (init.ts:1136 / :1149, plus the not-a-git-repo case at :1160). Because the reports worktree is deliberately a sibling (not nested), it survives the clone delete as a dangling husk.

Impact: every git op in the stale worktree throws fatal: not a git repository: .../.git/worktrees/reports-wt. commitAndPushReports catches it and returns false with a debug-only log — so member registration, pull auto-report, session save --push, and vote sync silently never reach origin, and reader commands show stale checkout data. No error surfaces to the user.

Reproduction (real git, run against this PR head — fails):

1. git-kind config → ensureReportsWorktree → write members/alice.yaml → commitAndPushReports pushes to origin/teamai-reports ✅
2. rm -rf <clone> then re-clone into the same localPath (mirrors init switching repo URL)
3. ensureReportsWorktree returns the same stale path (no rebuild); a member write → commitAndPushReports returns false, and origin/teamai-reports never gets bob.yaml. Manually rm -rf reports-wt and it heals — confirming the missing staleness check is the sole cause.

Suggested fix (verified: turns the repro green, full suite still passes):

if (await isGitRepo(wt)) {
  try {
    await createGit(wt).revparse(['--is-inside-work-tree']);
    return wt;
  } catch {
    // stale/dangling worktree link (clone was re-cloned/pruned) — recreate below.
  }
}

The existing pathExists(wt) → fse.remove(wt) + worktree prune lines right below already handle the rebuild once you fall through. A regression test around the re-clone path would be worth adding alongside git-kind-reports.test.ts.

[jimpablo]

Thanks for the review and the repro.

Fixed in d11e101: ensureReportsWorktree now probes revparse --is-inside-work-tree after isGitRepo, and falls through to the existing remove + prune + recreate path when the sibling gitdir is dangling.

Added a real-git regression in git-kind-reports.test.ts that matches your sequence (write alice → delete and re-clone → write bob); both files land on origin/teamai-reports after the rebuild.

CI on that commit is green.