# Shell split verification: the harness

Split out of `docs/shell-split-verification.md` when that file hit the 250-line
cap. That file covers **what makes a green split wrong**; this one covers the
tooling that catches it — the trace harness, its `--prefix` mode, the
regression fixtures, and two traps that have each cost a revert.

Read `docs/shell-split-verification.md` first: the hazards there decide whether
the harness can prove anything about your target at all.

### `trace_shell_split.sh` — the run, made routine

`meta/scripts/trace_shell_split.sh` (added 2026-08-16) does the stubbed run for
you. It shadows ~30 mutating binaries (`sudo`, `pacman`, `systemctl`, `adb`,
`nft`, `mount`, `mkinitcpio`, `reboot`, …) with `PATH` stubs that record each
call and exit 0, runs the script, and prints a diffable trace: exit status,
stubbed calls in order with arguments, stdout, stderr.

```bash
# baseline, in a detached worktree at the pre-split commit
meta/scripts/trace_shell_split.sh <script> --out /tmp/before.txt
# after the split
meta/scripts/trace_shell_split.sh <script> --out /tmp/after.txt
diff /tmp/before.txt /tmp/after.txt
```

It was verified against both bugs above: the `set -e` function tail shows up as
exit 1 with later calls missing, and the nameref as `circular name reference`
in stderr despite exit 0. A fixture calling `systemctl` left the unit untouched.

Four things to watch:

- **A silent stub produces an identical, meaningless trace.** This is the
  trap. `du` stubbed to print nothing makes `size=$(du -sk x)` empty, so every
  `((size > 0))` is false, every guarded branch is skipped, and two such traces
  match perfectly while exercising none of the code you changed. Give
  value-producing commands a plausible output: `--stub 'du=4096\t/some/dir'`.
  Measured on a fixture: the empty stub never reached the guarded
  `sudo rm -rf`; the valued stub printed `BRANCH TAKEN` and captured it. Ask of
  every trace: _did this actually go down the branch I touched?_
- **Stubs always exit 0.** A script that branches on a real failure status
  takes a path it would not take live. Hand-write a stub for those and say so
  in the split's evidence file.
- **Anything not stubbed runs for real.** `git`, `curl`, `wget` and `makepkg`
  are deliberately excluded from the default set, because many scripts read git
  state harmlessly and a blanket stub would change what they see. Grep the
  target for network and build verbs, then pass them:
  `--stub git,curl,makepkg`.
- **`pkill` is NOT stubbed and `--prefix` cannot contain it** — signals are not
  filesystem writes. `install_leechblock.sh:217` pkills every Chromium-family
  browser before seeding, so tracing it closes the user's browsers (measured:
  two probe runs fired it, harmless only because none was open). Grep a target
  for `pkill`/`killall`/`systemctl stop` and pass `--stub pkill` first. The
  mirror image also escapes the prefix: `setup_thorium_startup.sh` _launches_ a
  browser mid-run ("Browser launched with PID: …"). Process spawning is not a
  file write either — check `ps` after tracing anything that starts a program.

### `--prefix` — scripts whose job is placing files

Shell redirections (`cat > /etc/modprobe.d/x.conf`) are builtins, so no `PATH`
stub can intercept them. Stub `cp`/`chmod`/`ln` and the trace is nearly empty;
don't, and the run rewrites the live system. `--prefix` (added 2026-08-16)
resolves that in two layers — see `meta/scripts/lib/trace_prefix.sh`:

```bash
# $HOME-rooted destinations: no bwrap, no root needed
meta/scripts/trace_shell_split.sh install_usage_monitoring.sh --prefix /tmp/pfx-before

# hardcoded absolutes: bind each LEAF directory
meta/scripts/trace_shell_split.sh nvidia_troubleshoot.sh --prefix /tmp/pfx-nv \
  --bind-abs /etc/modprobe.d --bind-abs /etc/X11 --stub git,lsmod,nvidia-smi
```

**Every run needs its own empty prefix**, hence the `-before`/`-after` naming.
Reusing one leaves the first run's files in place, so a dropped write stops
showing up entirely — measured: the fixture pair below diffs _clean_ on a
reused prefix. The harness refuses a non-empty prefix rather than deleting it.

