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]>

Cargo.toml

@@ -12,7 +12,7 @@ members = [
 ]
 
 [workspace.package]
-version = "0.5.9"
+version = "0.6.2"
 edition = "2021"
 rust-version = "1.95"
 license = "MIT OR Apache-2.0"
@@ -20,21 +20,23 @@ repository = "https://gitea.milesward.dev/mward4/OpenPXE"
 authors = ["OpenPXE contributors"]
 
 [workspace.dependencies]
-tokio = { version = "1.40", features = ["full"] }
+tokio = { version = "1.52", features = ["full"] }
 tokio-util = { version = "0.7", features = ["io"] }
 tokio-stream = { version = "0.1", features = ["sync"] }
 futures = "0.3"
 async-trait = "0.1"
 
-dhcproto = "0.12"
+# v0.6.2: dhcproto 0.15 drops the deprecated trust-dns-proto dependency
-socket2 = { version = "0.5", features = ["all"] }
+# (replaced by hickory-proto) and carries three releases of DHCP option
+# coverage accumulated upstream — both directly relevant to the proxy core.
+dhcproto = "0.15"
+socket2 = { version = "0.6", features = ["all"] }
 bytes = "1.7"
-nom = "7.1"
 
 axum = { version = "0.7", features = ["macros", "multipart", "http2"] }
 tower = "0.5"
 tower-http = { version = "0.6", features = ["fs", "trace", "cors", "limit"] }
-hyper = "1.4"
+hyper = "1.9"
 reqwest = { version = "0.12", default-features = false, features = ["rustls-tls", "stream", "json"] }
 
 serde = { version = "1.0", features = ["derive"] }
@@ -52,9 +54,11 @@ thiserror = "2.0"
 clap = { version = "4.5", features = ["derive", "env"] }
 uuid = { version = "1.10", features = ["v4", "serde"] }
 time = { version = "0.3", features = ["serde", "serde-human-readable", "formatting", "macros"] }
+# sha2 stays 0.10 deliberately: bergshamra-crypto requires ^0.10, and
+# bumping to 0.11 would split the RustCrypto digest stack in the tree.
 sha2 = "0.10"
 hex = "0.4"
-bcrypt = "0.15"
+bcrypt = "0.19"
 parking_lot = "0.12"
 rust-embed = { version = "8.5", features = ["include-exclude"] }
 

README.md

@@ -14,7 +14,7 @@
 </p>
 
 <p align="center">
-  <img alt="release"   src="https://img.shields.io/badge/release-v0.5.8-2874d7" />
+  <img alt="release"   src="https://img.shields.io/badge/release-v0.6.0-2874d7" />
   <img alt="license"   src="https://img.shields.io/badge/license-MIT%20%7C%20Apache--2.0-59824f" />
   <img alt="rust"      src="https://img.shields.io/badge/built%20with-Rust-fb8841?logo=rust&logoColor=white" />
   <img alt="container" src="https://img.shields.io/badge/container--native-OCI%20%C2%B7%20OpenShift-2496ED?logo=docker&logoColor=white" />
@@ -32,7 +32,7 @@ Upload `.iso` files (or point at a remote share), and any machine on the network
 them — Linux installers, live tools, or stock Windows setup — with **zero iPXE knowledge
 required by the operator.**
 
-> **Status — v0.5.5, late pre-beta.** The full PXE stack, web UI, remote ISO libraries
+> **Status — v0.6.0, late pre-beta.** The full PXE stack, web UI, remote ISO libraries
 > (SMB/NFS/SFTP), Windows deployment, queued fleet rollout, SAML SSO, and Prometheus
 > metrics are implemented and test-covered. The release checklist gates every tag on the
 > full test suite + `clippy`. Currently in real-hardware validation.

crates/core/src/arch.rs

@@ -23,6 +23,25 @@ pub enum ClientArch {
     Unknown(u16),
 }
 
+/// Which iPXE network backend to advertise to a client (v0.6.1).
+///
+/// OpenPXE serves [`DriverMode::Firmware`] first (the firmware's own NIC
+/// stack, via `snponly`/`undionly`) and only escalates a specific MAC to
+/// [`DriverMode::Builtin`] (iPXE's bundled NIC drivers) automatically, when a
+/// firmware-net boot fails to chainload. There is no operator toggle — the
+/// DHCP proxy decides per client.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, Default, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case")]
+pub enum DriverMode {
+    /// Reuse the firmware UNDI/SNP NIC stack (`snponly.efi`, `undionly.kpxe`).
+    /// Default, smallest, most reliable for chainloading.
+    #[default]
+    Firmware,
+    /// iPXE's own bundled NIC drivers (`ipxe.efi`, `ipxe.pxe`). Fallback for
+    /// hardware whose firmware NIC stack is missing or buggy.
+    Builtin,
+}
+
 impl ClientArch {
     #[must_use]
     pub fn from_option_93(value: u16) -> Self {
@@ -39,18 +58,47 @@ impl ClientArch {
     /// Default iPXE binary filename to return via TFTP for this architecture.
     /// Uses `snponly` variants which reuse the firmware's UNDI/SNP network
     /// stack — smaller binaries and broader hardware compatibility than the
-    /// all-drivers-included `ipxe.efi`.
+    /// all-drivers-included `ipxe.efi`. Equivalent to
+    /// [`Self::ipxe_bootfile_mode`] with [`DriverMode::Firmware`]; kept as a
+    /// convenience for the common firmware-net path.
     #[must_use]
     pub fn ipxe_bootfile(self) -> Option<&'static str> {
-        Some(match self {
+        self.ipxe_bootfile_mode(DriverMode::Firmware)
-            Self::LegacyX86 => "undionly.kpxe",
+    }
-            Self::Ia32Uefi => "snponly-i386.efi",
+
-            Self::X64Uefi => "snponly.efi",
+    /// iPXE binary filename for this architecture under a given network
-            // ARM32 UEFI: upstream boot.ipxe.org does not publish a prebuilt
+    /// [`DriverMode`].
-            // snponly variant for this arch. We return None so the DHCP
+    ///
-            // proxy declines rather than advertising a file we can't serve.
+    /// * [`DriverMode::Firmware`] — the `snponly`/`undionly` builds that reuse
-            Self::Arm32Uefi | Self::Unknown(_) => return None,
+    ///   the firmware's UNDI/SNP NIC stack. Smallest, and the most reliable
-            Self::Arm64Uefi => "snponly-arm64.efi",
+    ///   choice for chainloading because the firmware just proved its network
+    ///   works by downloading the NBP. This is the default first attempt.
+    /// * [`DriverMode::Builtin`] — the all-drivers `ipxe.efi`/`ipxe.pxe`
+    ///   builds that carry iPXE's *own* NIC drivers. The automatic fallback
+    ///   for clients whose firmware NIC stack is missing or buggy (v0.6.1):
+    ///   the DHCP proxy escalates a MAC to this mode when a firmware-net boot
+    ///   never completes the iPXE handoff. iPXE still includes the `snp`
+    ///   driver here too, so it degrades gracefully.
+    #[must_use]
+    pub fn ipxe_bootfile_mode(self, mode: DriverMode) -> Option<&'static str> {
+        Some(match (self, mode) {
+            // Legacy x86 BIOS: UNDI (firmware) vs full native-driver build.
+            (Self::LegacyX86, DriverMode::Firmware) => "undionly.kpxe",
+            (Self::LegacyX86, DriverMode::Builtin) => "ipxe.pxe",
+            // IA32 UEFI.
+            (Self::Ia32Uefi, DriverMode::Firmware) => "snponly-i386.efi",
+            (Self::Ia32Uefi, DriverMode::Builtin) => "ipxe-i386.efi",
+            // x86_64 UEFI — the overwhelmingly common modern client.
+            (Self::X64Uefi, DriverMode::Firmware) => "snponly.efi",
+            (Self::X64Uefi, DriverMode::Builtin) => "ipxe.efi",
+            // ARM64 UEFI.
+            (Self::Arm64Uefi, DriverMode::Firmware) => "snponly-arm64.efi",
+            (Self::Arm64Uefi, DriverMode::Builtin) => "ipxe-arm64.efi",
+            // ARM32 UEFI: upstream boot.ipxe.org publishes no prebuilt binary
+            // for this arch in either mode. Unknown arches likewise. Return
+            // None so the DHCP proxy declines rather than advertising a file
+            // we can't serve.
+            (Self::Arm32Uefi | Self::Unknown(_), _) => return None,
         })
     }
 
@@ -133,6 +181,58 @@ mod tests {
         assert_eq!(ClientArch::Unknown(0xFFFF).ipxe_bootfile(), None);
     }
 
