v0.7.4: probe-based introspection — remote shares classify, gparted bug fixed, Storage pagination

Introspection (the headline): detection is now probe-based. Instead of
grepping raw sectors for filename strings, we walk the ISO9660
directory tree and check whether the well-known boot files actually
exist — and the same probes run over NFS READ3 / SFTP seek-reads, so
share-hosted ISOs finally classify instead of registering as Unknown.

iso-store:
- New iso_fs module: the read-only ISO9660 walker (generalized from
  http-api) over an IsoReadAt trait — local files, NFS, SFTP, and the
  in-memory test images all share it. Iterative walk, 4 MiB directory
  cap, strict-mastering trailing-dot normalization (VMLINUZ.;1 now
  matches /vmlinuz), CachingReadAt collapses repeated directory reads
  during the probe pass (~60 → ~6 round-trips per remote ISO).
- introspect.rs rewritten (INTROSPECT_REV 2): PVD label → El Torito →
  /sources/boot.wim probe → verified Linux kernel+initrd probe table →
  local-only 16 MiB UDF-Windows scan → filename-token fallback.
  * Fixes the false-Windows bug: any Linux ISO shipping GRUB/syslinux
    chainload modules contains the literal "bootmgr", so gparted-live
    classified as WindowsPe. Linux probes now run first; the byte scan
    only sees ISOs nothing else claimed. Local ISOs re-probe once on
    startup via the rev bump — no re-upload.
  * Kernel entries are emitted only when kernel+initrd verifiably
    exist (no more guessed paths that 404 at boot). Debian-live /
    d-i netinst / CoreOS shapes classify for the UI but keep their
    working sanboot entries (their boot protocols need args we don't
    render yet; CoreOS additionally needs its embedded ignition).
  * Label + filename vocab extended: rhcos/coreos/openshift/okd,
    gparted/clonezilla/kali/tails, almalinux/rocky, sles, manjaro.
- NFS + SFTP managers: per-ISO IsoReadAt readers (READ3-at-offset with
  short-read looping / seek+read_exact), background introspection pass
  after each scan — entries register instantly with a provisional
  filename-based report (rev 0, optimistic sanboot preserved) and
  upgrade in place as probes land (30s/ISO timeout, failures keep the
  provisional). locate_in_iso() exposes the walker to the HTTP layer.
- remote_cache: introspection results persisted per protocol keyed
  share/path@size and gated on INTROSPECT_REV — container restarts
  re-probe only new/replaced ISOs; upgrades re-probe exactly once.
- SMB: smbclient can't seek, so SMB ISOs get the filename-token family
  (rev stays 0 → sanboot entry + "awaiting introspection" label).
- IsoStore::update_external_introspection swaps in completed reports
  and regenerates boot entries, preserving category/password.

http-api:
- /iso/{id}/{*path} now serves files from inside NFS/SFTP-hosted ISOs
  (remote ISO9660 lookup + ranged share stream) — verified kernel
  entries on remote Linux ISOs are actually bootable, end to end.
- iso_fs.rs deleted in favor of the shared iso-store module.
- full_flow fixtures build real directory trees via the shared
  test-image builder (new iso-store feature) — a label-only blob no
  longer earns a kernel entry, by design.

webui:
- Available images: paged 5 per page with a quiet footer pager
  (Showing X–Y of N · Prev/Next), filter-then-paginate, page resets on
  search input. Fifty images is five clean pages, not a scroll wall.
- Hosts/Queue profile: "Unattended file (in Storage → Advanced)" so
  the picker says where the files live.
- Row badge keys on introspect_rev: probed remote ISOs read like local
  ones; un-probed say "awaiting introspection".

Validation: clippy pedantic clean, fmt clean, 316 workspace tests
green (+17: walker, probe shapes incl. gparted regression + CoreOS,
filename table, cache round-trips), webui syntax-checked.

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

Cargo.lock

@@ -2836,7 +2836,7 @@ checksum = "c08d65885ee38876c4f86fa503fb49d7b507c2b62552df7c70b2fce627e06381"
 
 [[package]]
 name = "openpxe"
-version = "0.7.1"
+version = "0.7.4"
 dependencies = [
  "anyhow",
  "axum",
@@ -2858,7 +2858,7 @@ dependencies = [
 
 [[package]]
 name = "openpxe-core"
-version = "0.7.1"
+version = "0.7.4"
 dependencies = [
  "anyhow",
  "base64",
@@ -2885,7 +2885,7 @@ dependencies = [
 
 [[package]]
 name = "openpxe-dhcp-proxy"
-version = "0.7.1"
+version = "0.7.4"
 dependencies = [
  "anyhow",
  "bytes",
@@ -2902,7 +2902,7 @@ dependencies = [
 
 [[package]]
 name = "openpxe-http-api"
-version = "0.7.1"
+version = "0.7.4"
 dependencies = [
  "anyhow",
  "axum",
@@ -2938,7 +2938,7 @@ dependencies = [
 
 [[package]]
 name = "openpxe-ipxe-assets"
-version = "0.7.1"
+version = "0.7.4"
 dependencies = [
  "openpxe-core",
  "rust-embed",
@@ -2948,7 +2948,7 @@ dependencies = [
 
 [[package]]
 name = "openpxe-iso-store"
-version = "0.7.1"
+version = "0.7.4"
 dependencies = [
  "anyhow",
  "bcrypt",
@@ -2977,7 +2977,7 @@ dependencies = [
 
 [[package]]
 name = "openpxe-tftp"
-version = "0.7.1"
+version = "0.7.4"
 dependencies = [
  "anyhow",
  "bytes",
@@ -2991,7 +2991,7 @@ dependencies = [
 
 [[package]]
 name = "openpxe-webui"
-version = "0.7.1"
+version = "0.7.4"
 
 [[package]]
 name = "p256"

Cargo.toml

@@ -12,7 +12,7 @@ members = [
 ]
 
 [workspace.package]
-version = "0.7.1"
+version = "0.7.4"
 edition = "2021"
 rust-version = "1.95"
 license = "MIT OR Apache-2.0"

crates/http-api/Cargo.toml

@@ -49,6 +49,10 @@ base64.workspace = true
 tokio = { workspace = true, features = ["macros", "rt", "rt-multi-thread", "time"] }
 tower = { workspace = true }
 tempfile = "3.12"
+# v0.7.4: probe-based introspection verifies kernel paths against the
+# real ISO9660 tree, so the full-flow tests synthesize images with the
+# shared test builder instead of label-only blobs.
+openpxe-iso-store = { workspace = true, features = ["test-image"] }
 serde_json = { workspace = true }
 time = { workspace = true }
 # v0.4.61: integration tests need to generate real PNG bytes for the

