Fix what the review of the held-branch work found

[skunkworker]

A /code-review of #7 found 15 issues (items 19 to 33 in docs/test-coverage-gaps.md). Each fix has a test that fails with the fix taken out. The git facts behind 19, 23 and 24 were reproduced in throwaway repos first.

• r and pull requests: r now refuses an upstream that is not a branch, such as refs/pull/1/head. Before, it pushed the delete of a teammate's branch with the tracking name.
• Protection: pruner.protect globs match both the name a row shows and the name on the remote. Each remote's trunk, from <remote>/HEAD, is protected, not only the default remote's.
• Confirm screen: it names the remote branch that a renamed row deletes. A refusal from r no longer stays on screen.
• Held branches: linked worktrees are listed, not globbed, so [, * or ? in the repo path no longer hides them. BISECT_START counts only while BISECT_LOG exists, as in git. The check no longer needs git 2.31.
• c on a remote row: a remote named team/a gives x from team/a/x, not a/x.
• Partial clones: git switch runs with no terminal, so ssh cannot prompt over the screen. Commit counts keep the local cap of 8, not the push cap of 3.
• No freeze: commit counts run off the update loop: every reload's merge queries, and the count d starts. d opens the confirmation screen when they land.
• Failed counts: a riskFailed flag, not -1 in the count field. Every "commits at risk" check uses atRisk().

One new item, 34, is open: the force retry still runs deletes and pushes inside Update.

Tests: 3 new, and new cases in existing tests. 108 in total, passing under -race.

🤖 Generated with Claude Code