+    #[test]
+    fn bootfile_default_is_firmware_mode() {
+        // The convenience method must equal the explicit Firmware mode.
+        for a in [
+            ClientArch::LegacyX86,
+            ClientArch::Ia32Uefi,
+            ClientArch::X64Uefi,
+            ClientArch::Arm64Uefi,
+            ClientArch::Arm32Uefi,
+            ClientArch::Unknown(0x99),
+        ] {
+            assert_eq!(
+                a.ipxe_bootfile(),
+                a.ipxe_bootfile_mode(DriverMode::Firmware)
+            );
+        }
+    }
+
+    #[test]
+    fn builtin_mode_maps_to_all_drivers_binaries() {
+        assert_eq!(
+            ClientArch::LegacyX86.ipxe_bootfile_mode(DriverMode::Builtin),
+            Some("ipxe.pxe")
+        );
+        assert_eq!(
+            ClientArch::X64Uefi.ipxe_bootfile_mode(DriverMode::Builtin),
+            Some("ipxe.efi")
+        );
+        assert_eq!(
+            ClientArch::Ia32Uefi.ipxe_bootfile_mode(DriverMode::Builtin),
+            Some("ipxe-i386.efi")
+        );
+        assert_eq!(
+            ClientArch::Arm64Uefi.ipxe_bootfile_mode(DriverMode::Builtin),
+            Some("ipxe-arm64.efi")
+        );
+        // No binary for ARM32 / unknown in either mode.
+        assert_eq!(
+            ClientArch::Arm32Uefi.ipxe_bootfile_mode(DriverMode::Builtin),
+            None
+        );
+        assert_eq!(
+            ClientArch::Unknown(0x99).ipxe_bootfile_mode(DriverMode::Builtin),
+            None
+        );
+    }
+
+    #[test]
+    fn driver_mode_default_is_firmware() {
+        assert_eq!(DriverMode::default(), DriverMode::Firmware);
+    }
+
     #[test]
     fn firmware_class_detects_ipxe_over_pxeclient() {
         let c = FirmwareClass::classify(Some(b"PXEClient:Arch:00007"), Some(b"iPXE"));

crates/core/src/auth.rs

@@ -125,9 +125,7 @@ impl AdminStore {
         {
             let mut g = self.inner.write();
             if g.admin.is_some() {
-                return Err(Error::Invalid(
+                return Err(Error::Invalid("admin account already configured".into()));
-                    "admin account already configured".into(),
-                ));
             }
             g.admin = Some(admin.clone());
         }
@@ -346,7 +344,7 @@ mod tests {
         assert!(s.bootstrap("", "hunter2hunter2").is_err());
         assert!(s.bootstrap("ad:min", "hunter2hunter2").is_err()); // ':' reserved
         assert!(s.bootstrap("admin", "short").is_err()); // <8 chars
-        // 65-char username is too long.
+                                                         // 65-char username is too long.
         let long = "a".repeat(65);
         assert!(s.bootstrap(&long, "hunter2hunter2").is_err());
     }

crates/core/src/client.rs

@@ -12,11 +12,9 @@ use time::OffsetDateTime;
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub enum ClientEvent {
     DhcpDiscover,
-    DhcpRequest,
     PxeBootServerRequest,
     TftpRead { file: String },
     HttpScriptFetch { target: String },
-    HttpIsoAsset { file: String },
 }
 
 #[derive(Debug, Clone, Serialize, Deserialize)]
@@ -32,8 +30,6 @@ pub struct ClientSnapshot {
     // Events are left with default serialization (9-tuple) — they're
     // diagnostic only and not consumed by the UI today.
     pub events: Vec<(OffsetDateTime, ClientEvent)>,
-    /// The boot target (ISO id) last selected via the iPXE menu, if any.
-    pub selected_target: Option<String>,
 }
 
 #[derive(Debug, Default)]
@@ -66,7 +62,6 @@ impl ClientRegistry {
                 first_seen: now,
                 last_seen: now,
                 events: Vec::new(),
-                selected_target: None,
             });
         entry.last_seen = now;
         if ip.is_some() {
@@ -84,13 +79,6 @@ impl ClientRegistry {
         }
     }
 
-    pub fn set_selected_target(&self, mac: &str, target: Option<String>) {
-        let mut guard = self.inner.write();
-        if let Some(c) = guard.get_mut(mac) {
-            c.selected_target = target;
-        }
-    }
-
     #[must_use]
     pub fn list(&self) -> Vec<ClientSnapshot> {
         let guard = self.inner.read();
@@ -99,9 +87,4 @@ impl ClientRegistry {
         v.sort_by_key(|c| std::cmp::Reverse(c.last_seen));
         v
     }
-
-    #[must_use]
-    pub fn get(&self, mac: &str) -> Option<ClientSnapshot> {
-        self.inner.read().get(mac).cloned()
-    }
 }

crates/core/src/config.rs

@@ -40,10 +40,6 @@ pub struct NetworkConfig {
     pub dhcp_port: u16,
     /// UDP port for PXE Boot Server discovery. Standard is 4011.
     pub pxe_port: u16,
-    /// Optional allowlist of client MAC prefixes (OUI). Empty = serve everyone.
-    pub mac_allowlist: Vec<String>,
-    /// Optional allowlist of subnets (CIDR). Empty = serve everyone.
-    pub subnet_allowlist: Vec<String>,
 }
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, Default)]
@@ -105,8 +101,6 @@ impl Default for NetworkConfig {
             dhcp_bind: IpAddr::V4(Ipv4Addr::UNSPECIFIED),
             dhcp_port: 67,
             pxe_port: 4011,
-            mac_allowlist: Vec::new(),
-            subnet_allowlist: Vec::new(),
         }
     }
 }

crates/core/src/lib.rs

@@ -21,7 +21,7 @@ pub mod settings;
 pub mod sso;
 pub mod wol;
 
-pub use arch::{ClientArch, FirmwareClass};
+pub use arch::{ClientArch, DriverMode, FirmwareClass};
 pub use auth::{AdminAccount, AdminPublic, AdminStore};
 pub use boot_log::{BootEvent, BootLog};
 pub use branding::{ext_for_mime, BrandingStore, LogoSlot, ALLOWED_LOGO_MIMES, MAX_LOGO_BYTES};

crates/core/src/notify.rs

@@ -303,7 +303,10 @@ mod tests {
             smtp_host: "smtp.example.com".into(),
             ..Default::default()
         });
-        assert!(matches!(r, Err(Error::Invalid(_))), "missing recipient should reject");
+        assert!(
+            matches!(r, Err(Error::Invalid(_))),
+            "missing recipient should reject"
+        );
         s.replace(NotifyConfig {
             enabled: true,
             kind: NotifyKind::Smtp,

crates/core/src/wol.rs

@@ -181,10 +181,7 @@ mod tests {
             Ipv4Addr::new(192, 168, 1, 255)
         );
         assert_eq!(
-            subnet_broadcast(
+            subnet_broadcast(Ipv4Addr::new(10, 5, 3, 7), Ipv4Addr::new(255, 255, 0, 0)),
-                Ipv4Addr::new(10, 5, 3, 7),
-                Ipv4Addr::new(255, 255, 0, 0)
-            ),
             Ipv4Addr::new(10, 5, 255, 255)
         );
     }
@@ -196,7 +193,8 @@ mod tests {
         // and confirm send_magic transmits the exact 102-byte packet.
         let rx = UdpSocket::bind(SocketAddrV4::new(Ipv4Addr::LOCALHOST, 0)).unwrap();
         let port = rx.local_addr().unwrap().port();
-        rx.set_read_timeout(Some(std::time::Duration::from_secs(2))).unwrap();
+        rx.set_read_timeout(Some(std::time::Duration::from_secs(2)))
+            .unwrap();
 
         let packet = magic_packet([0x0a, 0x1b, 0x2c, 0x3d, 0x4e, 0x5f]);
         let sent = send_magic(&packet, &[Ipv4Addr::LOCALHOST], port).unwrap();

crates/dhcp-proxy/Cargo.toml

@@ -18,3 +18,4 @@ tracing.workspace = true
 thiserror.workspace = true
 anyhow.workspace = true
 bytes.workspace = true
+parking_lot.workspace = true