crates/http-api/src/app.rs

@@ -19,7 +19,6 @@ use crate::ipxe_script::{
     render_entry, render_family_menu, render_local_hdd, render_menu, render_nic_info,
     render_queue_entry, render_shell, render_tools_menu, render_util,
 };
-use crate::iso_fs;
 use crate::log_stream;
 use crate::state::AppState;
 use crate::terminal;
@@ -36,6 +35,7 @@ use openpxe_core::{
     LogoSlot, NotifyConfig, Settings, SsoConfig, ALLOWED_LOGO_MIMES, MAX_LOGO_BYTES,
 };
 use openpxe_ipxe_assets::asset_slice;
+use openpxe_iso_store::iso_fs;
 use openpxe_iso_store::{
     render_template, IsoCategory, IsoMeta, IsoSource, NfsAddRequest, SftpAddRequest, SmbAddRequest,
     SmbState, UnattendedKind, UnattendedMeta,
@@ -1087,28 +1087,105 @@ async fn iso_file(
     State(state): State<AppState>,
     AxumPath((id, path)): AxumPath<(String, String)>,
 ) -> Response {
-    // In-ISO file extraction is only supported for local ISOs — it
-    // needs random-access reads into the ISO9660 directory tree, which
-    // smbclient's whole-file streaming can't do efficiently. SMB-
-    // sourced ISOs use the raw streaming endpoint above instead.
-    let Some(iso_path) = state.iso_store.iso_path_for(&id) else {
+    let Some(meta) = state.iso_store.get(&id) else {
         return (StatusCode::NOT_FOUND, "no such iso").into_response();
     };
-    let p = iso_path.clone();
     let in_path = format!("/{path}");
-    let loc = tokio::task::spawn_blocking(move || iso_fs::lookup(&p, &in_path))
-        .await
-        .ok()
-        .flatten();
-    let Some(loc) = loc else {
-        return (StatusCode::NOT_FOUND, "not found inside iso").into_response();
-    };
-    match stream_byte_range(&iso_path, loc.offset, loc.length).await {
-        Ok(r) => r,
-        Err(e) => (StatusCode::INTERNAL_SERVER_ERROR, format!("{e}")).into_response(),
+    match &meta.source {
+        IsoSource::Local => {
+            let Some(iso_path) = state.iso_store.local_path(&meta) else {
+                return (StatusCode::NOT_FOUND, "no such iso").into_response();
+            };
+            let p = iso_path.clone();
+            let loc = tokio::task::spawn_blocking(move || iso_fs::lookup_local(&p, &in_path))
+                .await
+                .ok()
+                .flatten();
+            let Some(loc) = loc else {
+                return (StatusCode::NOT_FOUND, "not found inside iso").into_response();
+            };
+            match stream_byte_range(&iso_path, loc.offset, loc.length).await {
+                Ok(r) => r,
+                Err(e) => (StatusCode::INTERNAL_SERVER_ERROR, format!("{e}")).into_response(),
+            }
+        }
+        // v0.7.4: remote ISOs serve in-ISO files too — the same ISO9660
+        // walk runs over NFS READ3 / SFTP seek-reads, then the located
+        // byte range streams through the share manager. This is what
+        // makes the verified kernel/initrd boot entries on share-hosted
+        // Linux ISOs actually bootable.
+        IsoSource::Nfs {
+            share_id,
+            relative_path,
+        } => {
+            match state
+                .nfs_shares
+                .locate_in_iso(share_id, relative_path, &in_path)
+                .await
+            {
+                Ok(Some(loc)) => {
+                    match state
+                        .nfs_shares
+                        .stream_iso(share_id, relative_path, loc.offset, Some(loc.length))
+                        .await
+                    {
+                        Ok(stream) => in_iso_stream_response(Body::from_stream(stream), loc.length),
+                        Err(e) => {
+                            (StatusCode::BAD_GATEWAY, format!("nfs stream: {e}")).into_response()
+                        }
+                    }
+                }
+                Ok(None) => (StatusCode::NOT_FOUND, "not found inside iso").into_response(),
+                Err(e) => (StatusCode::BAD_GATEWAY, format!("nfs lookup: {e}")).into_response(),
+            }
+        }
+        IsoSource::Sftp {
+            share_id,
+            relative_path,
+        } => {
+            match state
+                .sftp_shares
+                .locate_in_iso(share_id, relative_path, &in_path)
+                .await
+            {
+                Ok(Some(loc)) => {
+                    match state
+                        .sftp_shares
+                        .stream_iso(share_id, relative_path, loc.offset, Some(loc.length))
+                        .await
+                    {
+                        Ok(stream) => in_iso_stream_response(Body::from_stream(stream), loc.length),
+                        Err(e) => {
+                            (StatusCode::BAD_GATEWAY, format!("sftp stream: {e}")).into_response()
+                        }
+                    }
+                }
+                Ok(None) => (StatusCode::NOT_FOUND, "not found inside iso").into_response(),
+                Err(e) => (StatusCode::BAD_GATEWAY, format!("sftp lookup: {e}")).into_response(),
+            }
+        }
+        // smbclient streams sequentially — no seeks, no ISO9660 walk.
+        // SMB ISOs never emit kernel entries, so nothing requests this.
+        IsoSource::Smb { .. } => (
+            StatusCode::NOT_FOUND,
+            "in-ISO files are not available for SMB-sourced ISOs",
+        )
+            .into_response(),
     }
 }
 
+/// 200 response wrapping an in-ISO byte-range stream from a share
+/// manager. Content-Length is the located file's length — the stream is
+/// already bounded to exactly that range.
+fn in_iso_stream_response(body: Body, length: u64) -> Response {
+    Response::builder()
+        .status(StatusCode::OK)
+        .header(header::CONTENT_TYPE, "application/octet-stream")
+        .header(header::CONTENT_LENGTH, length)
+        .body(body)
+        .unwrap()
+}
+
 async fn stream_file_range(
     path: &std::path::Path,
     range: Option<&HeaderValue>,
@@ -2744,6 +2821,7 @@ async fn api_network(State(state): State<AppState>) -> Json<serde_json::Value> {
             .strip_prefix("http://")
             .unwrap_or(&state.public_base_url),
         "nic_name":   state.nic_name,
+        "nic_link":   state.nic_link,
         "subnet_mask": state.subnet_mask,
         "gateway":     state.gateway,
         "dns_server":  state.settings.snapshot().dns_server,