`HOME` **and every `XDG_*` base** are redirected. Exporting the XDG vars
explicitly is not belt-and-braces: `install_leechblock.sh` reads
`XDG_DATA_HOME` directly and then runs `rsync -a --delete` at it, so inheriting
the caller's real value would delete live data.

The trace gains a `=== files written (prefix)` section listing every file with
its size and a content hash. **That section is the point**: without it a
dropped `cat >` shows up as nothing at all — same exit status, same stubbed
calls, same stdout — and the traces match while one lost a file. Measured on
the `dropped_write_{before,after}.sh` fixture pair: with the manifest the lost
`.timer` is a one-line diff; stripped, the two traces are byte-identical.

Six traps, each of which cost a debugging round:

- **`/usr/bin` cannot be bound at all.** `--bind-abs /usr/bin` overlays the
  directory the sandbox needs to exec anything (`bwrap: execvp bash`). A script
  whose last act replaces a binary there (`install_pacman_wrapper.sh:304`,
  `ln -sf … /usr/bin/pacman`) therefore **stops at that line with exit 1** on
  EPERM. Stable across runs, so still a usable baseline — but it proves nothing
  at or after the failing line, so keep those statements in the entry script.
  The scanner misses `/usr/bin` here: it follows variables into `cat >`/`cp`,
  so a literal `ln -sf` (or `cp "$f" "$f.backup.$(date …)"`) escapes it.
- **Bind leaf directories, never `/etc`.** Binding `/etc` wholesale shadows
  `/etc/passwd` and `/etc/os-release`. Measured: `$SUDO_USER` lookup collapsed
  to `/home//pyroveil`, `/etc/profile` became "No such file or directory", and
  a nested `/etc/X11` bind was masked so its write vanished from the manifest.
  All three look like a broken split. The scanner therefore never emits a bare
  `/etc`. A `--bind-abs` naming a **file** (`/etc/profile`) is bound
  file-over-file, seeded with a copy of the original — bwrap cannot mount a
  directory over a file. The copy matters because a script that greps a config
  before appending would otherwise read an empty file and branch wrong. Note a
  file bind cannot cover new _siblings_: `cp "$f" "$f.backup.<stamp>"` still
  EPERMs, which is what truncates `nvidia_troubleshoot.sh`'s trace at step 3.
- **`require_root` truncates a stubbed run to nothing.** `lib/common.sh` does
  `exec sudo "$0" "$@"`; under the `sudo` stub that records one line and exits
  0 — a three-line trace that diffs clean against _any_ split. `--bind-abs`
  runs under `bwrap --unshare-user --uid 0`, so `$EUID` is 0 and the script
  proceeds for real. Without a bind there is no uid 0 and the truncation is
  silent, so check the trace actually entered the code you moved.
- **Unquoted heredocs bake the prefix into the artifact.**
  `install_usage_monitoring.sh:222` uses `<< SCRIPT`, so `$HOME` interpolates
  at write time. Hashes are taken over content with the prefix normalized to
  `@PREFIX@`; without that, a `mktemp -d` prefix makes every run differ.
- **A timestamp the target embeds still varies.** `nvidia_troubleshoot.sh`
  writes `$(date)` into its config, so its hash changes every run. Don't freeze
  the clock and don't fall back to comparing sizes — diff the written files'
  _contents_ with the date pattern normalized away. That still catches a
  dropped or altered line, which a size comparison does not.

Anything absolute that was _not_ bound fails read-only/EPERM rather than
succeeding, so Decision 6 holds by construction. `--prefix` **refuses to run**
when the target writes to absolute paths and no `--bind-abs` was given; the
scan follows variables (`MODPROBE_DIR="/etc/modprobe.d"` → `cat >"$CONFIG_FILE"`),
because a literal-only scan produced a silent empty manifest.

