quarantine_finding re-classifies without the graph, so it accepts a file the report calls risky

[Muawiya-contact]

quarantine_finding is documented as the safety net that does not trust
the frontend:

Nothing here trusts a verdict the frontend claims — quarantine_finding
re-classifies, and the restore/purge commands only ever touch paths the
manifest in that directory says Diskern put there itself.

The re-check is a rules lookup and nothing else:

let (_, verdict, _) = RulesDb::embedded().classify(&path);
if matches!(verdict, Verdict::Protected | Verdict::Risky) { ... }

That is the rule's base verdict. The verdict the report shows is the
base verdict after risk::downgrade has applied the impact graph, and the
whole point of #48 was that those two differ. So the command disagrees
with the report exactly where the graph is doing its job.

Scanning a live npm project — package.json beside its node_modules —
and then running what the command runs:

report verdict      : Risky
report reasons      : ["matched rule node-modules: ...", "referenced by 1 project", ...]
report reclaimable  : 0
command re-check    : Review
command would refuse: false
actions::quarantine : MOVED to .../quarantine/1788716271__tmp_..._node_modules_react_index.js

The file moved. actions::quarantine refuses Risky, but it is handed
Review, so its own guard never fires either — the caller decides the
verdict, which is the thing this design says never happens.

Not reachable from the shipped UI: FindingRow renders no button for a
risky row, so the command is never invoked with one today. That makes
this a layer that is documented, believed, and not there — the layer that
is supposed to catch a UI bug, a stale report row, or a second frontend.

The fix has to get graph evidence into the command. The cheap version is
to keep the referenced-project count from the scan that produced the
report and consult it here; the honest version is for the command to take
the finding's verdict from the report the backend itself produced rather
than re-deriving a weaker one. Either way docs/ARCHITECTURE.md and the
app README describe the result.

app/src-tauri/src/commands.rs:201 · crates/diskern-core/src/risk.rs:45