crates/http-api/src/iso_fs.rs

@@ -1,135 +0,0 @@
-//! Minimal read-only ISO9660 lookup. Given an uploaded ISO file and an
-//! in-ISO path (e.g. `/casper/vmlinuz`), locate the file and return a
-//! `(start_byte, length_bytes)` pair so the HTTP handler can stream just
-//! that range from the on-disk ISO without full extraction.
-//!
-//! We only implement what we need: the Primary Volume Descriptor and Rock
-//! Ridge / Joliet extensions are ignored. Paths are matched case-insensitive
-//! against plain ISO9660 filenames (uppercase, `;1` version suffix stripped).
-//! This is sufficient for the kernel/initrd and wimboot files we serve;
-//! if a requested path isn't found, the handler returns 404 and the user
-//! can still download the whole ISO via `/iso/<id>.iso`.
-
-use std::io::{Read, Seek, SeekFrom};
-use std::path::Path;
-
-const SECTOR: u64 = 2048;
-
-#[derive(Debug, Clone)]
-pub struct FileLocation {
-    pub offset: u64,
-    pub length: u64,
-}
-
-/// Look up `in_iso_path` (leading slash optional, case-insensitive) in the
-/// ISO at `iso_path`. Returns None on any parsing or IO failure.
-pub fn lookup(iso_path: &Path, in_iso_path: &str) -> Option<FileLocation> {
-    let mut f = std::fs::File::open(iso_path).ok()?;
-    let root = read_root_directory(&mut f)?;
-    let components: Vec<&str> = in_iso_path
-        .trim_start_matches('/')
-        .split('/')
-        .filter(|c| !c.is_empty())
-        .collect();
-    if components.is_empty() {
-        return None;
-    }
-    walk(&mut f, root.offset, root.length, &components)
-}
-
-fn read_root_directory(f: &mut std::fs::File) -> Option<FileLocation> {
-    // Primary Volume Descriptor at LBA 16.
-    let mut pvd = [0u8; 2048];
-    f.seek(SeekFrom::Start(16 * SECTOR)).ok()?;
-    f.read_exact(&mut pvd).ok()?;
-    if pvd[0] != 0x01 || &pvd[1..6] != b"CD001" {
-        return None;
-    }
-    // Root directory record is at offset 156, length 34.
-    let rec = &pvd[156..156 + 34];
-    let (offset, length) = parse_dir_record_ext(rec)?;
-    Some(FileLocation {
-        offset: offset * SECTOR,
-        length,
-    })
-}
-
-/// Walk components down the directory tree starting at `dir_offset`.
-fn walk(
-    f: &mut std::fs::File,
-    dir_offset: u64,
-    dir_len: u64,
-    components: &[&str],
-) -> Option<FileLocation> {
-    let mut dir = vec![0u8; dir_len as usize];
-    f.seek(SeekFrom::Start(dir_offset)).ok()?;
-    f.read_exact(&mut dir).ok()?;
-
-    let target = components[0];
-    let rest = &components[1..];
-    let mut i = 0;
-    while i < dir.len() {
-        let len = dir[i] as usize;
-        if len == 0 {
-            // Padding to sector boundary.
-            let next = (i / SECTOR as usize + 1) * SECTOR as usize;
-            if next <= i {
-                break;
-            }
-            i = next;
-            continue;
-        }
-        if i + len > dir.len() {
-            break;
-        }
-        let rec = &dir[i..i + len];
-        let name = dir_record_name(rec);
-        let is_dir = (rec.get(25).copied().unwrap_or(0) & 0x02) != 0;
-        // Skip "." (0x00) and ".." (0x01) pseudo-entries.
-        let is_pseudo = matches!(rec.get(32).copied(), Some(1))
-            && rec.get(33).copied() == Some(0x00)
-            || matches!(rec.get(32).copied(), Some(1)) && rec.get(33).copied() == Some(0x01);
-        if !is_pseudo && name.eq_ignore_ascii_case(target) {
-            let (child_off, child_len) = parse_dir_record_ext(rec)?;
-            if rest.is_empty() && !is_dir {
-                return Some(FileLocation {
-                    offset: child_off * SECTOR,
-                    length: child_len,
-                });
-            } else if !rest.is_empty() && is_dir {
-                return walk(f, child_off * SECTOR, child_len, rest);
-            }
-        }
-        i += len;
-    }
-    None
-}
-
-/// Extract (extent LBA, data length in bytes) from a directory record.
-/// Layout per ISO9660: bytes 2..10 extent LBA (LE+BE duplicate), 10..18
-/// data length (LE+BE duplicate). We trust the little-endian copy.
-fn parse_dir_record_ext(rec: &[u8]) -> Option<(u64, u64)> {
-    if rec.len() < 34 {
-        return None;
-    }
-    let lba = u32::from_le_bytes(rec[2..6].try_into().ok()?) as u64;
-    let len = u32::from_le_bytes(rec[10..14].try_into().ok()?) as u64;
-    Some((lba, len))
-}
-
-/// Extract the identifier from a directory record, stripping ISO9660's
-/// `;1` version suffix.
-fn dir_record_name(rec: &[u8]) -> String {
-    let name_len = *rec.get(32).unwrap_or(&0) as usize;
-    if name_len == 0 || rec.len() < 33 + name_len {
-        return String::new();
-    }
-    let raw = &rec[33..33 + name_len];
-    let s = String::from_utf8_lossy(raw).to_string();
-    // Strip `;N` version suffix.
-    if let Some(i) = s.rfind(';') {
-        s[..i].to_string()
-    } else {
-        s
-    }
-}

crates/http-api/src/lib.rs

@@ -9,8 +9,9 @@
 //!     and Linux kernel/initrd, without having to
 //!     re-extract on every request)
 //!
-//! The `<id>/<path>` handler uses a read-only ISO9660 shim (see `iso_fs`)
-//! that lseeks into the ISO on disk — so we never keep extracted copies.
+//! The `<id>/<path>` handler uses the read-only ISO9660 walker from
+//! `openpxe_iso_store::iso_fs` — seeking into the image wherever it
+//! lives (local disk, NFS, SFTP), so we never keep extracted copies.
 #![forbid(unsafe_code)]
 
 pub mod app;
@@ -18,7 +19,6 @@ pub mod auth;
 pub mod error;
 pub mod grub_script;
 pub mod ipxe_script;
-pub mod iso_fs;
 pub mod log_stream;
 pub mod notify;
 pub mod saml_routes;

crates/http-api/src/state.rs

@@ -116,6 +116,10 @@ pub struct AppState {
     /// `enp1s0`). Surfaced read-only on the Network tab. Empty if the
     /// interface couldn't be identified.
     pub nic_name: String,
+    /// v0.7.2: physical link summary for that NIC (operstate, speed,
+    /// duplex, port MAC) — read from sysfs at startup; empty where
+    /// unavailable. Helps confirm which port answers PXE.
+    pub nic_link: String,
     /// Subnet mask of the public interface in dotted-quad form.
     pub subnet_mask: String,
     /// Default gateway IPv4 address.

crates/http-api/tests/full_flow.rs

@@ -18,23 +18,16 @@ use openpxe_iso_store::{IsoStore, NfsShareManager, SftpShareManager, SmbShareMan
 use tempfile::tempdir;
 use tower::ServiceExt;
 
-/// Build a tiny valid ISO9660 blob with volume label "ALPINE-TEST" so
-/// introspection identifies it as Alpine.
+/// Build a tiny Alpine-shaped ISO9660 image: volume label "ALPINE-TEST"
+/// plus the real `/boot/vmlinuz-lts` + `/boot/initramfs-lts` tree.
+/// v0.7.4's probe-based introspection verifies those paths exist before
+/// emitting a kernel boot entry — a label-only blob no longer counts.
 fn fake_alpine_iso() -> Vec<u8> {
-    let mut buf = vec![0u8; 32 * 2048];
-    let off = 16 * 2048;
-    buf[off] = 0x01;
-    buf[off + 1..off + 6].copy_from_slice(b"CD001");
-    buf[off + 6] = 0x01;
-    let label = b"ALPINE-TEST".to_vec();
-    let mut padded = label.clone();
-    padded.resize(32, b' ');
-    buf[off + 40..off + 40 + 32].copy_from_slice(&padded);
-    let term = 17 * 2048;
-    buf[term] = 0xFF;
-    buf[term + 1..term + 6].copy_from_slice(b"CD001");
-    buf[term + 6] = 0x01;
-    buf
+    openpxe_iso_store::iso_fs::testiso::TestIsoBuilder::new("ALPINE-TEST")
+        .el_torito(true)
+        .file("/boot/vmlinuz-lts", b"fake-kernel-bytes")
+        .file("/boot/initramfs-lts", b"fake-initramfs-bytes")
+        .build()
 }
 