crates/dhcp-proxy/src/escalation.rs

@@ -0,0 +1,237 @@
+//! Automatic per-MAC NIC driver-mode escalation (v0.6.1).
+//!
+//! OpenPXE serves the firmware-net iPXE build (`snponly`/`undionly`) by
+//! default — it's the most reliable choice for chainloading because the
+//! firmware just proved its network works by downloading the NBP. A minority
+//! of NICs have a missing or buggy firmware UNDI/SNP stack; those clients
+//! TFTP the binary fine, but then iPXE can't bring the link up, so the
+//! tell-tale second DHCP DISCOVER carrying the `iPXE` user-class never arrives
+//! and the machine eventually re-PXE-boots.
+//!
+//! We detect exactly that: a *fresh* firmware DISCOVER from a MAC whose
+//! previous firmware attempt was never confirmed by an iPXE handoff means the
+//! firmware-net build failed → escalate that MAC to [`DriverMode::Builtin`]
+//! (iPXE's own NIC drivers). The decision is sticky — once a MAC settles on a
+//! mode that completes the handoff, later boots go straight to it. There is no
+//! operator toggle; it just works, and the default (firmware) path is
+//! unchanged so hardware that already boots never regresses.
+
+use openpxe_core::DriverMode;
+use parking_lot::Mutex;
+use std::collections::HashMap;
+use std::time::{Duration, Instant};
+
+/// Multiple DISCOVERs within this window belong to the *same* boot (DHCP
+/// retransmits, plus the :4011 PXE Boot Server query that follows the :67
+/// DISCOVER). They must not be mistaken for a failed-and-retried boot.
+const SAME_BOOT_DEBOUNCE: Duration = Duration::from_secs(8);
+
+/// Forget a MAC's state after this long with no activity, so a transient
+/// escalation doesn't pin a client to Builtin forever and the map stays
+/// bounded over a long-running deployment.
+const ENTRY_TTL: Duration = Duration::from_mins(30);
+
+/// Hard cap on tracked MACs. Past this we evict the least-recently-seen
+/// entry — escalation is best-effort, never a memory-growth vector.
+const MAX_ENTRIES: usize = 4096;
+
+/// How often (at most) the whole map is swept for expired entries.
+/// Correctness doesn't depend on the sweep — a stale entry is also
+/// detected inline when its MAC next appears — so the sweep only bounds
+/// memory for MACs that never return, and amortizing it keeps the
+/// per-packet path O(1) instead of O(map).
+const PRUNE_INTERVAL: Duration = Duration::from_mins(1);
+
+#[derive(Debug, Clone, Copy)]
+struct Entry {
+    mode: DriverMode,
+    /// True once we've served `mode` and are waiting for the iPXE handoff to
+    /// confirm it worked. A *new* boot arriving while this is still true means
+    /// the previous attempt failed and we should escalate.
+    awaiting_confirm: bool,
+    last_seen: Instant,
+}
+
+#[derive(Debug)]
+struct Inner {
+    map: HashMap<String, Entry>,
+    /// When the last full TTL sweep ran — see [`PRUNE_INTERVAL`].
+    last_prune: Instant,
+}
+
+impl Default for Inner {
+    fn default() -> Self {
+        Self {
+            map: HashMap::new(),
+            last_prune: Instant::now(),
+        }
+    }
+}
+
+/// Tracks per-MAC driver-mode escalation. Cheap to share via `Arc`.
+#[derive(Debug, Default)]
+pub struct DriverEscalation {
+    inner: Mutex<Inner>,
+}
+
+impl DriverEscalation {
+    #[must_use]
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    /// Decide the driver mode for a firmware (PXEClient/HTTPClient) boot from
+    /// `mac`. `primary` is true for the main DHCP DISCOVER (:67) and false for
+    /// the PXE Boot Server query (:4011); only the primary path drives
+    /// escalation, and only when it's clearly a *new* boot (outside the
+    /// same-boot debounce). The :4011 path just echoes the current mode.
+    pub fn mode_for_firmware_attempt(&self, mac: &str, primary: bool) -> DriverMode {
+        self.decide_at(mac, primary, Instant::now())
+    }
+
+    /// Record that `mac` completed the iPXE handoff (a DISCOVER carrying the
+    /// `iPXE` user-class). The mode we last served worked, so stop awaiting
+    /// confirmation and keep it sticky for next time.
+    pub fn mark_ipxe_success(&self, mac: &str) {
+        self.confirm_at(mac, Instant::now());
+    }
+
+    fn decide_at(&self, mac: &str, primary: bool, now: Instant) -> DriverMode {
+        let mut g = self.inner.lock();
+        if now.duration_since(g.last_prune) >= PRUNE_INTERVAL {
+            g.map
+                .retain(|_, e| now.duration_since(e.last_seen) < ENTRY_TTL);
+            g.last_prune = now;
+        }
+        // Inline staleness check: a MAC whose entry outlived the TTL starts
+        // fresh even when the amortized sweep above hasn't caught it yet.
+        if g.map
+            .get(mac)
+            .is_some_and(|e| now.duration_since(e.last_seen) >= ENTRY_TTL)
+        {
+            g.map.remove(mac);
+        }
+
+        match g.map.get_mut(mac) {
+            None => {
+                g.map.insert(
+                    mac.to_owned(),
+                    Entry {
+                        mode: DriverMode::Firmware,
+                        // Only the primary DISCOVER opens a confirmation window.
+                        awaiting_confirm: primary,
+                        last_seen: now,
+                    },
+                );
+                if g.map.len() > MAX_ENTRIES {
+                    evict_oldest(&mut g.map);
+                }
+                DriverMode::Firmware
+            }
+            Some(entry) => {
+                let recent = now.duration_since(entry.last_seen) < SAME_BOOT_DEBOUNCE;
+                if primary && !recent {
+                    // A genuinely new boot. If the previous attempt was never
+                    // confirmed, the firmware-net build failed → escalate to
+                    // the all-drivers build. Builtin is the most capable build
+                    // we have, so it's the single escalation target (and a MAC
+                    // already on Builtin simply stays there).
+                    if entry.awaiting_confirm {
+                        entry.mode = DriverMode::Builtin;
+                    }
+                    entry.awaiting_confirm = true;
+                }
+                entry.last_seen = now;
+                entry.mode
+            }
+        }
+    }
+
+    fn confirm_at(&self, mac: &str, now: Instant) {
+        let mut g = self.inner.lock();
+        if let Some(e) = g.map.get_mut(mac) {
+            e.awaiting_confirm = false;
+            e.last_seen = now;
+        }
+    }
+}
+
+fn evict_oldest(map: &mut HashMap<String, Entry>) {
+    if let Some(oldest) = map
+        .iter()
+        .min_by_key(|(_, e)| e.last_seen)
+        .map(|(k, _)| k.clone())
+    {
+        map.remove(&oldest);
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn firmware_first_then_escalates_on_unconfirmed_retry() {
+        let e = DriverEscalation::new();
+        let t0 = Instant::now();
+        // Boot 1, primary DISCOVER: firmware.
+        assert_eq!(e.decide_at("aa", true, t0), DriverMode::Firmware);
+        // Same boot's :4011 query (+1s, within debounce): still firmware, no escalation.
+        assert_eq!(
+            e.decide_at("aa", false, t0 + Duration::from_secs(1)),
+            DriverMode::Firmware
+        );
+        // Firmware net failed → no iPXE handoff → machine re-PXE-boots much
+        // later: escalate to builtin drivers.
+        assert_eq!(
+            e.decide_at("aa", true, t0 + Duration::from_mins(1)),
+            DriverMode::Builtin
+        );
+    }
+
+    #[test]
+    fn builtin_is_sticky_after_success() {
+        let e = DriverEscalation::new();
+        let t0 = Instant::now();
+        assert_eq!(e.decide_at("bb", true, t0), DriverMode::Firmware);
+        assert_eq!(
+            e.decide_at("bb", true, t0 + Duration::from_mins(1)),
+            DriverMode::Builtin
+        );
+        // Builtin worked this time — confirm the handoff.
+        e.confirm_at("bb", t0 + Duration::from_secs(61));
+        // Next cold boot goes straight to builtin (no wasted firmware attempt).
+        assert_eq!(
+            e.decide_at("bb", true, t0 + Duration::from_mins(2)),
+            DriverMode::Builtin
+        );
+    }
+
+    #[test]
+    fn confirmed_firmware_never_escalates() {
+        let e = DriverEscalation::new();
+        let t0 = Instant::now();
+        assert_eq!(e.decide_at("cc", true, t0), DriverMode::Firmware);
+        // snponly worked: handoff confirmed.
+        e.confirm_at("cc", t0 + Duration::from_secs(2));
+        // A later boot stays on firmware — no spurious escalation.
+        assert_eq!(
+            e.decide_at("cc", true, t0 + Duration::from_mins(5)),
+            DriverMode::Firmware
+        );
+    }
+
+    #[test]
+    fn stale_entry_is_forgotten_and_resets_to_firmware() {
+        let e = DriverEscalation::new();
+        let t0 = Instant::now();
+        assert_eq!(e.decide_at("dd", true, t0), DriverMode::Firmware);
+        assert_eq!(
+            e.decide_at("dd", true, t0 + Duration::from_mins(1)),
+            DriverMode::Builtin
+        );
+        // After the TTL with no activity the entry is pruned → fresh firmware.
+        let later = t0 + Duration::from_mins(1) + ENTRY_TTL + Duration::from_secs(1);
+        assert_eq!(e.decide_at("dd", true, later), DriverMode::Firmware);
+    }
+}

crates/dhcp-proxy/src/lib.rs

@@ -17,7 +17,9 @@
 //!   clients silently drop them.
 #![forbid(unsafe_code)]
 
+pub mod escalation;
 pub mod reply;
 pub mod server;
 
+pub use escalation::DriverEscalation;
 pub use server::DhcpProxyServer;

crates/dhcp-proxy/src/reply.rs

@@ -9,7 +9,7 @@
 //!     pass, or the HTTP URL of the boot script once iPXE has chained.
 
