fix: resolve bash executable path for Windows test subprocess calls

[Dakshkweera]

Closes #665

Problem

subprocess.run(["bash", ...]) only resolves bash on Windows when the calling shell already has Git's usr/bin on PATH (i.e. inside a Git Bash session itself). A plain PowerShell/cmd session doesn't have it on PATH at all, so the bare string raised FileNotFoundError: [WinError 2].

Fix

Added a small _resolve_bash() helper (duplicated in each affected test file, since this pytest config's import mode doesn't reliably support cross-file imports from conftest.py) that:

1. Tries shutil.which("bash") first (works on Linux/macOS/Git Bash where it's on PATH).
2. Falls back to deriving Git's install root from the git executable (reliably on PATH via shutil.which("git")) and checks bin/bash.exe / usr/bin/bash.exe relative to it.
3. Falls back to the bare "bash" string if neither resolves, preserving current behavior everywhere else.

Verified

Tested on Windows 11 in a plain PowerShell session (not Git Bash) — the actual repro conditions from #665. Diffed the full test suite's failure list before/after:

• 4 tests fixed, 0 regressions (test_claude_hook_memsearch_disable_exits_before_writing_memory, test_claude_session_start_does_not_create_journal, test_claude_session_start_falls_back_for_older_cli, test_session_start_shows_skill_candidate_hint)

Note on scope

Some other tests in these same files (test_codex_parse_rollout.py, test_codex_stop_hook_utf8.py) still fail on Windows for an unrelated reason: the .sh scripts they exercise call python3 internally, and Windows only ships python.exe (no python3 alias), so it falls through to the Microsoft Store app-execution-alias stub. That's a separate, pre-existing issue affecting ~33 files across the plugins — out of scope for this PR, happy to file it separately if useful.