Extend guard_master.py's worktree gate to Edit/Write/NotebookEdit

[WilfordGrimley]

Description

guard_master.py's PreToolUse hook only ever matched Bash commands. Four
independent sessions on 2026-07-23 and 2026-07-24 accidentally edited the
shared main checkout via absolute-path Read/Edit/Write calls run from
a worker worktree — the failure mode docs/lessons.md's "Absolute paths to
the repo root silently target the wrong checkout in a worktree session"
entry already documents, but via a file tool instead of git/gh in
Bash, so the existing rules never saw it. All four were self-caught
before landing anything, but the near-misses stacked up fast enough to
close the gap at the tool layer instead of relying on every session to
catch its own mistake.

This PR:

• Extends .claude/settings.json's PreToolUse matcher to also fire on
  Edit, Write, and NotebookEdit (was Bash only).
• Adds check_worktree_write_guard() to guard_master.py: when the
  session's cwd is under a worker worktree (.claude/worktrees/<name>/...)
  and the tool's target path (file_path for Edit/Write, notebook_path
  for NotebookEdit) resolves inside the main checkout root but outside
  .claude/worktrees/, the call is blocked with a message mirroring the
  existing Bash-rule wording.
• Preserves the documented WORKERS.md/journal/ exception (per
  docs/lessons.md and CLAUDE.local.md's multi-worker protocol) —
  those are gitignored files that intentionally live in the main
  checkout, so writes there from a worktree session stay allowed.
• Leaves Read entirely ungated — blocking reads would break legitimate
  cross-referencing against the main checkout's on-disk state, and a read
  can't silently land content on the wrong branch the way a write can.
• Does NOT block: paths inside the session's own worktree, paths inside
  any other worktree, or paths outside the repo entirely (/tmp, the
  orchestration repo, the memory dir).

Checklist

• I have installed pre-commit and installed the hooks with pre-commit install before creating any commits.
• I have updated any related tests for code I modified or added new tests where appropriate — extended the existing .claude/hooks/test_guard_master.py suite (there was already a test precedent for this hook) with 13 new end-to-end cases and 9 new unit cases against the new helper functions; no new --self-test mode was needed since the precedent already existed. This suite has no CI wiring (same as before this change) — it's a local-only dev-loop check, run manually.
• I have manually tested my changes as follows:
  • python3 .claude/hooks/test_guard_master.py — all 58 cases pass (36 pre-existing + 22 new).
  • python3 -m py_compile .claude/hooks/guard_master.py .claude/hooks/test_guard_master.py
  • black --check and the repo's pre-commit hooks (ruff, isort, black, mypy, prettier) all pass.
  • python3 .github/scripts/docs_lint.py — clean.
• I have updated any relevant documentation or created new documentation where appropriate — extended the existing .claude/hooks/guard_master.py guard_master.py entry in docs/troubleshooting.md, and added a "closed at the tool layer" note to the existing worktree-path-trap entry in docs/lessons.md.