 use dhcproto::v4::{DhcpOption, Message, MessageType, Opcode, OptionCode};
-use openpxe_core::{ClientArch, FirmwareClass};
+use openpxe_core::{ClientArch, DriverMode, FirmwareClass};
 use std::net::Ipv4Addr;
 
 /// Where the reply directs the client next.
@@ -31,6 +31,11 @@ pub struct ReplyContext<'a> {
     pub our_ip: Ipv4Addr,
     pub arch: ClientArch,
     pub class: FirmwareClass,
+    /// Which iPXE network backend to advertise for this client. The DHCP
+    /// proxy fills this from the automatic per-MAC escalation state: normally
+    /// [`DriverMode::Firmware`], escalated to [`DriverMode::Builtin`] for a
+    /// MAC whose firmware-net boot failed to chainload (v0.6.1).
+    pub driver_mode: DriverMode,
     /// Public base URL (scheme://host[:port]) used in HTTP directives.
     pub public_base_url: &'a str,
 }
@@ -52,9 +57,13 @@ pub fn decide(ctx: &ReplyContext<'_>) -> BootDirective {
         },
         FirmwareClass::HttpClient => {
             // UEFI HTTP boot: client wants an http:// URL in option 67
-            // pointing at an EFI executable. We serve ipxe.efi over HTTP;
+            // pointing at an EFI executable. We serve the iPXE EFI build for
-            // it'll then do the same script-fetch the iPXE path does.
+            // the negotiated driver mode over HTTP; it'll then do the same
-            let name = ctx.arch.ipxe_bootfile().unwrap_or("snponly.efi");
+            // script-fetch the iPXE path does.
+            let name = ctx
+                .arch
+                .ipxe_bootfile_mode(ctx.driver_mode)
+                .unwrap_or("snponly.efi");
             BootDirective::HttpScript {
                 url: format!(
                     "{}/ipxe/{}",
@@ -63,7 +72,7 @@ pub fn decide(ctx: &ReplyContext<'_>) -> BootDirective {
                 ),
             }
         }
-        FirmwareClass::PxeClient => match ctx.arch.ipxe_bootfile() {
+        FirmwareClass::PxeClient => match ctx.arch.ipxe_bootfile_mode(ctx.driver_mode) {
             Some(name) => BootDirective::TftpIpxe {
                 filename: name.to_string(),
             },

crates/dhcp-proxy/src/server.rs

@@ -1,10 +1,11 @@
 //! UDP listener loop for the DHCP proxy. Accepts on :67 (and :4011 on a
 //! second socket) and dispatches each datagram through the pure reply logic.
 
+use crate::escalation::DriverEscalation;
 use crate::reply::{build_reply, decide, BootDirective, ReplyContext};
 use dhcproto::v4::{DhcpOption, Message, OptionCode};
 use dhcproto::{Decodable, Decoder, Encodable, Encoder};
-use openpxe_core::{ClientArch, ClientEvent, ClientRegistry, FirmwareClass};
+use openpxe_core::{ClientArch, ClientEvent, ClientRegistry, DriverMode, FirmwareClass};
 use socket2::{Domain, Protocol, Socket, Type};
 use std::net::{IpAddr, Ipv4Addr, SocketAddr, SocketAddrV4};
 use std::sync::Arc;
@@ -18,6 +19,9 @@ pub struct DhcpProxyServer {
     public_base_url: String,
     clients: Arc<ClientRegistry>,
     metrics: openpxe_core::Metrics,
+    /// Automatic per-MAC NIC driver-mode escalation (v0.6.1). Shared across
+    /// the :67 and :4011 listener tasks via the server `Arc`.
+    escalation: DriverEscalation,
 }
 
 impl DhcpProxyServer {
@@ -38,6 +42,7 @@ impl DhcpProxyServer {
             public_base_url,
             clients,
             metrics,
+            escalation: DriverEscalation::new(),
         }
     }
 
@@ -126,11 +131,30 @@ impl DhcpProxyServer {
             },
         );
 
+        // Automatic NIC driver-mode selection (v0.6.1). The default is
+        // firmware-net (snponly/undionly). A successful iPXE handoff confirms
+        // the current mode works for this MAC; a fresh firmware boot whose
+        // predecessor never handed off escalates the MAC to iPXE's built-in
+        // NIC drivers. No operator toggle — the firmware path is unchanged so
+        // hardware that already boots never regresses.
+        let driver_mode = match class {
+            FirmwareClass::IpxeUserClass => {
+                self.escalation.mark_ipxe_success(&mac);
+                DriverMode::Firmware // unused: this path serves the HTTP script
+            }
+            FirmwareClass::PxeClient | FirmwareClass::HttpClient => self
+                .escalation
+                .mode_for_firmware_attempt(&mac, label == "67"),
+            // Unreachable: FirmwareClass::Other returned above.
+            FirmwareClass::Other => DriverMode::Firmware,
+        };
+
         let ctx = ReplyContext {
             request: &request,
             our_ip: self.our_ip,
             arch,
             class,
+            driver_mode,
             public_base_url: &self.public_base_url,
         };
         let directive = decide(&ctx);
@@ -154,7 +178,7 @@ impl DhcpProxyServer {
         sock.send_to(&out, dest).await?;
         tracing::info!(
             target: "openpxe::dhcp",
-            mac=%mac, arch=arch.as_str(), class=?class, dest=%dest, directive=?directive,
+            mac=%mac, arch=arch.as_str(), class=?class, driver=?driver_mode, dest=%dest, directive=?directive,
             "PXE reply sent"
         );
         Ok(())
@@ -213,11 +237,16 @@ fn bind_udp(bind: IpAddr, port: u16, broadcast: bool) -> anyhow::Result<UdpSocke
 }
 
 fn format_mac(chaddr: &[u8]) -> String {
-    let take = chaddr.iter().take(6).copied().collect::<Vec<_>>();
+    use std::fmt::Write;
-    take.iter()
+    // One allocation — this runs for every PXE datagram we answer.
-        .map(|b| format!("{b:02x}"))
+    let mut s = String::with_capacity(17);
-        .collect::<Vec<_>>()
+    for (i, b) in chaddr.iter().take(6).enumerate() {
-        .join(":")
+        if i > 0 {
+            s.push(':');
+        }
+        let _ = write!(s, "{b:02x}");
+    }
+    s
 }
 
 /// Walk raw DHCP options looking for option 93 (Client System Architecture)

