fix(security): normalize the program path in the destructive-command screen

[D41910]

Why

Follow-up to #195 (which closed the whitespace / flag-order bypass in issue #128), and of the same kind as #176. screen_command matches the program name against _classify using tokens[0].lower(), so it classifies on the literal token rather than on the program it names. Every other screen in the module resolves the program through _program() first:

core/harness/command_guard.py:467    if _program(tokens[0]) not in _FETCHERS:      # screen_egress
core/harness/command_guard.py:600    program = _program(tokens[0])                 # screen_install
core/harness/command_guard.py:609    program = _program(args[1])                   # screen_install
core/harness/command_guard.py:158    reason = _classify(tokens[0].lower(), ...)    # screen_command

_program() is the module's own definition of "what a program is called" — it drops the directory and any Windows extension:

def _program(token: str) -> str:
    """The bare program name, without directory or Windows extension."""
    name = token.replace("\\", "/").rsplit("/", 1)[-1].lower()
    for suffix in (".exe", ".cmd", ".bat", ".ps1"):
        ...

Because screen_command skipped it, the identical command was caught when typed bare and passed when typed with its path:

command	screen_command	reality
rm -rf /	recursive force remove	destructive
/bin/rm -rf /	None ❌	same binary, same flags
sudo reboot	privilege escalation	destructive
/usr/bin/sudo reboot	None ❌	same
dd if=/dev/zero of=/dev/sda	raw disk write	destructive
/bin/dd if=/dev/zero of=/dev/sda	None ❌	same
mkfs.ext4 /dev/sda1	raw disk write	destructive
/sbin/mkfs.ext4 /dev/sda1	None ❌	same
chmod 777 /etc	world-writable permissions	destructive
/bin/chmod 777 /etc	None ❌	same
ls | /usr/bin/rm -rf /	None ❌	still split segment-by-segment

This is the exact gap the module docstring says it exists to close: "That closes the whitespace / flag-order / flag-spelling gaps without pretending to be exhaustive." Path qualification is one more spelling gap, and the helper for it was already in the file.

Change

One line — resolve the program through _program() before classifying, matching the two sibling screens:

-        reason = _classify(tokens[0].lower(), tokens[1:])
+        reason = _classify(_program(tokens[0]), tokens[1:])

_program() already lowercases, so the existing RM -rf case is unaffected.

Still defense-in-depth, not the boundary

Unchanged from #195, and deliberately so: the sandbox in core.harness.sandbox remains the enforcement boundary, screen_command still never raises, and an untokenisable command is still passed through. This only makes the shallow screen catch what it already claims to. I did not touch the sandbox.

Tests

Added to tests/test_command_guard.py (no new file):

• test_blocks_path_qualified_destructive_commands — 13 cases: absolute (/bin/rm, /usr/bin/sudo, /sbin/mkfs.ext4), relative (./rm --recursive --force), Windows extension stripping (/usr/bin/rm.exe, C:/Windows/System32/shutdown.exe), and path-qualified stages inside a sequence and a pipeline.
• test_allows_path_qualified_benign_commands — 4 cases, so normalising the program does not turn every absolute path into a hit (/usr/bin/rm notes.txt stays allowed, as before).

Test-validity self-check. With the one-line fix reverted and the new tests kept, all 13 destructive cases fail on the unmodified main; restoring the fix makes all 53 pass:

FAILED test_blocks_path_qualified_destructive_commands[/bin/rm -rf /]
FAILED ... [/usr/bin/rm -rf /tmp/x]        ... 13 failed, 40 passed   (fix reverted)
53 passed                                                              (fix applied)

Differential fuzz. 28,200 generated commands (programs × flag spellings × paths × shell operators, plus 4,000 random multi-segment pipelines) run through the pre-fix and post-fix implementations: 635 newly blocked, 0 newly allowed. Every one of the 635 contains a segment whose normalised program is genuinely destructive (checked programmatically against the classified set) — no unintended behaviour change, and spot-checked that touch rm-rf-notes.txt, grep -r foo ./src, cat chmod.txt, git commit -m 'drop sudo' still pass.

Full suite: 2279 passed, 11 skipped in 2m26s, on the fix applied — same as baseline. ruff check and ruff format --check clean on both touched files.

Note on Windows backslash paths

C:\Windows\System32\shutdown.exe still passes, for an unrelated reason: shlex.split(..., posix=True) consumes the backslashes as escapes, so the token arrives as C:WindowsSystem32shutdown.exe before _program() ever sees it. That is a tokenizer-level behaviour rather than a program-name mismatch, and fixing it means deciding how this POSIX-oriented tokenizer should treat Windows paths — a larger design question than this PR's one-line scope, so I left it alone rather than fold it in silently.