findings of the security review of stages 2-6

The review of 4433688..0808779 (read-only, a separate agent) found one
medium and some lows; all fixed here.

Medium, proven by the reviewer: A1's userKey put the typed username into
the limiter map whole, and the public forms had no body bound beyond
net/http's 10 MiB for a parsed form - each refused request kept up to
10 MiB in memory for fifteen minutes. userKey now holds a 96-bit hash of
the folded name, and sign-in, reset, a reset link's page and setup read at
most 64 KiB (limitForm, http.MaxBytesReader).
TestPublicFormsAreBoundedAndKeysShort.

Lows:
- Refund skipped the request that filled a bucket (count < max), so a
  right password at the ceiling left the shared address blocked; Refund
  now always undoes the Allow it follows.
- The reset form did not give the address its attempt back when the
  name's bucket refused (sign-in did); it does.
- A3's day of five links is spent only by a link that goes out, and the
  price of the pause - anyone knowing a username can pause that owner's
  reset for a day - is recorded in security.md.
- longHeaderLine counted a trailing CR on a header's first line.
- test.yml declares permissions: contents: read.

Not changed, as the review's nits: the store matches names exactly while
the limiter folds them (stricter, said in the comment now); a session from
a longer window is honoured until its next Touch (the next request).

go test ./internal/web/auth ./internal/milter.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

.github/workflows/test.yml

@@ -5,6 +5,11 @@ on:
     branches: [main]
   pull_request:
 
+# The job reads the repository and nothing else (security audit 2026-10-17,
+# review of stages 2-6).
+permissions:
+  contents: read
+
 jobs:
   test:
     runs-on: ubuntu-latest