crates/http-api/src/app.rs

@@ -35,7 +35,7 @@ use openpxe_core::{
     encoding::pct_encode, ext_for_mime, wol, BootEvent, ClientEvent, DeployProfile, Error,
     LogoSlot, NotifyConfig, Settings, SsoConfig, ALLOWED_LOGO_MIMES, MAX_LOGO_BYTES,
 };
-use openpxe_ipxe_assets::asset_bytes;
+use openpxe_ipxe_assets::asset_slice;
 use openpxe_iso_store::{
     render_template, IsoCategory, IsoMeta, IsoSource, NfsAddRequest, SftpAddRequest, SmbAddRequest,
     SmbState, UnattendedKind, UnattendedMeta,
@@ -406,6 +406,22 @@ fn bundled_logo_response() -> Response {
 /// brand mark falls back to the *default* background for the PXE screen
 /// (the WebUI still renders the SVG natively in the top-left).
 async fn ui_pxe_logo(State(state): State<AppState>) -> Response {
+    // The composite is a pure function of the uploaded logo, so the
+    // encoded PNG is cached keyed on the branding revision — an upload
+    // or clear bumps the rev and invalidates it. The response headers
+    // stay `no-cache` (clients must refetch); only the server-side
+    // ~50-200 ms decode/compose/encode is skipped per boot.
+    let rev = state.branding.logo_rev();
+    let cached = state
+        .pxe_bg_cache
+        .lock()
+        .as_ref()
+        .filter(|(r, _)| *r == rev)
+        .map(|(_, png)| png.clone());
+    if let Some(png) = cached {
+        return pxe_png_response(png);
+    }
+
     // Resolve the operator's raster upload, if any and if it's a format
     // iPXE/our compositor can consume. SVG (or a missing/unreadable
     // file) yields `None`, which composes the default background.
@@ -450,6 +466,12 @@ async fn ui_pxe_logo(State(state): State<AppState>) -> Response {
                 .into_response();
         }
     };
+    let png = bytes::Bytes::from(composed);
+    *state.pxe_bg_cache.lock() = Some((rev, png.clone()));
+    pxe_png_response(png)
+}
+
+fn pxe_png_response(png: bytes::Bytes) -> Response {
     (
         [
             (header::CONTENT_TYPE, HeaderValue::from_static("image/png")),
@@ -460,7 +482,7 @@ async fn ui_pxe_logo(State(state): State<AppState>) -> Response {
                 HeaderValue::from_static("no-cache, max-age=0"),
             ),
         ],
-        composed,
+        png,
     )
         .into_response()
 }
@@ -638,7 +660,23 @@ async fn boot_sub(
                                     ));
                                 }
                                 Some(token) => {
-                                    match state.iso_store.verify_password(&iso.id, token) {
+                                    // bcrypt verify costs ~100-200 ms of pure
+                                    // CPU and this path is unauthenticated —
+                                    // run it on the blocking pool so password
+                                    // probes can't stall the workers that are
+                                    // streaming ISO bytes to imaging machines.
+                                    let store = state.iso_store.clone();
+                                    let iso_id = iso.id.clone();
+                                    let tok = token.to_string();
+                                    let verdict = match tokio::task::spawn_blocking(move || {
+                                        store.verify_password(&iso_id, &tok)
+                                    })
+                                    .await
+                                    {
+                                        Ok(v) => v,
+                                        Err(e) => Err(openpxe_core::Error::Other(e.into())),
+                                    };
+                                    match verdict {
                                         Ok(true) => { /* fall through to render the entry */ }
                                         Ok(false) => {
                                             // Don't log the candidate — just the
@@ -735,9 +773,15 @@ async fn ipxe_binary(AxumPath(name): AxumPath<String>) -> Response {
     if name.contains('/') || name.contains('\\') {
         return (StatusCode::BAD_REQUEST, "invalid name").into_response();
     }
-    let Some(bytes) = asset_bytes(&name) else {
+    let Some(data) = asset_slice(&name) else {
         return (StatusCode::NOT_FOUND, "no such ipxe asset").into_response();
     };
+    // Release builds embed the asset in rodata — serve it without the
+    // ~1 MiB per-request heap copy `into_owned` would cost.
+    let bytes = match data {
+        std::borrow::Cow::Borrowed(b) => bytes::Bytes::from_static(b),
+        std::borrow::Cow::Owned(v) => bytes::Bytes::from(v),
+    };
     (
         [
             (
@@ -768,7 +812,10 @@ async fn iso_raw(
     };
     match &meta.source {
         IsoSource::Local => {
-            let Some(path) = state.iso_store.iso_path_for(id) else {
+            // `local_path(&meta)` reuses the meta we already cloned —
+            // `iso_path_for(id)` would re-lock and deep-clone it again,
+            // hundreds of times per sanboot install.
+            let Some(path) = state.iso_store.local_path(&meta) else {
                 return (StatusCode::NOT_FOUND, "no such iso").into_response();
             };
             match stream_file_range(&path, headers.get(header::RANGE)).await {
@@ -1027,15 +1074,25 @@ fn parse_range(h: Option<&HeaderValue>, total: u64) -> Option<(u64, u64, bool)>
             return Some((total.saturating_sub(n), total.saturating_sub(1), true));
         }
     }
-    let mut parts = spec.splitn(2, '-');
+    // RFC 7233 §3.1: a Range header we can't parse is *ignored* (200 +
-    let start = parts
+    // full body), never coerced into a bogus 206 claiming the whole
-        .next()
+    // file. Only `first-pos[-last-pos]` with numeric positions reaches
-        .and_then(|s| s.parse::<u64>().ok())
+    // the partial path; `None` is reserved for syntactically valid but
-        .unwrap_or(0);
+    // unsatisfiable ranges (→ 416).
-    let end = parts
+    let full = Some((0, total.saturating_sub(1), false));
-        .next()
+    let Some((start_s, end_s)) = spec.split_once('-') else {
-        .and_then(|s| s.parse::<u64>().ok())
+        return full;
-        .unwrap_or(total.saturating_sub(1));
+    };
+    let Ok(start) = start_s.trim().parse::<u64>() else {
+        return full;
+    };
+    let end = if end_s.trim().is_empty() {
+        total.saturating_sub(1)
+    } else if let Ok(e) = end_s.trim().parse::<u64>() {
+        e
+    } else {
+        return full;
+    };
     if start >= total {
         return None;
     }

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.
+    // as Sonarr/Radarr. The bcrypt verify is ~100-200 ms of pure CPU on
-    let pub_ = match state.admin.verify(&body.username, &body.password) {
+    // 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(
+    // Two bcrypt operations (verify current + hash new) plus a sync disk
-        &body.current_password,
+    // persist — run the lot on the blocking pool.
-        body.new_username.as_deref(),
+    let admin = state.admin.clone();
-        body.new_password.as_deref(),
+    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();

crates/http-api/src/notify.rs

@@ -105,7 +105,11 @@ async fn send_email(cfg: &NotifyConfig, subject: &str, body: &str) -> Result<(),
             .trim()
             .parse()
             .map_err(|e| format!("invalid To address '{}': {e}", cfg.smtp_to))?)
-        .subject(if subject.is_empty() { "OpenPXE" } else { subject })
+        .subject(if subject.is_empty() {
+            "OpenPXE"
+        } else {
+            subject
+        })
         .body(body.to_string())
         .map_err(|e| format!("could not build email: {e}"))?;
 

crates/http-api/src/state.rs

@@ -11,6 +11,10 @@ use openpxe_iso_store::{
 use std::sync::Arc;
 use time::OffsetDateTime;
 
+/// Cached composited PXE boot-menu background: `(logo_rev, encoded PNG)`.
+/// See `AppState::pxe_bg_cache`.
+pub type PxeBgCache = Arc<parking_lot::Mutex<Option<(u64, bytes::Bytes)>>>;
+
 #[derive(Clone)]
 pub struct AppState {
     pub iso_store: IsoStore,
@@ -29,6 +33,12 @@ pub struct AppState {
     /// operator hasn't uploaded anything, the WebUI serves the bundled
     /// rainbow-horizon mark.
     pub branding: BrandingStore,
+    /// v0.6.2: cache of the composited PXE boot-menu background PNG,
+    /// keyed on the branding logo revision. Composing costs ~50-200 ms
+    /// of image decode/encode and **every** booting client fetches it
+    /// for `console --picture` — caching makes that one compose per
+    /// logo change instead of one per boot.
+    pub pxe_bg_cache: PxeBgCache,
     /// Forms-auth admin record + first-run bootstrap state. When
     /// `admin.is_configured() == false`, the auth middleware passes
     /// every request through and `/api/me` reports `setup_required`.

crates/http-api/tests/full_flow.rs

@@ -116,6 +116,7 @@ async fn build_state() -> (AppState, tempfile::TempDir) {
         hosts,
         boot_log,
         branding,
+        pxe_bg_cache: openpxe_http_api::state::PxeBgCache::default(),
         admin,
         sessions,
         sso,

crates/ipxe-assets/src/lib.rs

@@ -6,43 +6,40 @@
 //! missing, that architecture simply won't have PXE support — we log at
 //! startup and serve what we have.
 //!
-//! Filename convention (matches `ClientArch::ipxe_bootfile`):
+//! Filename convention (matches `ClientArch::ipxe_bootfile_mode`):
+//!
+//! DriverMode::Firmware (default — reuse the firmware UNDI/SNP NIC stack):
 //!   - `undionly.kpxe`       — Legacy x86 BIOS
 //!   - `snponly-i386.efi`    — IA32 UEFI
 //!   - `snponly.efi`         — x86_64 UEFI
-//!   - `snponly-arm32.efi`   — ARM32 UEFI
 //!   - `snponly-arm64.efi`   — ARM64 UEFI
-//!   - `ipxe.efi` (fallback) — UEFI with bundled drivers, if snponly fails on a NIC
+//!
+//! DriverMode::Builtin (v0.6.1 automatic fallback — iPXE's own NIC drivers,
+//! advertised when a firmware-net boot fails to chainload):
+//!   - `ipxe.pxe`            — Legacy x86 BIOS
+//!   - `ipxe-i386.efi`       — IA32 UEFI
+//!   - `ipxe.efi`            — x86_64 UEFI (built from source with PNG)
+//!   - `ipxe-arm64.efi`      — ARM64 UEFI
+//!
 //!   - `wimboot`             — Windows boot shim (fetched separately for WIM chains)
 #![forbid(unsafe_code)]
 
-use openpxe_core::ClientArch;
+use openpxe_core::{ClientArch, DriverMode};
 use rust_embed::Embed;
 
 #[derive(Embed)]
 #[folder = "../../assets/ipxe/"]
 #[include = "*.kpxe"]
 #[include = "*.efi"]
+#[include = "*.pxe"]
 #[include = "wimboot"]
 pub struct IpxeAssets;
 
