fix(labs): strict numeric parsing, atomic confirm guard, test integrity

- normalize.ts num(): parseFloat truncated at the first comma, so "1,200"
  silently became 1 (1000x error) and "5,5" became 5. Now strictly matches
  either US thousands-grouping or a plain number spanning the whole string;
  anything else (incl. ambiguous "5,5") returns null instead of a wrong value.
- labs.ts confirm handler: the pending-status check ran before the request
  body was read, so two concurrent confirms could both pass it and
  double-insert. Added a guarded UPDATE ... WHERE status = 'pending' as the
  first statement inside the existing synchronous transaction; zero rows
  affected throws and the route returns 409, with the fast-path check kept
  for the common case.
- Added missing `await` on two rejects.toThrow assertions (llm.test.ts,
  extract.test.ts) that were previously resolving before the assertion
  settled.
- Bumped the 11th-failed-login rate-limit test to a 30s timeout — 10
  sequential argon2id verifies can exceed bun:test's 5s default under load.

server/test/llm.test.ts

@@ -46,6 +46,6 @@ describe("chatJSON", () => {
 
   test("non-2xx throws llm_error", async () => {
     const mock = (async () => new Response("nope", { status: 401 })) as unknown as typeof fetch;
-    expect(chatJSON(deps(mock), { system: "s", user: "u" })).rejects.toThrow(/llm_error/);
+    await expect(chatJSON(deps(mock), { system: "s", user: "u" })).rejects.toThrow(/llm_error/);
   });
 });