"Fix it for me" audit

The desktop app may show a "Fix it for me" button only for a module that passes every one of these tests:

1. The change is reversible, by the person, with steps we can write in plain words.
2. It is low impact: nothing else stops working, no data is lost.
3. It does exactly what one sentence says ("Turns on your Mac's firewall").
4. It needs no administrator password.
5. It touches only the setting or file the finding is about.
6. It cannot be wrongly aimed (no ids or paths that go stale between the scan and the fix).

Any doubt is a NO. A wrong automatic fix on a frightened person's computer is worse than none; the "How to fix this" steps remain for everything declined here.

Only modules listed with verdict OFFER may appear in rescue/app_fixes.json. tests/test_app_fixes.py fails if a module with a state-changing fix() is missing from this table, if the table lists a module whose fix() is guidance only, or if the allowlist names a module that is not marked OFFER here.

Method: every modules/*/*/__init__.py was parsed, and fix() plus every helper reachable from it was searched for state-changing calls (subprocess, kill, unlink, rmtree, write, move, mkdir, chmod, ActionKind.MUTATION). The other modules' fix() only return guidance text (some contain sudo ... commands in strings for the person to read; none run them).

Result: 11 modules have a state-changing fix(). 0 are offerable today. One engine-level action, disable_launch_agent, is offered instead (last row). The feature, the allowlist and the gate are in place and tested, so a module can be added later once its fix() is narrowed and re-audited.

Module	What fix() really does	Verdict	Reason
startup_auditor	Runs launchctl unload -w ~/Library/LaunchAgents/<label>.plist for EVERY matched finding in one go. The list it matches mixes real scareware (MacKeeper, Genieo, Pirrit) with updaters (Google Keystone, Microsoft AutoUpdate, Java) and a working antivirus (Avast).	NO	Fixes all findings at once, so one card's button would change other items. Several matches are security updaters or an antivirus whose removal leaves the person less safe. Unloading adware does not remove it, so "fixed" would overpromise. Candidate for a future per-item version limited to the scareware/adware entries.
firewall_audit	Runs socketfilterfw --setglobalstate on and --setstealthmode on.	NO	Needs administrator rights (its own error text says so), so it would fail for a normal user. Stealth mode can also break device discovery.
win_firewall	Runs netsh advfirewall set <profile> state on.	NO	Needs Administrator. Turning a profile on can cut off a network the person relies on (remote desktop, shared printers).
win_defender	Runs Set-MpPreference -DisableRealtimeMonitoring $false or Update-MpSignature.	NO	Needs Administrator. Real-time protection is often off on purpose when another antivirus is installed; forcing it on creates conflicts. Tamper Protection can silently refuse.
win_antivirus_status	Same two PowerShell commands as win_defender.	NO	Same reasons as win_defender.
win_updates	Runs Start-Service -Name wuauserv.	NO	Needs Administrator. The service is often stopped deliberately by policy or a management tool.
ai_worm_filesystem	Moves files flagged "high confidence" into ~/.rescue_quarantine.	NO	Moves files, and the person may be dealing with an active compromise or want to keep evidence. Detection confidence is a heuristic. Needs a human decision per file.
ai_worm_git_ssh	Unsets global git config keys, edits .npmrc, edits MCP server config files, moves repo hook files.	NO	Edits configuration files a developer may rely on, with no undo copy. Wrong edits can break their tools or destroy evidence.
ai_worm_lateral	Deletes files flagged as malicious artifacts.	NO	Permanent deletion with no undo. Rotating exposed credentials must come first and cannot be automated.
ai_worm_network	Sends SIGTERM to a process id taken from the scan, and rewrites the hosts file.	NO	The process id can be reused by a different program between the scan and the fix, so it could close the wrong program. Rewriting /etc/hosts needs admin.
ai_worm_persistence	Unloads and deletes launch agents, systemd units and scheduled tasks, and rewrites shell profiles, settings hooks.	NO	Deletes persistence entries and rewrites files with no undo copy, and the order matters for dead-man-switch malware. Needs human guidance.
disable_launch_agent	Engine action, not a module fix(). Stops one background job (launchctl bootout gui/<uid>/<label>) and moves its plist from the person's own ~/Library/LaunchAgents to ~/Library/Application Support/Device Rescue/Quarantine, with a JSON manifest (original path, label, time, sha256). Undo moves it back and runs launchctl bootstrap.	OFFER	1. Reversible in one click (Undo), or by moving the file back. 2. One job stops; nothing is deleted. 3. One sentence says it. 4. No password: it is the person's own user domain and their own file. 5. Touches only the one named plist. 6. The path is re-validated at run time (regular file, not a symlink, owned by the user, directly inside their LaunchAgents, parseable, Label equals file name), and never /Library or /System; the scan is not trusted. Offered only for warning/critical findings from the persistence audits. Caveat: whoever set the job up may notice it stopped, so the app adds a warning for the "someone is watching me" situation.

Adding a module later

1. Narrow fix() so it acts on exactly one named setting and cannot touch more.
2. Add its row here with verdict OFFER and the reason it passes all six tests.
3. Add an entry to rescue/app_fixes.json (will_change, how_to_undo).
4. pytest tests/test_app_fixes.py must pass.