Verified untouched after the runs above: `~/.config/systemd/user`,
`~/.local/bin`, `~/.local/share`, `/etc/modprobe.d` (identical listings), and
`systemctl --user list-unit-files | grep -c usage-report` unchanged at 2.

`block_compulsive_opening.sh` remains out of scope: `install_all` copies the
running script into `/usr/local/bin`, and the handoff records that a bad split
there breaks three daily-use apps plus the pacman rewrap hook.

### Regression fixtures

`meta/scripts/fixtures/` holds runnable reproductions of every trap above, so a
harness change can be checked instead of trusted:

| Fixture                           | Reproduces                                      |
| --------------------------------- | ----------------------------------------------- |
| `set_e_function_tail.sh`          | exit 1, later calls missing from the trace      |
| `circular_nameref.sh`             | exit 0, correct stdout, warning only in stderr  |
| `valued_stub_branch.sh`           | the silent-stub trap (`--stub du` vs `du=4096`) |
| `dropped_write_{before,after}.sh` | a split that silently loses one write           |
| `xdg_destinations.sh`             | writes addressed via `XDG_*`, not `$HOME`       |

A lint fix once broke one of these invisibly: adding a trailing `echo` to
satisfy SC2034 made _that_ echo the function's last statement, so the fixture
exited 0 and reproduced nothing while still "passing". Re-run the fixtures
after touching them, and check the exit status is still the one documented.

**Run the baseline from the script's own directory** when it reads relative
paths. `fresh-install/main.sh` reads `aur_packages.txt` and
`pacman_packages.txt` from the cwd; tracing it from elsewhere dies at the first
read and looks exactly like a broken split.

Note it does not need a `main()`, which matters: none of the remaining
near-miss shell targets (`install_usage_monitoring.sh`,
`install_pacman_wrapper.sh`, `disk_cleanup_check.sh`, `fresh-install/main.sh`)
have one, so `verify_shell_split.sh` cannot check any of them.

Note the knock-on: renaming the nameref local hides the array from shellcheck
(SC2034), because it is then only ever named as a bare word. Pass the name
through a variable (`CODE_FILES_MAP="LANG_CODE_FILES"`) so the reference is
real, rather than suppressing.

### Verify through every entry point, especially symlinks

Adding a `source "$SCRIPT_DIR/lib/..."` line to `meta/lint_python.sh` broke
`./lint_python.sh` from the repo root, because that root path is a **symlink**
into `meta/` (so are `pyproject.toml`, `requirements.txt`, `run.sh`, `.fvmrc`).
`SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"` does not resolve
symlinks — `dirname` is string manipulation — so it became the repo root, where
`lib/` does not exist, and `set -euo pipefail` turned that into an instant exit.

It was invisible because every check used `bash meta/lint_python.sh`, the one
path that cannot fail. Before splitting a script, ask how else it is reachable:

```bash
find . -maxdepth 2 -type l -lname '*<script>*'   # symlinks pointing at it
grep -rn '<script>' .github/workflows/ .pre-commit-config.yaml meta/run.sh
```

The fix is the pattern already used elsewhere in this repo:

```bash
SCRIPT_DIR="$(dirname "$(readlink -f "${BASH_SOURCE[0]}")")"
```

Note the pre-split script had the same unresolved `SCRIPT_DIR`; it only fed
`PROJECT_ROOT`, so the bug was latent. Adding a `source` line is what turns a
quiet wrong value into a hard failure — which is exactly what a split does.

### Do not make a clean file shfmt-dirty

157 of 298 `linux_configuration` shell files already fail `shfmt` and there is
no `.editorconfig` or shfmt pre-commit hook, so `shell_check.sh`'s `run_linters`
exits 1 on the full tree regardless. That is not licence to add more: check
`shfmt -d <file>` before and after a split. `analyze_repo.sh` was clean and a
split made it dirty over one double blank line.

`shfmt -w` also **deletes comment banners** it considers stray — it removed a
`#====` STEP header. Re-count the banners after running it.