-/// Return the embedded iPXE binary for `arch`, or `None` if we didn't bundle
+/// Return a named embedded asset (e.g. `snponly.efi`, `wimboot`) as a
-/// one for that architecture.
+/// `Cow` over the embedded bytes. In release builds the data is borrowed
-#[must_use]
+/// straight from the binary's rodata — **zero copy** — which matters
-pub fn bootfile_bytes(arch: ClientArch) -> Option<Vec<u8>> {
+/// because the TFTP and HTTP serving paths hit this for every boot
-    let name = arch.ipxe_bootfile()?;
+/// (`ipxe.efi` is ~1 MiB). Debug builds read from disk and return Owned.
-    IpxeAssets::get(name).map(|f| f.data.into_owned())
-}
-
-/// Return a named asset directly (e.g. `wimboot`, or a fallback `ipxe.efi`).
-#[must_use]
-pub fn asset_bytes(name: &str) -> Option<Vec<u8>> {
-    IpxeAssets::get(name).map(|f| f.data.into_owned())
-}
-
-/// Same as [`asset_bytes`] but returns the embedded slice directly,
-/// avoiding the heap copy when the caller only needs to read the
-/// payload. Falls back to None for unknown names.
 #[must_use]
 pub fn asset_slice(name: &str) -> Option<std::borrow::Cow<'static, [u8]>> {
     IpxeAssets::get(name).map(|f| f.data)
@@ -56,25 +53,42 @@ pub fn list_assets() -> Vec<String> {
         .collect()
 }
 
-/// Log at startup which iPXE binaries are present and which are missing.
+/// Log at startup which iPXE binaries are present and which are missing, for
+/// both driver modes. The Firmware-mode binaries are required for PXE on each
+/// arch; the Builtin-mode binaries are the optional automatic NIC-driver
+/// fallback (v0.6.1) — without one, escalation simply can't help that arch.
 pub fn log_availability() {
     let have: std::collections::HashSet<String> = list_assets().into_iter().collect();
-    let needed = [
+    let arches = [
-        (ClientArch::LegacyX86, "undionly.kpxe"),
+        ClientArch::LegacyX86,
-        (ClientArch::Ia32Uefi, "snponly-i386.efi"),
+        ClientArch::Ia32Uefi,
-        (ClientArch::X64Uefi, "snponly.efi"),
+        ClientArch::X64Uefi,
-        // ARM32 UEFI deferred — no upstream snponly binary published.
+        // ARM32 UEFI deferred — no upstream binary published in either mode.
-        (ClientArch::Arm64Uefi, "snponly-arm64.efi"),
+        ClientArch::Arm64Uefi,
     ];
-    for (arch, name) in needed {
+    for arch in arches {
-        if have.contains(name) {
+        for mode in [DriverMode::Firmware, DriverMode::Builtin] {
-            tracing::info!(target: "openpxe::ipxe", "bundled iPXE for {}: {}", arch.as_str(), name);
+            let Some(name) = arch.ipxe_bootfile_mode(mode) else {
-        } else {
+                continue;
-            tracing::warn!(
+            };
-                target: "openpxe::ipxe",
+            if have.contains(name) {
-                "MISSING iPXE binary for {}: {} — clients of this arch will not PXE boot",
+                tracing::info!(
-                arch.as_str(), name
+                    target: "openpxe::ipxe",
-            );
+                    "bundled iPXE for {} [{mode:?}]: {name}", arch.as_str()
+                );
+            } else if mode == DriverMode::Firmware {
+                tracing::warn!(
+                    target: "openpxe::ipxe",
+                    "MISSING iPXE binary for {} [{mode:?}]: {name} — clients of this arch will not PXE boot",
+                    arch.as_str()
+                );
+            } else {
+                tracing::info!(
+                    target: "openpxe::ipxe",
+                    "no built-in-driver fallback for {} [{mode:?}]: {name} — auto NIC driver escalation unavailable for this arch",
+                    arch.as_str()
+                );
+            }
         }
     }
 }

crates/iso-store/src/entry.rs

@@ -25,9 +25,11 @@ pub enum BootKind {
         wimboot_url: String,
         files: Vec<(String, String)>,
     },
-    /// Last-resort: SAN-boot the ISO as an emulated CD. Only works for small
+    /// SAN-boot the raw ISO as an emulated CD (iPXE `sanboot`). The emulated
-    /// ISOs (<~1 GiB) and older distros. Kept for completeness, not the
+    /// CD is backed by on-demand HTTP range reads, so ISO size is *not* a
-    /// default.
+    /// constraint — this is the primary path for Windows (v0.5.8) and for any
+    /// El Torito-bootable image we don't special-case: ESXi/VMvisor
+    /// installers, BSDs, firmware/diagnostic tools, custom spins (v0.6.0).
     SanBootIso { iso_url: String },
 }
 

crates/iso-store/src/introspect.rs

@@ -103,7 +103,13 @@ pub fn introspect(path: &Path) -> IntrospectionReport {
     let scan_bytes = 64 * 1024 * 1024;
     let mut buf = vec![0u8; 1024 * 1024];
     let mut read_total = 0usize;
-    let mut haystack = Vec::with_capacity(scan_bytes.min(32 * 1024 * 1024));
+    // Size the haystack to what will actually be read — the scan cap or
+    // the file itself, whichever is smaller — so the fill never reallocs
+    // and a small ISO doesn't reserve the full 64 MiB.
+    let file_len = f.metadata().map_or(usize::MAX, |m| {
+        usize::try_from(m.len()).unwrap_or(usize::MAX)
+    });
+    let mut haystack = Vec::with_capacity(scan_bytes.min(file_len));
     while read_total < scan_bytes {
         let n = f.read(&mut buf).unwrap_or(0);
         if n == 0 {

crates/iso-store/src/nfs_share.rs

@@ -63,8 +63,8 @@ use bytes::Bytes;
 use nfs3_client::tokio::TokioConnector;
 use nfs3_client::Nfs3ConnectionBuilder;
 use nfs3_types::nfs3::{
-    self as nfs3, diropargs3, entry3, filename3, nfs_fh3, GETATTR3args, LOOKUP3args,
+    self as nfs3, diropargs3, entry3, filename3, nfs_fh3, GETATTR3args, LOOKUP3args, Nfs3Result,
-    Nfs3Result, READ3args, READDIR3args,
+    READ3args, READDIR3args,
 };
 use nfs3_types::rpc::{auth_unix, opaque_auth};
 use nfs3_types::xdr_codec::Opaque;
@@ -220,10 +220,7 @@ impl NfsShareManager {
     /// Register an NFS share. Validates, probes connectivity by
     /// performing a real MOUNT3 + READDIR3, and registers the
     /// resulting ISOs with the store.
-    pub async fn add(
+    pub async fn add(&self, req: NfsAddRequest) -> std::result::Result<NfsShare, NfsShareError> {
-        &self,
-        req: NfsAddRequest,
-    ) -> std::result::Result<NfsShare, NfsShareError> {
         let server = normalize_server(&req.server);
         let export = req.export.trim().to_string();
         if server.is_empty() {
@@ -258,7 +255,9 @@ impl NfsShareManager {
         if let Err(e) = self.rescan_inner(&id).await {
             let m = self.get(&id);
             return Err(NfsShareError {
-                error: m.as_ref().and_then(|m| m.last_error.clone())
+                error: m
+                    .as_ref()
+                    .and_then(|m| m.last_error.clone())
                     .unwrap_or_else(|| e.to_string()),
                 stderr: String::new(),
                 hint: m.and_then(|m| m.last_hint),
@@ -333,8 +332,7 @@ impl NfsShareManager {
             return Err(Error::Invalid(format!("invalid filename '{filename}'")));
         }
 
-        let (tx, rx) =
+        let (tx, rx) = tokio::sync::mpsc::channel::<std::io::Result<Bytes>>(STREAM_BUFFER_DEPTH);
-            tokio::sync::mpsc::channel::<std::io::Result<Bytes>>(STREAM_BUFFER_DEPTH);
         let server = share.server.clone();
         let export = share.export.clone();
         let port = share.port;
@@ -358,16 +356,11 @@ impl NfsShareManager {
             if let Err(e) = result {
                 // Best-effort signal of the error to the consumer.
                 // If the receiver has already dropped we just exit.
-                let _ = tx
+                let _ = tx.send(Err(std::io::Error::other(e.to_string()))).await;
-                    .send(Err(std::io::Error::other(e.to_string())))
-                    .await;
             }
         });
 
-        Ok(NfsStream {
+        Ok(NfsStream { rx, _task: task })
-            rx,
-            _task: task,
-        })
     }
 
     // ── internals ─────────────────────────────────────────────────────
@@ -982,8 +975,14 @@ mod tests {
         // the allow-list and the secure/insecure angle.
         let h = hint_for("connect failed: MNT3ERR_ACCES").unwrap();
         let lc = h.to_lowercase();
-        assert!(lc.contains("insecure") || lc.contains("privileged"), "got: {h}");
+        assert!(
-        assert!(lc.contains("allow") || lc.contains("permission"), "got: {h}");
+            lc.contains("insecure") || lc.contains("privileged"),
+            "got: {h}"
+        );
+        assert!(
+            lc.contains("allow") || lc.contains("permission"),
+            "got: {h}"
+        );
     }
 
