## What is wrong?

On this machine (`lua` 5.5.1, PUC-Rio), a Lua program that dies on an assertion exits **0** when read from stdin but **1** when read from a file or `-e`:

```text
$ lua /tmp/lua-fail.lua >/dev/null 2>&1; echo $?
1
$ lua < /tmp/lua-fail.lua >/dev/null 2>&1; echo $?
0
```

(where `/tmp/lua-fail.lua` is just `assert(false, "boom")`).

Every `test/shell.d/*-test.sh` that runs its assertions through `lua <<'LUA'` therefore **passes even when every assertion fails**: the traceback prints to stderr, `set -e` never fires (exit status is 0), and the trailing `pass` line prints. I found this while writing coverage for #14836 — the new assertions failed against the unfixed code exactly as designed, yet the file exited 0 with `ok`.

## Scope

Only two files use the pattern:

- `test/shell.d/hyprland-startup-cursor-test.sh` (worked around in #14837 by writing the program to a temp file and running `lua "$file"`)
- `test/shell.d/dictation-test.sh` (two `lua <<'LUA'` blocks — still silently non-failing)

## Suggested fix

Either run those programs from temp files (as #14837 does for its file), or add a runner guard in `base-test.sh` so a lua traceback without a matching pass fails the file. Happy to do the dictation-test follow-up once this lands.