 fn multipart_iso_body(filename: &str, bytes: &[u8]) -> (String, Vec<u8>) {
@@ -135,6 +128,7 @@ async fn build_state() -> (AppState, tempfile::TempDir) {
         started_at: time::OffsetDateTime::now_utc(),
         public_base_url: "http://127.0.0.1".into(),
         nic_name: "lo".into(),
+        nic_link: String::new(),
         subnet_mask: "255.0.0.0".into(),
         gateway: "127.0.0.1".into(),
     };
@@ -1247,9 +1241,10 @@ async fn chunked_upload_writes_progressively_and_finishes_iso() {
                 .method("POST")
                 .uri("/api/uploads")
                 .header("content-type", "application/json")
-                .body(Body::from(
-                    r#"{"filename":"chunked-alpine.iso","size_bytes":65536}"#,
-                ))
+                .body(Body::from(format!(
+                    r#"{{"filename":"chunked-alpine.iso","size_bytes":{}}}"#,
+                    iso.len()
+                )))
                 .unwrap(),
         )
         .await

crates/iso-store/Cargo.toml

@@ -49,3 +49,11 @@ futures = { workspace = true }
 
 [dev-dependencies]
 tempfile = "3.12"
+
+[features]
+# v0.7.4: exposes the in-memory ISO9660 test-image builder
+# (`iso_fs::testiso`) to other crates' integration tests, so http-api's
+# full-flow tests can synthesize ISOs with real directory trees — the
+# probe-based introspection no longer classifies label-only blobs.
+# Never enabled in production builds.
+test-image = []

crates/iso-store/src/introspect.rs

@@ -1,16 +1,33 @@
 //! ISO introspection — identify the distro family and locate kernel/initrd.
 //!
-//! We avoid a full ISO9660/Joliet/Rock-Ridge parser by reading a small number
-//! of well-known files via `isoinfo` (from cdrtools/genisoimage) when it's on
-//! the path. As a pure-Rust fallback we do a crude scan: read the volume
-//! descriptor at offset 0x8000 to grab the volume label, and grep for known
-//! filenames by scanning raw sectors — good enough to tell Debian from RHEL
-//! most of the time, without shelling out.
+//! v0.7.4 rewrite: detection is **probe-based**. Instead of grepping raw
+//! sectors for filename strings (which false-positived — any Linux ISO
+//! shipping GRUB/syslinux chainload modules contains the literal
+//! "bootmgr", so gparted-live classified as Windows), we walk the
+//! ISO9660 directory tree via [`crate::iso_fs`] and check whether the
+//! well-known boot files actually exist. The same probes run over local
+//! files and remote NFS/SFTP shares — remote ISOs finally classify
+//! instead of registering as `Unknown`.
 //!
-//! The returned `IntrospectionReport` is what `BootEntry`s get generated from.
+//! Layered, first-decisive-answer-wins:
+//!   1. PVD volume label → family hint.
+//!   2. El Torito boot-catalog presence (the "bootable at all" signal).
+//!   3. `/sources/boot.wim` directory probe → Windows install media.
+//!   4. Linux probe table → verified kernel+initrd paths. A probe match
+//!      both classifies the family and (for the families whose boot
+//!      arguments we render) yields kernel paths that are *known to
+//!      exist* — no more guessed paths that 404 at boot.
+//!   5. Bulk byte scan for UDF Windows markers — local images only
+//!      (modern Windows ISOs hide their tree from ISO9660; remote scans
+//!      skip this so a share rescan doesn't stream 16 MiB per ISO).
+//!   6. Filename tokens — the last-resort hint, and the only signal
+//!      available for SMB shares (smbclient cannot seek).
+//!
+//! The returned `IntrospectionReport` is what `BootEntry`s get generated
+//! from.
 
+use crate::iso_fs::{self, CachingReadAt, FileReadAt, IsoReadAt, SECTOR};
 use serde::{Deserialize, Serialize};
-use std::io::{Read, Seek, SeekFrom};
 use std::path::Path;
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, Default)]
@@ -30,18 +47,26 @@ pub enum DistroFamily {
 /// re-classify already-uploaded ISOs. On startup the store re-runs
 /// `introspect` on any *local* ISO whose persisted report predates this
 /// revision (see `IsoStore::load_from_disk`), so an upgrade fixes stale
-/// metadata — e.g. a Windows 11 ISO tagged `Unknown` by an older binary —
-/// without the operator having to delete and re-upload it.
+/// metadata without the operator having to delete and re-upload.
 ///
 /// rev 1 (v0.5.9): added El Torito boot-catalog detection + broadened
-/// Windows (UDF/UTF-16) detection becomes retroactive.
-pub const INTROSPECT_REV: u32 = 1;
+/// Windows (UDF/UTF-16) detection.
+/// rev 2 (v0.7.4): probe-based detection. Fixes Linux live ISOs that
+/// classified as Windows via the raw "bootmgr" byte grep, verifies
+/// kernel/initrd paths exist before emitting them, and adds the Debian
+/// live / netinst / CoreOS shapes. Remote (NFS/SFTP) introspection
+/// caches key off this rev too, so the cache self-invalidates.
+pub const INTROSPECT_REV: u32 = 2;
 
 #[derive(Debug, Clone, Default, Serialize, Deserialize)]
 pub struct IntrospectionReport {
     pub family: DistroFamily,
     pub volume_label: Option<String>,
-    /// Kernel path inside the ISO (e.g. `/casper/vmlinuz`, `/isolinux/vmlinuz`).
+    /// Kernel path inside the ISO (e.g. `/casper/vmlinuz`). v0.7.4: only
+    /// set when the path was verified to exist *and* the family's boot
+    /// arguments are known-good for direct kernel boot; families we can
+    /// only classify (Debian live, CoreOS live) leave it `None` so the
+    /// entry generator falls back to sanboot instead of a broken boot.
     pub kernel_path: Option<String>,
     /// Initrd path(s) inside the ISO. May be multiple for multi-initrd setups.
     pub initrd_paths: Vec<String>,
@@ -55,124 +80,242 @@ pub struct IntrospectionReport {
     /// appliance bundle) has no boot catalog and reports `false`. v0.5.9.
     #[serde(default)]
     pub el_torito: bool,
-    /// Revision of the introspection logic that produced this report. Old
-    /// `meta.json` files without the field deserialize as 0, which is
-    /// below [`INTROSPECT_REV`], triggering a one-time re-introspect on
-    /// the next startup. v0.5.9.
+    /// Revision of the introspection logic that produced this report.
+    /// `0` means "never introspected" (pre-v0.5.9 metadata, or a remote
+    /// ISO whose probe hasn't run / can't run) — the entry generator
+    /// treats those optimistically (sanboot) and the UI labels them.
     #[serde(default)]
     pub introspect_rev: u32,
 }
 
