fix: close the gaps a security review found in the default install

Reported against 2.2.1 by someone reading the source. Every finding held.

**Sign-in had no ceiling.** Failures were logged with the caller's address
and nothing more. Two counters now — one per address, one per account —
because the two attacks look different: one source working through many
accounts is caught by the first, many sources working on one account by the
second, and behind a proxy only the second still means anything.

The count is kept in the process that serves the request. PLANKA needs no
Redis and the stock deployment is one container; run several and each keeps
its own count, which multiplies the ceiling by their number. That trade is
written where the limits are configured.

**The second factor could be guessed at leisure.** Six digits, and a
pending token that stayed valid for its full ten minutes however many codes
were wrong. Wrong codes are now counted on the session row — in the
database, so the count survives a restart and holds across every process —
and when the budget is spent the session is destroyed. After that even the
right code is refused and the login starts over from the password.

**Avatars, background images and favicons** checked the token's signature
and nothing else, so a revoked session, a deactivated account or a changed
password all kept working there for as long as the signature lasted, which
is a year by default. The five checks the API makes now live in one helper
that both use, rather than the shortened copy that had drifted from it.

**A link attachment's favicon** was fetched from wherever the URL pointed.
Storing a link is harmless — it is a string the user typed — but fetching
its icon is a request the server makes to an address the user chose, and
whether an icon came back reported on what is reachable from inside the
network. Server-side fetches now refuse private, loopback and link-local
addresses, `169.254.169.254` among them. The attachment is still created:
linking to an internal wiki is a legitimate thing to do, and it was the
server's own request that had to stop.

**The signing key.** Our own compose file ships `notsecretkey`, and it is
printed in the documentation — so on any instance that copied it, anyone can
sign a token for any account. PLANKA now says so on every start, and keeps
saying it, along with a key that is missing or shorter than 32 characters.
The placeholder carries the warning inline, where it is copied from.

**The backup script** wrote password hashes, live sessions, TOTP secrets and
SMTP credentials to an unencrypted archive. `BACKUP_PASSPHRASE` now encrypts
it, and without one the script says what it just put on disk. It also says
what it is — an example for the stock compose stack, not a backup concept —
and names the window between the database dump and the file copy, which no
ordering closes.

server/db/migrations/20260827130000_add_totp_attempt_limit.js

@@ -0,0 +1,25 @@
+/*!
+ * Copyright (c) 2026 PLANKA Software GmbH
+ * Licensed under the Fair Use License: https://github.com/plankanban/planka/blob/master/LICENSE.md
+ */
+
+// Counts wrong second-factor codes against the one half-finished login they
+// were entered into. A pending token stays valid for ten minutes, and six
+// digits fall to patience — without a counter those ten minutes are ten
+// minutes of free guessing. When the count runs out the session is destroyed
+// and the login starts over from the password.
+//
+// It lives on the row rather than in memory so it survives a restart and holds
+// across every process serving the instance.
+
+module.exports.up = (knex) =>
+  knex.schema.alterTable('session', (table) => {
+    // The default belongs on the column: sessions are also created by raw
+    // inserts that never pass through the model, and those would write NULL.
+    table.integer('pending_token_attempts').notNullable().defaultTo(0);
+  });
+
+module.exports.down = (knex) =>
+  knex.schema.alterTable('session', (table) => {
+    table.dropColumn('pending_token_attempts');
+  });