review follow-up: close the gaps left by the phase1-6 pass

Verified the tree against docs/plans/code-review-2026-08.md and finished the
tasks that were only partly done, plus two defects the pass introduced.

- bcrypt: the setup account was still hashed at DefaultCost (10). The work
  factor now has one definition, auth.BcryptCost = 12, used everywhere.
- rate limiter: protecting blocked buckets from eviction was unbounded — an
  attacker with many source addresses could grow the map by making each one
  fail its way to blocked. Blocked buckets are now evicted past
  maxBuckets * 4, nearest expiry first, logged once per window. Tests cover
  both halves.
- postqueue: the 5s timeout left the output size unbounded; reads are now
  capped at 4 MiB and report truncation.
- spec references: the ~76 `spec N` citations CHANGELOG [0.5.0] claimed were
  gone are removed from build/, deploy/, the first two migrations and the
  release workflow, with the affected comment paragraphs re-wrapped. Comments
  only; shell/YAML/SQL syntax re-checked.
- schema-migrations.md: schema section was headed v9 with the chain at v10,
  and three links pointed at deleted plan files.
- internal/configsafe: the shared empty/forbidden-character check behind the
  OpenDKIM and Postfix config writers; each writer keeps its own forbidden set.
- dropped the parseDomainRateLimitForm/parseAppRateLimitForm pass-throughs
  left behind by the parser merge (only their own tests still called them).
- plan doc now carries a status table: what shipped, what was already in
  place, what was declined (//go:build linux for the /proc reader) and the
  open CSRF-token question.

Not done, each its own phase: typed template data structs, log/slog, DMARC
service and handler tests, error-page/corrupted-backup tests, backup re-auth.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

internal/web/auth/ratelimit.go

@@ -7,6 +7,16 @@ import (
 
 const defaultMaxBuckets = 4096
 
+// blockedBucketCeiling is how far past maxBuckets the map may grow while every
+// bucket in it is blocked. Blocked buckets are normally protected from
+// eviction (see evictNearest), but protecting them unconditionally would let
+// an attacker with many source addresses grow the map without bound — each
+// address only has to fail max times. Past the ceiling blocked buckets are
+// evicted too, nearest expiry first: bounded memory wins over holding every
+// block, and the addresses that lose their bucket have to spend max attempts
+// again to get back to blocked.
+const blockedBucketCeiling = 4
+
 // rateLimiter is a simple fixed-window per-key counter used to throttle the
 // setup and login routes (security.md). Keys are client IPs.
 type rateLimiter struct {
@@ -14,8 +24,9 @@ type rateLimiter struct {
 	window     time.Duration
 	maxBuckets int
 
-	mu      sync.Mutex
-	buckets map[string]*rlBucket
+	mu           sync.Mutex
+	buckets      map[string]*rlBucket
+	lastOverflow time.Time
 }
 
 type rlBucket struct {
@@ -68,22 +79,38 @@ func (r *rateLimiter) makeRoom(now time.Time) {
 		return
 	}
 	r.sweep(now)
+	hardCap := r.maxBuckets * blockedBucketCeiling
 	for len(r.buckets) >= r.maxBuckets {
-		if !r.evictNearest() {
+		if r.evictNearest(false) {
+			continue
+		}
+		// Nothing evictable is left: every remaining bucket is blocked. Hold
+		// them until the map reaches the hard ceiling, then start evicting
+		// blocked buckets as well so memory stays bounded.
+		if len(r.buckets) < hardCap {
 			break
 		}
+		if !r.evictNearest(true) {
+			break
+		}
+		if now.Sub(r.lastOverflow) >= r.window {
+			r.lastOverflow = now
+			logf("panel: login rate limiter holds %d blocked buckets (cap %d) — evicting blocked entries; "+
+				"this means many distinct client addresses are failing authentication", len(r.buckets), r.maxBuckets)
+		}
 	}
 }
 
-// evictNearest removes the bucket whose window ends soonest, skipping buckets
-// that are currently blocked (count >= max) so an attacker cannot flush active
-// rate-limit state by filling the map with new keys.
-func (r *rateLimiter) evictNearest() bool {
+// evictNearest removes the bucket whose window ends soonest and reports whether
+// it removed one. Blocked buckets (count >= max) are skipped unless evictBlocked
+// is set, so an attacker cannot flush active rate-limit state by filling the map
+// with new keys.
+func (r *rateLimiter) evictNearest(evictBlocked bool) bool {
 	var bestKey string
 	var bestEnds time.Time
 	first := true
 	for k, b := range r.buckets {
-		if b.count >= r.max {
+		if !evictBlocked && b.count >= r.max {
 			continue
 		}
 		if first || b.windowEnds.Before(bestEnds) {