-/// Probe an ISO file on disk. Never fails — on unrecoverable IO error we log
-/// and return an `Unknown` family so the uploader still sees a record.
+/// One row of the Linux detection table.
+///
+/// `emit_kernel` distinguishes "we can boot this directly" from "we can
+/// only classify it". Families marked `false` have boot protocols our
+/// cmdline renderer doesn't speak yet (Debian-live `boot=live fetch=`,
+/// d-i netinst, CoreOS `coreos.live.rootfs_url=`) — for those the probe
+/// sets the family for the UI/menu but leaves `kernel_path` unset so the
+/// ISO keeps its (working) sanboot entry instead of gaining a broken
+/// kernel one. Strictly fewer broken boots than guessing.
+struct LinuxProbe {
+    family: DistroFamily,
+    kernel: &'static str,
+    initrd_candidates: &'static [&'static str],
+    emit_kernel: bool,
+}
+
+const LINUX_PROBES: &[LinuxProbe] = &[
+    // Ubuntu and friends (casper) — the classic direct-boot shape.
+    LinuxProbe {
+        family: DistroFamily::DebianUbuntu,
+        kernel: "/casper/vmlinuz",
+        initrd_candidates: &["/casper/initrd", "/casper/initrd.lz", "/casper/initrd.gz"],
+        emit_kernel: true,
+    },
+    // Debian-live derivatives: gparted-live, Clonezilla, Kali live, tails.
+    // Classification only — live-boot needs `boot=live fetch=<squashfs>`
+    // which we don't render yet; sanboot of these images works today.
+    LinuxProbe {
+        family: DistroFamily::DebianUbuntu,
+        kernel: "/live/vmlinuz",
+        initrd_candidates: &["/live/initrd.img", "/live/initrd"],
+        emit_kernel: false,
+    },
+    // Debian installer (netinst/DVD). Classification only for the same
+    // reason — d-i sanboots fine.
+    LinuxProbe {
+        family: DistroFamily::DebianUbuntu,
+        kernel: "/install.amd/vmlinuz",
+        initrd_candidates: &["/install.amd/initrd.gz"],
+        emit_kernel: false,
+    },
+    // Anaconda family: RHEL, CentOS, Alma, Rocky, Fedora — and their
+    // many derivatives (Cisco ISE, Nagios appliances, …). The CoreOS
+    // variant of this shape is special-cased after the table.
+    LinuxProbe {
+        family: DistroFamily::RhelFedora,
+        kernel: "/images/pxeboot/vmlinuz",
+        initrd_candidates: &["/images/pxeboot/initrd.img"],
+        emit_kernel: true,
+    },
+    LinuxProbe {
+        family: DistroFamily::OpenSuse,
+        kernel: "/boot/x86_64/loader/linux",
+        initrd_candidates: &["/boot/x86_64/loader/initrd"],
+        emit_kernel: true,
+    },
+    LinuxProbe {
+        family: DistroFamily::Arch,
+        kernel: "/arch/boot/x86_64/vmlinuz-linux",
+        initrd_candidates: &["/arch/boot/x86_64/initramfs-linux.img"],
+        emit_kernel: true,
+    },
+    LinuxProbe {
+        family: DistroFamily::Alpine,
+        kernel: "/boot/vmlinuz-lts",
+        initrd_candidates: &["/boot/initramfs-lts"],
+        emit_kernel: true,
+    },
+];
+
+/// CoreOS-style live images (RHCOS, FCOS, OpenShift agent ISOs) carry
+/// the anaconda pxeboot layout *plus* a rootfs image. Direct kernel boot
+/// of those requires `coreos.live.rootfs_url=` (and for agent ISOs, the
+/// ignition config embedded in the ISO device) — neither of which a
+/// plain `inst.repo=` cmdline provides. Their sanboot path works, so
+/// they classify as RHEL-family but keep the sanboot entry.
+const COREOS_ROOTFS: &str = "/images/pxeboot/rootfs.img";
+
+/// Probe an ISO file on disk. Never fails — on unrecoverable IO error we
+/// log and return an `Unknown` family so the uploader still sees a record.
 pub fn introspect(path: &Path) -> IntrospectionReport {
+    let filename = path
+        .file_name()
+        .map(|s| s.to_string_lossy().into_owned())
+        .unwrap_or_default();
+    let Ok(f) = std::fs::File::open(path) else {
+        tracing::warn!(target: "openpxe::iso", "cannot open ISO for introspection: {}", path.display());
+        return IntrospectionReport {
+            introspect_rev: INTROSPECT_REV,
+            ..Default::default()
+        };
+    };
+    let len = f.metadata().map_or(0, |m| m.len());
+    // `FileReadAt` completes every read inline (no real awaits), so this
+    // light-weight block_on never parks; callers already run us on the
+    // blocking pool.
+    futures::executor::block_on(introspect_reader(
+        &mut FileReadAt::new(f),
+        len,
+        &filename,
+        true,
+    ))
+}
+
+/// The detection core, generic over any random-access source. `total_len`
+/// is the image size (every caller knows it — file metadata locally, the
+/// share listing remotely) and bounds the bulk scan, since `IsoReadAt`
+/// reads are exact-or-error. `filename` feeds the last-resort token
+/// heuristics; `allow_bulk_scan` gates the 16 MiB UDF-Windows byte scan
+/// (local files only — remote shares would stream that much per ISO per
+/// rescan).
+pub async fn introspect_reader<R: IsoReadAt + Send>(
+    r: &mut R,
+    total_len: u64,
+    filename: &str,
+    allow_bulk_scan: bool,
+) -> IntrospectionReport {
     let mut report = IntrospectionReport {
         introspect_rev: INTROSPECT_REV,
         ..Default::default()
     };
 
-    let Ok(mut f) = std::fs::File::open(path) else {
-        tracing::warn!(target: "openpxe::iso", "cannot open ISO for introspection: {}", path.display());
-        return report;
-    };
-
-    // ISO9660 Primary Volume Descriptor at LBA 16 (offset 0x8000), 2048 bytes.
-    // Bytes 40..72 are the Volume Identifier (space-padded, d-characters).
-    let mut pvd = [0u8; 2048];
-    if f.seek(SeekFrom::Start(0x8000)).is_ok() && f.read_exact(&mut pvd).is_ok() {
-        // Byte 0 must be 0x01 (primary descriptor), bytes 1..6 = "CD001".
+    // ISO9660 Primary Volume Descriptor at LBA 16. Bytes 40..72 are the
+    // volume identifier (space-padded).
+    if let Ok(pvd) = r.read_at(16 * SECTOR, 2048).await {
         if pvd[0] == 0x01 && &pvd[1..6] == b"CD001" {
-            let label_raw = &pvd[40..72];
-            let label = String::from_utf8_lossy(label_raw).trim().to_string();
+            let label = String::from_utf8_lossy(&pvd[40..72]).trim().to_string();
             if !label.is_empty() {
-                report.volume_label = Some(label.clone());
                 report.family = family_from_label(&label);
+                report.volume_label = Some(label);
             }
         }
     }
 
-    // Does the ISO have an El Torito boot catalog? This is what decides
-    // whether an ISO we *can't* otherwise classify is bootable at all —
-    // a bootable ISO sanboots; a data/appliance ISO (no catalog) can't.
-    report.el_torito = detect_el_torito(&mut f);
+    report.el_torito = detect_el_torito(r).await;
 
