[BUG] Agent tool `isolation: "worktree"` creates worktree from stale session-start snapshot (old HEAD + old dirty files), not current HEAD

## Environment
- Claude Code v2.1.220, macOS (darwin 24.5.0), CLI
- Model: claude-fable-5 (coordinator) spawning subagents via the Agent tool

## Bug
When a subagent is launched with `isolation: "worktree"` late in a long session (session had been running for hours and went through `/compact` at least once), the auto-created worktree is **not** based on the current HEAD of the repo. Instead it reproduces a **snapshot of the repo state from the moment the session started**: the old HEAD *plus the exact set of dirty/untracked files that existed at session start*.

## Evidence (one session, 3 occurrences on 2026-07-26)
Main tree at delegation time: HEAD `70f5ead7`, `tests/aaf` = 31 files.
Session start state (from the session's initial gitStatus snapshot): HEAD `2208e60d`, 5 modified + 2 untracked files.

All three auto-created worktrees (`.claude/worktrees/agent-*`) contained:
- HEAD `2208e60d` (= session-start HEAD, dozens of commits behind current)
- `tests/aaf` = 19 files (session-start count)
- `git status` showing **exactly the same 5 modified + 2 untracked files** as the session-start snapshot — files that had long been committed in the main tree

The dirty-file match makes it clear this is a cached session-start snapshot being replayed, not a race or a wrong-branch checkout.

## Impact
In an earlier occurrence (same day, before we added guards) a subagent received the stale base and, in good faith, **re-implemented 15 commits of already-finished work**, which then had to be detected and discarded during review. Silent data-duplication risk for any workflow that delegates writing work to isolated subagents in long sessions.

## Workarounds we use now
- Every writing-subagent brief starts with a base assert (`git log -1` == expected SHA + a file-count signature; mismatch → STOP). This caught occurrences 2 and 3 before any edit.
- Coordinator creates the worktree manually (`git worktree add --detach /tmp/wt-task `) and passes the path in the brief instead of using `isolation: "worktree"`.

## Expected behavior
`isolation: "worktree"` should create the worktree from the repo's **current** HEAD (and arguably current dirty state), or at minimum refresh its snapshot instead of reusing the session-start one after `/compact`/long uptime.

Reproduce the issue by running a long CLI session through /compact, making commits after session start, and launching an Agent with isolation: "worktree". Compare the auto-created worktree's git log and git status with the current repository and the session-start state. Done means the worktree reflects the current HEAD rather than replaying stale commits or dirty files.

Written by the indexing model from the issue text.