     #[test]

crates/iso-store/src/smb_share.rs

@@ -231,10 +231,7 @@ impl SmbShareManager {
 
     /// Add or refresh a share. Validates the input, writes a creds
     /// file, probes connectivity, and scans for ISOs.
-    pub async fn add(
+    pub async fn add(&self, req: SmbAddRequest) -> std::result::Result<SmbShare, SmbShareError> {
-        &self,
-        req: SmbAddRequest,
-    ) -> std::result::Result<SmbShare, SmbShareError> {
         let server = normalize_server(&req.server);
         let share = req.share.trim().trim_start_matches('/').to_string();
         if server.is_empty() {
@@ -309,7 +306,9 @@ impl SmbShareManager {
         if let Err(e) = self.rescan_inner(&id).await {
             let m = self.get(&id);
             return Err(SmbShareError {
-                error: m.as_ref().and_then(|m| m.last_error.clone())
+                error: m
+                    .as_ref()
+                    .and_then(|m| m.last_error.clone())
                     .unwrap_or_else(|| e.to_string()),
                 stderr: String::new(),
                 hint: m.and_then(|m| m.last_hint),
@@ -365,11 +364,7 @@ impl SmbShareManager {
     /// throttling concurrent smbclients) would need to await without
     /// changing the call sites.
     #[allow(clippy::unused_async)]
-    pub async fn stream_iso(
+    pub async fn stream_iso(&self, share_id: &str, filename: &str) -> Result<SmbStream> {
-        &self,
-        share_id: &str,
-        filename: &str,
-    ) -> Result<SmbStream> {
         let share = self
             .get(share_id)
             .ok_or_else(|| Error::Invalid(format!("no such SMB share '{share_id}'")))?;
@@ -377,9 +372,7 @@ impl SmbShareManager {
         // share root. smbclient itself accepts only filenames at the
         // share root in our `get` form, but belt-and-suspenders.
         if filename.contains('/') || filename.contains('\\') || filename.contains("..") {
-            return Err(Error::Invalid(format!(
+            return Err(Error::Invalid(format!("invalid filename '{filename}'")));
-                "invalid filename '{filename}'"
-            )));
         }
         let creds = share
             .creds_path
@@ -537,10 +530,7 @@ impl SmbShareManager {
                 } else {
                     String::new()
                 };
-                return Err((
+                return Err((format!("could not exec smbclient: {e}"), stderr));
-                    format!("could not exec smbclient: {e}"),
-                    stderr,
-                ));
             }
         };
         if !output.status.success() {
@@ -751,10 +741,7 @@ fn parse_ls_iso(out: &str) -> Vec<SmbListEntry> {
 
 /// Pre-flight TCP probe to `server:port`. Format matches v0.4.64 NFS
 /// probe so the UI banner reads consistently.
-async fn tcp_probe(
+async fn tcp_probe(server: &str, port: u16) -> std::result::Result<(), (String, String)> {
-    server: &str,
-    port: u16,
-) -> std::result::Result<(), (String, String)> {
     use tokio::net::TcpStream;
     let addr = format!("{server}:{port}");
     match tokio::time::timeout(PROBE_TIMEOUT, TcpStream::connect(&addr)).await {
@@ -931,8 +918,8 @@ mod tests {
         // exec error in `error` plus an empty `stderr`. The
         // SmbShareError constructor's hint_for fallback checks error
         // too, so this pattern needs to translate as well.
-        let h2 = hint_for("could not exec smbclient: No such file or directory (os error 2)")
+        let h2 =
-            .unwrap();
+            hint_for("could not exec smbclient: No such file or directory (os error 2)").unwrap();
         assert!(h2.contains("smbclient"));
     }
 

crates/iso-store/src/store.rs

@@ -342,9 +342,18 @@ impl IsoStore {
     /// SMB sources or when the file is missing.
     pub fn iso_path_for(&self, id: &str) -> Option<PathBuf> {
         let meta = self.get(id)?;
+        self.local_path(&meta)
+    }
+
+    /// Same resolution as [`Self::iso_path_for`], but for a meta the
+    /// caller already holds — skips the second registry lock + deep
+    /// clone, which matters on the per-range-request ISO serving path
+    /// (a sanboot install issues hundreds of those).
+    #[must_use]
+    pub fn local_path(&self, meta: &IsoMeta) -> Option<PathBuf> {
         match &meta.source {
             IsoSource::Local => {
-                let path = self.iso_path(id);
+                let path = self.iso_path(&meta.id);
                 if path.exists() {
                     Some(path)
                 } else {
@@ -639,15 +648,33 @@ fn generate_boot_entries(id: &str, filename: &str, r: &IntrospectionReport) -> V
             }]
         }
         _ => {
-            // Last-resort SAN boot. Won't work for large modern ISOs, but
+            // No Windows-install media and no Linux kernel/initrd. Decide
-            // lets the ISO at least appear in the menu.
+            // whether the ISO is bootable at all (v0.6.0):
-            vec![BootEntry {
+            //   * `el_torito` — it carries a boot catalog, so iPXE sanboots
-                id: format!("{id}-sanboot"),
+            //     the raw image as an emulated CD: BSDs, ESXi/VMvisor
-                title: format!("{title} (SAN boot — may fail for >1GiB ISOs)"),
+            //     installers, firmware tools, custom spins. The emulated CD
-                kind: BootKind::SanBootIso {
+            //     is backed by HTTP range reads, so ISO size is a non-issue
-                    iso_url: format!("iso/{id}.iso"),
+            //     (this is the same path Windows uses since v0.5.8) — hence
-                },
+            //     no more "may fail for >1GiB ISOs" disclaimer.
-            }]
+            //   * `introspect_rev == 0` — a remote-share ISO we couldn't
+            //     introspect (SMB/NFS/SFTP listings don't seek into the ISO).
+            //     Offer sanboot optimistically rather than hide a
+            //     likely-bootable installer.
+            // Otherwise it's a local image we *did* introspect and found to
+            // carry no boot catalog — a data/appliance ISO (e.g. a VMware
+            // vCenter Server Appliance bundle). It genuinely cannot boot, so
+            // we expose no menu entry; the dashboard flags it instead.
+            if r.el_torito || r.introspect_rev == 0 {
+                vec![BootEntry {
+                    id: format!("{id}-sanboot"),
+                    title,
+                    kind: BootKind::SanBootIso {
+                        iso_url: format!("iso/{id}.iso"),
+                    },
+                }]
+            } else {
+                Vec::new()
+            }
         }
     }
 }
@@ -721,6 +748,49 @@ mod tests {
         assert!(!s.contains(" --- "), "stray ---: {s}");
     }
 
+    #[test]
+    fn boot_entries_respect_el_torito_and_source() {
+        use crate::introspect::INTROSPECT_REV;
+
+        // ESXi / VMvisor installer shape: bootable (carries an El Torito
+        // catalog) but not classifiable as Windows or Linux. Must yield a
+        // single sanboot entry so it's selectable + boots via emulated CD.
+        let esxi = IntrospectionReport {
+            family: DistroFamily::Unknown,
+            volume_label: Some("ESXI-7.0U3".into()),
+            el_torito: true,
+            introspect_rev: INTROSPECT_REV,
+            ..Default::default()
+        };
+        let e = generate_boot_entries("esxi", "VMware-VMvisor-Installer-7.0U3n.iso", &esxi);
+        assert_eq!(e.len(), 1, "ESXi should get exactly one boot entry");
+        assert!(matches!(e[0].kind, BootKind::SanBootIso { .. }));
+        // Clean title — no stale ">1GiB may fail" disclaimer.
+        assert!(!e[0].title.contains("may fail"), "title: {}", e[0].title);
+
+        // VCSA / data-appliance shape: locally introspected (rev set), no
+        // boot catalog, not Windows/Linux. Genuinely unbootable → no entry,
+        // so it stays out of the iPXE menu (the dashboard flags it instead).
+        let vcsa = IntrospectionReport {
+            family: DistroFamily::Unknown,
+            el_torito: false,
+            introspect_rev: INTROSPECT_REV,
+            ..Default::default()
+        };
+        assert!(
+            generate_boot_entries("vcsa", "VMware-VCSA-all-8.0.iso", &vcsa).is_empty(),
+            "data/appliance ISO must produce no boot entry"
+        );
+
+        // Remote-share ISO: never introspected (rev 0, no random access over
+        // SMB/NFS/SFTP). Assume bootable and offer sanboot rather than hide a
+        // likely-bootable installer.
+        let remote = IntrospectionReport::default();
+        let r = generate_boot_entries("remote", "unknown-remote.iso", &remote);
+        assert_eq!(r.len(), 1, "remote (uninspected) ISO keeps a sanboot entry");
+        assert!(matches!(r[0].kind, BootKind::SanBootIso { .. }));
+    }
+
     fn fake_meta(id: &str) -> IsoMeta {
         IsoMeta {
             id: id.into(),

crates/openpxe/src/main.rs

@@ -192,6 +192,7 @@ async fn main() -> anyhow::Result<()> {
         hosts: hosts.clone(),
         boot_log: boot_log.clone(),
         branding: branding.clone(),
+        pxe_bg_cache: openpxe_http_api::state::PxeBgCache::default(),
         admin: admin.clone(),
         sessions: sessions.clone(),
         sso: sso.clone(),

crates/tftp/src/server.rs

@@ -7,12 +7,12 @@
 //! `tftpd`/`in.tftpd` works and is why TFTP is awkward behind stateful NAT:
 //! the ephemeral ports must be reachable from the client.
 //!
-//! We only serve files from `openpxe_ipxe_assets::asset_bytes` — that is,
+//! We only serve files from `openpxe_ipxe_assets::asset_slice` — that is,
 //! the bundled iPXE binaries and wimboot. No filesystem is ever opened, so
 //! `../` path traversal attempts simply return ENOENT.
 
