v0.6.2: Mythos Validation — full-codebase polish, hot-path optimizations, dhcproto 0.15

Codebase-wide review pass: finish or remove every loose end, take the
safe performance wins on the serving hot paths, and refresh the
dependency tree for reliability. No behavior changes for working
clients; legacy clients get clearer protocol errors.

Finalize / cleanup:
- Remove mac_allowlist/subnet_allowlist config fields — parsed but never
  enforced since introduction; the operator wants line-of-sight serving,
  so the honest fix is deletion, not wiring.
- Remove dead ClientRegistry API (get, set_selected_target,
  always-None selected_target field, never-emitted DhcpRequest/
  HttpIsoAsset events).
- TFTP: reject WRQ with ERR_ILLEGAL_OP and non-octet modes with a clear
  error instead of silent timeouts (legacy-client friendliness); fold
  plan_window into cfg(test); drop the unused-constant keep-alive hack.
- rustfmt sweep over the six files with accumulated drift.

Hot-path optimizations (all behavior-preserving):
- Serve embedded iPXE binaries zero-copy (Cow over rodata) on both TFTP
  and HTTP — was a ~1 MiB heap copy per boot file request.
- Cache the composited PXE boot-menu background PNG keyed on the
  branding logo revision — was ~50-200 ms of image work per booting
  client; now one compose per logo change.
- Run bcrypt verify/hash on the blocking pool (boot password gate,
  login, setup, credential rotation) so CPU-heavy auth can't stall the
  workers streaming ISO ranges to imaging machines.
- iso_raw: reuse the already-cloned IsoMeta for path resolution instead
  of a second registry lock + deep clone per range request.
- DriverEscalation: amortize the TTL sweep (1-min interval + inline
  staleness check) instead of an O(map) retain per DHCP packet.
- format_mac: one allocation instead of four per datagram.
- Introspection haystack sized to min(scan cap, file size) — was
  guaranteed a 32 MiB realloc on every large-ISO probe.

Robustness:
- parse_range: malformed Range headers are now ignored per RFC 7233
  (200 + full body) instead of answered with a bogus 206.

Dependencies:
- dhcproto 0.12 -> 0.15: drops the deprecated/unmaintained
  trust-dns-proto from the tree (hickory-proto), three releases of DHCP
  option coverage. Compiles + passes the full suite unchanged.
- socket2 0.6 (dedupes tree), bcrypt 0.19, tower-http 0.6.11 (sheds
  iri-string), tokio 1.52.3 / hyper 1.10 lockfile refresh; dead nom
  workspace entry removed; requested versions synced to shipped reality.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

crates/http-api/src/auth.rs

@@ -154,10 +154,15 @@ pub fn session_cookie(session: &str) -> String {
 
 fn parse_cookie(headers: &axum::http::HeaderMap) -> Option<String> {
     // `Cookie: a=b; c=d` parsing — small enough not to drag in a crate.
+    // Two-step strip (name, then '=') keeps this allocation-free per
+    // candidate and can't match a longer cookie name sharing the prefix.
     let raw = headers.get(header::COOKIE)?.to_str().ok()?;
     for part in raw.split(';') {
         let part = part.trim();
-        if let Some(v) = part.strip_prefix(&format!("{SESSION_COOKIE}=")) {
+        if let Some(v) = part
+            .strip_prefix(SESSION_COOKIE)
+            .and_then(|rest| rest.strip_prefix('='))
+        {
             return Some(v.to_string());
         }
     }
@@ -246,7 +251,14 @@ pub async fn api_setup(State(state): State<AppState>, Json(body): Json<SetupBody
         )
             .into_response();
     }
-    match state.admin.bootstrap(&body.username, &body.password) {
+    // bcrypt hashing is ~100-200 ms of pure CPU (and `bootstrap` also
+    // persists to disk synchronously) — keep it off the async workers.
+    let admin = state.admin.clone();
+    let result =
+        tokio::task::spawn_blocking(move || admin.bootstrap(&body.username, &body.password))
+            .await
+            .unwrap_or_else(|e| Err(openpxe_core::Error::Other(e.into())));
+    match result {
         Ok(pub_) => {
             let session = state.sessions.create(&pub_.username);
             login_response(StatusCode::CREATED, &pub_, &session)
@@ -271,8 +283,13 @@ pub struct LoginBody {
 pub async fn api_login(State(state): State<AppState>, Json(body): Json<LoginBody>) -> Response {
     // Brief, deliberately vague — "invalid credentials" rather than
     // "no such user" / "wrong password". Same anti-enumeration posture
-    // as Sonarr/Radarr.
-    let pub_ = match state.admin.verify(&body.username, &body.password) {
+    // as Sonarr/Radarr. The bcrypt verify is ~100-200 ms of pure CPU on
+    // an unauthenticated endpoint, so it runs on the blocking pool.
+    let admin = state.admin.clone();
+    let verdict = tokio::task::spawn_blocking(move || admin.verify(&body.username, &body.password))
+        .await
+        .unwrap_or_else(|e| Err(openpxe_core::Error::Other(e.into())));
+    let pub_ = match verdict {
         Ok(Some(u)) => u,
         Ok(None) => {
             return (
@@ -394,11 +411,18 @@ pub async fn api_update_credentials(
         )
             .into_response();
     }
-    let result = state.admin.update_credentials(
-        &body.current_password,
-        body.new_username.as_deref(),
-        body.new_password.as_deref(),
-    );
+    // Two bcrypt operations (verify current + hash new) plus a sync disk
+    // persist — run the lot on the blocking pool.
+    let admin = state.admin.clone();
+    let result = tokio::task::spawn_blocking(move || {
+        admin.update_credentials(
+            &body.current_password,
+            body.new_username.as_deref(),
+            body.new_password.as_deref(),
+        )
+    })
+    .await
+    .unwrap_or_else(|e| Err(openpxe_core::Error::Other(e.into())));
     match result {
         Ok(pub_) => {
             state.sessions.revoke_all();