feat: run shell commands through the detected user shell

[wowi42]

What

lecode executed shell commands through hardcoded shells — the agent's bash tool used /bin/sh -c, and the interactive !cmd / !!cmd shell-out used bash -c — so commands ran without the user's shell environment. Both now detect and use the user's shell.

How

• New extras/shell.py: user_shell() resolves $SHELL via shutil.which (accepts absolute paths and bare names), falling back to /bin/sh when $SHELL is unset, empty, missing, or not executable — i.e. the previous behavior is the fallback
• agent/tools/bash.py: BashTool resolves the shell at construction (shell= param for injection) and _run_shell executes it; background shell tasks inherit the shell since they run through _run_shell
• The tool description names the resolved shell (Run a shell command (/bin/zsh -c)…) so the model can adapt its syntax (e.g. fish vs POSIX)
• tui/app.py: !cmd / !!cmd run through user_shell() -c

Deliberately unchanged: hooks keep their documented /bin/sh -c contract (docs/hooks.md), and loop mode's configured command still runs via bash.

Tests

• New tests/test_shell.py: unset/empty/missing/non-executable $SHELL → /bin/sh; absolute path used; bare name resolved on PATH
• test_tool_bash.py: the tool runs commands through $SHELL -c (pinned with /bin/echo, which prints its argv) and the description names the detected shell
• test_tui_app.py: !cmd uses the detected shell
• Verified end-to-end with SHELL=/bin/zsh: description reads /bin/zsh -c, $0 and $ZSH_VERSION confirm zsh executes the command