-    // Cheap content scan: read the first ~64 MiB, look for signature filenames.
-    // This is enough to identify `sources/boot.wim` (Windows) and common
-    // kernel/initrd paths for the major Linux distros.
-    let _ = f.seek(SeekFrom::Start(0));
-    let scan_bytes = 64 * 1024 * 1024;
-    let mut buf = vec![0u8; 1024 * 1024];
-    let mut read_total = 0usize;
-    // 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 {
-            break;
-        }
-        haystack.extend_from_slice(&buf[..n]);
-        read_total += n;
-    }
-
-    // `sources/boot.wim` is the definitive Windows-install-media marker
-    // when the ISO exposes ASCII (ISO9660/Joliet) names. `contains_ascii`
-    // is case-insensitive, so one form covers BOOT.WIM / boot.wim and the
-    // backslash variant.
-    if contains_ascii(&haystack, b"sources/boot.wim")
-        || contains_ascii(&haystack, b"sources\\boot.wim")
+    // Directory-tree probes. The caching wrapper collapses the repeated
+    // root/subdirectory reads the probe table would otherwise issue —
+    // over NFS/SFTP that's the difference between ~6 and ~60 round-trips.
     {
-        report.has_boot_wim = true;
-        report.family = DistroFamily::WindowsPe;
-    }
+        let mut cr = CachingReadAt::new(r);
 
-    // v0.5.8: broaden Windows detection. Modern Windows 10/11 ISOs are
-    // UDF — filenames are stored as UTF-16 (so the ASCII scan above misses
-    // them) and the volume label is a cryptic Microsoft string (so
-    // `family_from_label` misses it too). Booting is via HTTP sanboot of
-    // the raw ISO (no boot.wim extraction), so we only need the *family*.
-    // Catch the common cases: well-known Windows markers in either ASCII
-    // or UTF-16LE within the first 16 MiB, plus a filename hint.
-    if report.family == DistroFamily::Unknown {
-        let head = &haystack[..haystack.len().min(16 * 1024 * 1024)];
-        let ascii_markers: [&[u8]; 4] = [
-            b"bootmgr",
-            b"sources/install.wim",
-            b"sources/install.esd",
-            b"efi/microsoft",
-        ];
-        let utf16_markers = ["bootmgr", "boot.wim", "install.wim", "microsoft"];
-        let looks_windows = ascii_markers.iter().any(|m| contains_ascii(head, m))
-            || utf16_markers.iter().any(|m| contains_utf16le_ci(head, m))
-            || filename_looks_windows(path);
-        if looks_windows {
+        if iso_fs::exists(&mut cr, "/sources/boot.wim").await {
+            report.has_boot_wim = true;
             report.family = DistroFamily::WindowsPe;
+        } else {
+            for probe in LINUX_PROBES {
+                if !iso_fs::exists(&mut cr, probe.kernel).await {
+                    continue;
+                }
+                let mut initrd = None;
+                for cand in probe.initrd_candidates {
+                    if iso_fs::exists(&mut cr, cand).await {
+                        initrd = Some((*cand).to_string());
+                        break;
+                    }
+                }
+                let Some(initrd) = initrd else { continue };
+                // Content beats label: a rebadged derivative (volume
+                // label "ISE-3.2") with the anaconda layout is
+                // RHEL-family no matter what the label says.
+                report.family = probe.family;
+                let coreos = probe.family == DistroFamily::RhelFedora
+                    && iso_fs::exists(&mut cr, COREOS_ROOTFS).await;
+                if probe.emit_kernel && !coreos {
+                    report.kernel_path = Some(probe.kernel.to_string());
+                    report.initrd_paths = vec![initrd];
+                }
+                break;
+            }
         }
     }
 
-    // Best-effort kernel/initrd path guess from family. These paths are what
-    // distro ISOs conventionally ship at — we don't verify extraction here;
-    // that happens in the store after introspection.
-    let (k, i) = guess_kernel_initrd(report.family);
-    report.kernel_path = k.map(str::to_string);
-    report.initrd_paths = i.iter().map(std::string::ToString::to_string).collect();
+    // Modern Windows 10/11 ISOs are UDF — their tree is invisible to the
+    // ISO9660 walk and the volume label is a cryptic Microsoft string.
+    // Scan the first 16 MiB for well-known markers, ASCII and UTF-16LE.
+    // Runs after the Linux probes so a Linux ISO that *contains* the
+    // string "bootmgr" (GRUB/syslinux chainload modules do) has already
+    // classified and never reaches this — that ordering is the v0.7.4
+    // gparted-misdetection fix.
+    if report.family == DistroFamily::Unknown && allow_bulk_scan {
+        if let Some(win) = bulk_windows_scan(r, total_len).await {
+            report.family = DistroFamily::WindowsPe;
+            report.has_boot_wim = win;
+        }
+    }
+
+    // Last resort: filename tokens. The only signal for SMB-sourced ISOs
+    // and renamed/UDF images that defeated everything above.
+    if report.family == DistroFamily::Unknown {
+        report.family = family_from_filename(filename);
+    }
 
     report
 }
 
+/// Provisional report for a remote ISO that hasn't been (or can't be)
+/// content-probed yet: family from the filename, `introspect_rev` left
+/// at 0 so the entry generator keeps the optimistic sanboot entry and
+/// the UI shows it as awaiting introspection. Used by all three share
+/// managers at registration; NFS/SFTP upgrade it in the background.
+#[must_use]
+pub fn provisional_report(filename: &str) -> IntrospectionReport {
+    IntrospectionReport {
+        family: family_from_filename(filename),
+        ..Default::default()
+    }
+}
+
 fn family_from_label(label: &str) -> DistroFamily {
     let l = label.to_ascii_lowercase();
-    if l.contains("ubuntu") || l.contains("debian") || l.contains("mint") {
+    if l.contains("ubuntu")
+        || l.contains("debian")
+        || l.contains("mint")
+        || l.contains("kali")
+        || l.contains("gparted")
+        || l.contains("clonezilla")
+    {
         DistroFamily::DebianUbuntu
     } else if l.contains("rhel")
         || l.contains("centos")
         || l.contains("fedora")
         || l.contains("rocky")
         || l.contains("alma")
+        || l.contains("rhcos")
+        || l.contains("coreos")
+        || l.contains("openshift")
+        || l.contains("okd")
     {
         DistroFamily::RhelFedora
     } else if l.contains("suse") || l.contains("opensuse") {
@@ -188,23 +331,91 @@ fn family_from_label(label: &str) -> DistroFamily {
     }
 }
 