 use openpxe_core::{ClientEvent, ClientRegistry};
-use openpxe_ipxe_assets::asset_bytes;
+use openpxe_ipxe_assets::asset_slice;
 use socket2::{Domain, Protocol, Socket, Type};
 use std::net::{IpAddr, SocketAddr};
 use std::sync::Arc;
@@ -21,6 +21,7 @@ use tokio::net::UdpSocket;
 
 // TFTP opcodes.
 const OP_RRQ: u16 = 1;
+const OP_WRQ: u16 = 2;
 const OP_DATA: u16 = 3;
 const OP_ACK: u16 = 4;
 const OP_ERROR: u16 = 5;
@@ -89,16 +90,34 @@ async fn handle_rrq(
     metrics: openpxe_core::Metrics,
 ) -> anyhow::Result<()> {
     let Some(req) = parse_rrq(&packet) else {
+        // Not a well-formed RRQ. A WRQ deserves an explicit refusal —
+        // legacy clients retry a silently-dropped write until they time
+        // out; an ERROR packet fails them fast with a readable reason.
+        if packet.len() >= 2 && u16::from_be_bytes([packet[0], packet[1]]) == OP_WRQ {
+            let sock = bind_udp(bind_ip, 0)?;
+            let _ = send_error(&sock, peer, ERR_ILLEGAL_OP, "writes not supported").await;
+        }
         return Ok(());
     };
     let Request {
-        filename, options, ..
+        filename,
+        mode,
+        options,
     } = req;
 
     // Per-transfer ephemeral socket.
     let sock = bind_udp(bind_ip, 0)?;
 
-    let Some(file_bytes) = asset_bytes(&filename) else {
+    // We serve binary boot artifacts; netascii line-ending translation
+    // would corrupt them. Refuse loudly instead of timing out silently —
+    // matters for legacy clients that default to netascii.
+    if !mode.eq_ignore_ascii_case("octet") {
+        let _ = send_error(&sock, peer, ERR_NOT_DEFINED, "only octet mode is supported").await;
+        tracing::info!(target: "openpxe::tftp", peer=%peer, %mode, "rejected non-octet transfer");
+        return Ok(());
+    }
+
+    let Some(file_bytes) = asset_slice(&filename) else {
         let _ = send_error(&sock, peer, ERR_FILE_NOT_FOUND, "no such file").await;
         tracing::info!(target: "openpxe::tftp", peer=%peer, file=%filename, "404");
         clients.record(
@@ -262,7 +281,6 @@ async fn handle_rrq(
 #[derive(Debug)]
 struct Request {
     filename: String,
-    #[allow(dead_code)]
     mode: String,
     options: Vec<(String, String)>,
 }
@@ -393,17 +411,13 @@ fn bind_udp(bind: IpAddr, port: u16) -> anyhow::Result<UdpSocket> {
     Ok(UdpSocket::from_std(std_sock)?)
 }
 
-#[allow(dead_code)]
+/// Pure-logic mirror of `handle_rrq`'s windowing math, exercised by the
-const _UNUSED: (u16, u16) = (ERR_NOT_DEFINED, ERR_ILLEGAL_OP);
+/// unit tests below. Given a position in the file and the window, return
-
+/// the (block_no, chunk_len) list this window will emit — tested against
-/// Pure-logic helper used by the unit tests below and (in a refactor) by
+/// edge cases (exact-blksize tail, short tail, single-block window,
-/// `handle_rrq`. Given a position in the file and the window, return the
+/// block-number wraparound).
-/// (block_no, chunk_len) list this window will emit. Useful as a sanity
+#[cfg(test)]
-/// check that our windowing math matches the wire behavior the spec
+fn plan_window(
-/// requires — tested against edge cases (exact-blksize tail, short tail,
-/// single-block window).
-#[must_use]
-pub fn plan_window(
     total: usize,
     offset: usize,
     blksize: usize,

crates/webui/src/app.js

@@ -365,7 +365,7 @@
       const settings = status.settings;
       const problems = isos.map(i => ({i, b: bootability(i, settings)})).filter(x => !x.b.ok);
       const problemsBlock = problems.length ? el('div', {class:'card'}, [
-        el('header', {}, [el('h2', {}, 'Images that won\'t boot with current settings')]),
+        el('header', {}, [el('h2', {}, 'Non-bootable images')]),
         el('div', {class:'body'},
           problems.map(({i, b}) => el('div', {class:'row-warn'},
             '⚠ ' + i.filename + ' — ' + b.reason)))

scripts/build-ipxe.sh

@@ -41,15 +41,31 @@ DEST="${1:-$ROOT/assets/ipxe}"
 WORK="$(mktemp -d)"
 trap 'rm -rf "$WORK"' EXIT
 
-# Pinned upstream iPXE. Rolling master is fine functionally, but a pin
+# Pinned upstream iPXE. Rolling master is fine functionally, but a pin keeps
-# keeps builds reproducible and protects against a transient master
+# builds reproducible, protects against a transient master breakage, and —
-# breakage. Bump deliberately.
+# crucially for the Docker image — busting this value invalidates the cached
+# ipxe-build layer so an "update iPXE" release actually recompiles from the
+# new upstream. Bump deliberately to a recent master commit.
+#
+# v0.6.1: ipxe/ipxe master @ 2026-06-09 (newer NIC drivers + EFI fixes;
+# mirrors iVentoy 1.0.35 "Update iPXE").
 IPXE_REPO="https://github.com/ipxe/ipxe.git"
-IPXE_REF="${IPXE_REF:-master}"
+IPXE_REF="${IPXE_REF:-95ffbf4745553e8a207922389929e1943c0237c0}"
 
-echo ">> cloning iPXE ($IPXE_REF)"
+echo ">> fetching iPXE ($IPXE_REF)"
-git clone --depth 1 --branch "$IPXE_REF" "$IPXE_REPO" "$WORK/ipxe" 2>/dev/null \
+# Shallow-fetch the exact ref: works for a full commit SHA (GitHub allows
-  || git clone "$IPXE_REPO" "$WORK/ipxe"
+# reachable-SHA1-in-want) and for branch/tag names. Fall back to a full
+# clone + checkout if the server refuses a direct fetch of this ref.
+git init -q "$WORK/ipxe"
+git -C "$WORK/ipxe" remote add origin "$IPXE_REPO"
+if git -C "$WORK/ipxe" fetch -q --depth 1 origin "$IPXE_REF"; then
+  git -C "$WORK/ipxe" checkout -q FETCH_HEAD
+else
+  echo "   direct fetch failed; falling back to full clone + checkout"
+  rm -rf "$WORK/ipxe"
+  git clone -q "$IPXE_REPO" "$WORK/ipxe"
+  git -C "$WORK/ipxe" checkout -q "$IPXE_REF"
+fi
 SRC="$WORK/ipxe/src"
 
 echo ">> applying OpenPXE config overrides (PNG + framebuffer + console cmd)"

scripts/fetch-ipxe.sh

@@ -29,11 +29,19 @@ mkdir -p "$DEST"
 # Upstream uses arch-scoped subdirectories; we flatten to the names our
 # ClientArch::ipxe_bootfile() expects.
 declare -a MAP=(
+  # DriverMode::Firmware (default) — reuse the firmware UNDI/SNP NIC stack.
   "undionly.kpxe=undionly.kpxe"
   "snponly.efi=x86_64-efi/snponly.efi"
   "snponly-i386.efi=i386-efi/snponly.efi"
   "snponly-arm64.efi=arm64-efi/snponly.efi"
-  "ipxe.efi=x86_64-efi/ipxe.efi"          # fallback with bundled drivers
+  # DriverMode::Builtin (v0.6.1 automatic fallback) — iPXE's own all-drivers
+  # builds, advertised by the DHCP proxy to a MAC whose firmware NIC stack
+  # failed to chainload. (x86_64 ipxe.efi is rebuilt from source with PNG in
+  # build-ipxe.sh and overlaid on top of this fetched baseline.)
+  "ipxe.efi=x86_64-efi/ipxe.efi"
+  "ipxe.pxe=ipxe.pxe"
+  "ipxe-i386.efi=i386-efi/ipxe.efi"
+  "ipxe-arm64.efi=arm64-efi/ipxe.efi"
 )
 
 BASE="https://boot.ipxe.org"