harden(auth): LOW + Nit batch — bootstrap audit, crypto/rand, XFF trust, CSRF check, protocol-prefix unify (Batch 1)

Audit 2026-05-10 — close 8 LOWs + 2 Nits in-bundle. Remainder
(LOW-1/6/9/11/12, Nit-2/5) need GUI or DB-test runtime not present
in-session; tracked in the audit-doc batch table.

LOW-2: bootstrap.ValidateAndMint now emits 'bootstrap.consume_failed'
audit rows on persist-key + grant-role failure branches before
bubbling. Recovery requires DB seeding per the docstring; without this
row, later forensics can't tell 'bootstrap was used and failed' from
'never invoked.'

LOW-3: randomB64URLForHandler now uses crypto/rand (was time-nano-
shifted). Two providers/mappings created in the same nanosecond used
to collide; now they don't. Time-nano fallback retained for the
unlikely crypto/rand-broken path.

LOW-4: breakglass.verifyDummy uses s.readRand(salt) for the dummy
Argon2id verify. Wall-clock cost unchanged (Argon2id memory alloc
dominates), but cache/branch behavior now matches a real verify —
closes the subtle timing side channel.

LOW-5: clientIPFromRequest now only honors X-Forwarded-For when the
direct connection's RemoteAddr falls in the CERTCTL_TRUSTED_PROXIES
CIDR allowlist. Default-deny: empty list means XFF is ignored.
SetTrustedProxies wired in cmd/server/main.go from cfg.Auth.TrustedProxies.

LOW-7: internal/auth/protocol_endpoints.go::ProtocolEndpointPrefixes
now carries /scep-mtls + /.well-known/est-mtls (previously only in
router.AuthExemptDispatchPrefixes; the two lists had drifted). The
canonical-prefix coverage test in Phase 12 still pins the set.

LOW-8: docs/operator/rbac.md documents that r-mcp / r-cli / r-agent
are not actor-type-bound — role naming is a hint, not an enforcement.
Operators wanting hard binding must apply periodic audit queries.
Native binding is on the v2 roadmap.

LOW-10: Session.Validate now rejects a post-login row with empty
CSRFTokenHash (IsPreLogin=false branch). validSession test fixture
updated with a valid 64-hex CSRF hash.

Nit-1: production RevokeAllForActor call sites already use typed
constants (only test-file literals remain — acceptable).

Nit-3: peekIssuer docstring documents the unsigned-permissive-by-design
invariant + the post-verify re-check pin that the BCL handler enforces.
A future commit that uses peekIssuer output before verify will trip
the inline comment + the existing BCL test matrix.

Status table updated in cowork/auth-bundles-audit-2026-05-10.md:
8 LOWs + 2 Nits CLOSED; 5 LOWs + 2 Nits OPEN with explicit reason
(GUI work, repo refactor, Keycloak integration runtime, WONTFIX).

Refs: cowork/auth-bundles-audit-2026-05-10.md LOW-2/3/4/5/7/8/10
      cowork/auth-bundles-audit-2026-05-10.md Nit-1/3

internal/auth/breakglass/service.go

@@ -532,7 +532,14 @@ func verifyPassword(plaintext, encoded string) (bool, error) {
 // paths take statistically indistinguishable time. The result is
 // discarded.
 func (s *Service) verifyDummy(plaintext string) bool {
-	dummySalt := make([]byte, argon2SaltSize) // all-zeros — fine for timing parity
+	// Audit 2026-05-10 LOW-4 closure — was an all-zeros salt; while the
+	// wall-clock cost matched a real verify (the 64MiB Argon2id
+	// allocation dominates), cache/branch behavior differed enough to
+	// give a subtle timing side channel. Use crypto/rand for the dummy
+	// salt too. If RNG fails, fall back to all-zeros (the timing parity
+	// is still preserved by the dominant Argon2id memory cost).
+	dummySalt := make([]byte, argon2SaltSize)
+	_, _ = s.readRand(dummySalt)
 	_ = argon2.IDKey([]byte(plaintext), dummySalt,
 		uint32(argon2Iterations), uint32(argon2Memory),
 		uint8(argon2Parallelism), uint32(argon2OutputSize))