-fn guess_kernel_initrd(family: DistroFamily) -> (Option<&'static str>, Vec<&'static str>) {
-    match family {
-        DistroFamily::DebianUbuntu => (Some("/casper/vmlinuz"), vec!["/casper/initrd"]),
-        DistroFamily::RhelFedora => (
-            Some("/images/pxeboot/vmlinuz"),
-            vec!["/images/pxeboot/initrd.img"],
-        ),
-        DistroFamily::OpenSuse => (
-            Some("/boot/x86_64/loader/linux"),
-            vec!["/boot/x86_64/loader/initrd"],
-        ),
-        DistroFamily::Arch => (
-            Some("/arch/boot/x86_64/vmlinuz-linux"),
-            vec!["/arch/boot/x86_64/initramfs-linux.img"],
-        ),
-        DistroFamily::Alpine => (Some("/boot/vmlinuz-lts"), vec!["/boot/initramfs-lts"]),
-        DistroFamily::WindowsPe | DistroFamily::Unknown => (None, Vec::new()),
+/// Filename token heuristic — `AlmaLinux-9.5-x86_64-dvd.iso` says what
+/// it is even when we can't read a byte of it. Tokens are the filename
+/// split on every non-alphanumeric character, so "almalinux", "rhel",
+/// "win11" match without "search" tripping the "arch" token.
+pub fn family_from_filename(filename: &str) -> DistroFamily {
+    if filename_looks_windows(filename) {
+        return DistroFamily::WindowsPe;
+    }
+    let lower = filename.to_ascii_lowercase();
+    let tokens: Vec<&str> = lower
+        .split(|c: char| !c.is_ascii_alphanumeric())
+        .filter(|t| !t.is_empty())
+        .collect();
+    let has = |t: &str| tokens.contains(&t);
+    if has("ubuntu")
+        || has("debian")
+        || has("mint")
+        || has("kali")
+        || has("gparted")
+        || has("clonezilla")
+        || has("tails")
+    {
+        DistroFamily::DebianUbuntu
+    } else if has("rhel")
+        || has("centos")
+        || has("almalinux")
+        || has("alma")
+        || has("rocky")
+        || has("rockylinux")
+        || has("fedora")
+        || has("rhcos")
+        || has("coreos")
+        || has("openshift")
+        || has("okd")
+    {
+        DistroFamily::RhelFedora
+    } else if has("opensuse") || has("suse") || has("sles") {
+        DistroFamily::OpenSuse
+    } else if has("arch") || has("archlinux") || has("manjaro") {
+        DistroFamily::Arch
+    } else if has("alpine") {
+        DistroFamily::Alpine
+    } else {
+        DistroFamily::Unknown
+    }
+}
+
+/// Scan the first 16 MiB (or the whole image when smaller) for Windows
+/// markers. Returns `Some(has_boot_wim)` on a hit, `None` when nothing
+/// Windows-shaped is found.
+async fn bulk_windows_scan<R: IsoReadAt + Send>(r: &mut R, total_len: u64) -> Option<bool> {
+    const SCAN_BYTES: u64 = 16 * 1024 * 1024;
+    const CHUNK: u64 = 1024 * 1024;
+    let budget = SCAN_BYTES.min(total_len);
+    let mut haystack = Vec::with_capacity(usize::try_from(budget).unwrap_or(0));
+    let mut offset = 0u64;
+    while offset < budget {
+        // Reads are exact-or-error, so clamp the final chunk to what the
+        // image actually has — netboot.xyz is 2.3 MB, not 16.
+        let want = u32::try_from(CHUNK.min(budget - offset)).unwrap_or(u32::MAX);
+        let Ok(chunk) = r.read_at(offset, want).await else {
+            break; // read error: scan what we have
+        };
+        offset += chunk.len() as u64;
+        haystack.extend_from_slice(&chunk);
+    }
+    if haystack.is_empty() {
+        return None;
+    }
+    let boot_wim = contains_ascii(&haystack, b"sources/boot.wim")
+        || contains_ascii(&haystack, b"sources\\boot.wim")
+        || contains_utf16le_ci(&haystack, "boot.wim");
+    if boot_wim {
+        return Some(true);
+    }
+    let ascii_markers: [&[u8]; 3] = [b"bootmgr", b"sources/install.wim", b"sources/install.esd"];
+    let utf16_markers = ["bootmgr", "install.wim", "microsoft"];
+    let hit = ascii_markers.iter().any(|m| contains_ascii(&haystack, m))
+        || utf16_markers
+            .iter()
+            .any(|m| contains_utf16le_ci(&haystack, m));
+    if hit {
+        Some(false)
+    } else {
+        None
     }
 }
 
@@ -237,14 +448,10 @@ fn contains_utf16le_ci(haystack: &[u8], ascii: &str) -> bool {
 
 /// Filename heuristic: a stock Windows ISO almost always carries an obvious
 /// token in its name (e.g. `..._windows_11_...`, `Win10`, `winserver`).
-/// Used only as a last-resort family hint when the content scan and volume
-/// label are inconclusive. v0.5.8.
-fn filename_looks_windows(path: &Path) -> bool {
-    let name = path
-        .file_name()
-        .and_then(|s| s.to_str())
-        .unwrap_or("")
-        .to_ascii_lowercase();
+/// v0.7.4: takes the bare filename instead of a `Path` so the same check
+/// runs against remote share listings.
+fn filename_looks_windows(filename: &str) -> bool {
+    let name = filename.to_ascii_lowercase();
     const TOKENS: [&str; 6] = [
         "windows",
         "winpe",
@@ -263,19 +470,18 @@ const EL_TORITO_ID: &[u8] = b"EL TORITO SPECIFICATION";
 /// Detect an El Torito boot catalog — the marker that an ISO is bootable
 /// by BIOS/UEFI firmware (and thus by iPXE `sanboot`).
 ///
-/// The ISO9660 Volume Descriptor Set starts at LBA 16 (offset 0x8000) and
-/// runs one 2048-byte descriptor per sector until a Set Terminator
-/// (type 0xFF). A Boot Record descriptor (type 0x00) whose 32-byte boot
-/// system identifier reads "EL TORITO SPECIFICATION" means the image
-/// declares an El Torito boot catalog. We only confirm its presence — we
-/// don't parse the catalog (sanboot/the firmware does that). The walk is
-/// capped so a malformed/huge image can't spin us. v0.5.9.
-fn detect_el_torito(f: &mut std::fs::File) -> bool {
-    let mut vd = [0u8; 2048];
+/// The ISO9660 Volume Descriptor Set starts at LBA 16 and runs one
+/// 2048-byte descriptor per sector until a Set Terminator (type 0xFF).
+/// A Boot Record descriptor (type 0x00) whose 32-byte boot system
+/// identifier reads "EL TORITO SPECIFICATION" means the image declares a
+/// boot catalog. We only confirm its presence — we don't parse the
+/// catalog (sanboot/the firmware does that). The walk is capped so a
+/// malformed image can't spin us. v0.5.9; reader-generic since v0.7.4.
+async fn detect_el_torito<R: IsoReadAt + Send>(r: &mut R) -> bool {
     for lba in 16u64..32 {
-        if f.seek(SeekFrom::Start(lba * 2048)).is_err() || f.read_exact(&mut vd).is_err() {
+        let Ok(vd) = r.read_at(lba * SECTOR, 2048).await else {
             return false;
-        }
+        };
         // Every descriptor in the set carries the "CD001" magic; once it's
         // missing we've walked off the end of a valid set.
         if &vd[1..6] != b"CD001" {
@@ -297,6 +503,12 @@ fn detect_el_torito(f: &mut std::fs::File) -> bool {
 #[cfg(test)]
 mod tests {
     use super::*;
+    use crate::iso_fs::testiso::{MemReadAt, TestIsoBuilder};
+
+    fn introspect_mem(img: Vec<u8>, filename: &str, bulk: bool) -> IntrospectionReport {
+        let len = img.len() as u64;
+        futures::executor::block_on(introspect_reader(&mut MemReadAt(img), len, filename, bulk))
+    }
 
     #[test]
     fn label_matching() {
@@ -313,13 +525,158 @@ mod tests {
             DistroFamily::OpenSuse
         );
         assert_eq!(family_from_label("ARCH_202604"), DistroFamily::Arch);
+        assert_eq!(
+            family_from_label("GParted-live"),
+            DistroFamily::DebianUbuntu
+        );
+        assert_eq!(family_from_label("rhcos-417"), DistroFamily::RhelFedora);
         assert_eq!(family_from_label("weird-custom"), DistroFamily::Unknown);
     }
 
+    #[test]
+    fn gparted_shape_is_not_windows() {
+        // The v0.7.4 regression test: a Debian-live image whose payload
+        // contains the literal string "bootmgr" (as GRUB/syslinux
+        // chainload modules do). The old byte-grep classified this as
+        // WindowsPe; the probe order must classify Debian first.
+        let img = TestIsoBuilder::new("GParted-live")
+            .el_torito(true)
+            .file("/live/vmlinuz", b"KERNEL")
+            .file("/live/initrd.img", b"INITRD")
+            .file("/boot/grub/chain.mod", b"xxx bootmgr xxx")
+            .build();
+        let r = introspect_mem(img, "gparted-live-1.8.1-3-amd64.iso", true);
+        assert_eq!(r.family, DistroFamily::DebianUbuntu);
+        // Classification only — live-boot args aren't rendered yet, so no
+        // kernel entry; sanboot (via el_torito) keeps working.
+        assert!(r.kernel_path.is_none());
+        assert!(r.el_torito);
+        assert_eq!(r.introspect_rev, INTROSPECT_REV);
+    }
+
+    #[test]
+    fn casper_shape_verifies_kernel_and_initrd() {
+        let img = TestIsoBuilder::new("Ubuntu-Server 24.04.1 LTS amd64")
+            .el_torito(true)
+            .file("/casper/vmlinuz", b"K")
+            .file("/casper/initrd", b"I")
+            .build();
+        let r = introspect_mem(img, "ubuntu-24.04.1-live-server-amd64.iso", true);
+        assert_eq!(r.family, DistroFamily::DebianUbuntu);
+        assert_eq!(r.kernel_path.as_deref(), Some("/casper/vmlinuz"));
+        assert_eq!(r.initrd_paths, vec!["/casper/initrd".to_string()]);
+    }
+
+    #[test]
+    fn anaconda_shape_emits_verified_paths() {
+        let img = TestIsoBuilder::new("AlmaLinux-9-5-x86_64-dvd")
+            .el_torito(true)
+            .file("/images/pxeboot/vmlinuz", b"K")
+            .file("/images/pxeboot/initrd.img", b"I")
+            .build();
+        let r = introspect_mem(img, "AlmaLinux-9.5-x86_64-dvd.iso", true);
+        assert_eq!(r.family, DistroFamily::RhelFedora);
+        assert_eq!(r.kernel_path.as_deref(), Some("/images/pxeboot/vmlinuz"));
+    }
+
+    #[test]
+    fn coreos_shape_classifies_but_keeps_sanboot() {
+        // RHCOS / OpenShift agent ISOs: anaconda layout + rootfs.img.
+        // Direct kernel boot needs coreos.live.rootfs_url (and agent
+        // ISOs their embedded ignition), so kernel_path must stay None.
+        let img = TestIsoBuilder::new("rhcos-417.94.202501")
+            .el_torito(true)
+            .file("/images/pxeboot/vmlinuz", b"K")
+            .file("/images/pxeboot/initrd.img", b"I")
+            .file("/images/pxeboot/rootfs.img", b"R")
+            .build();
+        let r = introspect_mem(img, "rhcos-live.x86_64.iso", true);
+        assert_eq!(r.family, DistroFamily::RhelFedora);
+        assert!(
+            r.kernel_path.is_none(),
+            "CoreOS must not get a kernel entry"
+        );
+        assert!(r.el_torito);
+    }
+
+    #[test]
+    fn boot_wim_probe_classifies_windows() {
+        let img = TestIsoBuilder::new("CCCOMA_X64FRE_EN-US_DV9")
+            .el_torito(true)
+            .file("/sources/boot.wim", b"MSWIMMSWIM")
+            .build();
+        let r = introspect_mem(img, "whatever.iso", false);
+        assert_eq!(r.family, DistroFamily::WindowsPe);
+        assert!(r.has_boot_wim);
+    }
+
+    #[test]
+    fn label_only_linux_without_verified_kernel_gets_no_kernel_path() {
+        // Label says RHEL but the tree has no pxeboot files — the old
+        // code guessed `/images/pxeboot/vmlinuz` and emitted an entry
+        // that 404'd at boot. Now: family yes, kernel paths no.
+        let img = TestIsoBuilder::new("RHEL-9-5-CUSTOM")
+            .el_torito(true)
+            .file("/readme.txt", b"hi")
+            .build();
+        let r = introspect_mem(img, "rhel-custom.iso", true);
+        assert_eq!(r.family, DistroFamily::RhelFedora);
+        assert!(r.kernel_path.is_none());
+        assert!(r.initrd_paths.is_empty());
+    }
+
+    #[test]
+    fn remote_skips_bulk_scan_but_filename_still_hints() {
+        // No ISO9660 signatures at all (e.g. pure-UDF image read over a
+        // share), bulk scan off: filename is the only signal.
+        let img = vec![0u8; 64 * 1024];
+        let r = introspect_mem(img.clone(), "Win11_24H2_English_x64.iso", false);
+        assert_eq!(r.family, DistroFamily::WindowsPe);
+        let r2 = introspect_mem(img, "mystery.iso", false);
+        assert_eq!(r2.family, DistroFamily::Unknown);
+    }
+
+    #[test]
+    fn filename_family_table() {
+        use DistroFamily::*;
+        let cases = [
+            ("AlmaLinux-9.5-x86_64-dvd.iso", RhelFedora),
+            ("CentOS-Stream-10-latest-x86_64-dvd1.iso", RhelFedora),
+            ("rhel-9.0-x86_64-boot.iso", RhelFedora),
+            ("rhcos-live.x86_64.iso", RhelFedora),
+            ("openshift-4-21-9.agent.x86_64.iso", RhelFedora),
+            ("ubuntu-24.04-desktop.iso", DebianUbuntu),
+            ("gparted-live-1.8.1-3-amd64.iso", DebianUbuntu),
+            ("archlinux-2026.05.01-x86_64.iso", Arch),
+            ("arch-2026.05.01.iso", Arch),
+            ("alpine-standard-3.21.0-x86_64.iso", Alpine),
+            ("openSUSE-Leap-15.6-DVD-x86_64.iso", OpenSuse),
+            ("en-us_windows_11_iot_enterprise.iso", WindowsPe),
+            ("Win10_22H2_English_x64.iso", WindowsPe),
+            ("netboot.xyz.iso", Unknown),
+            ("ise-3.2.0.542a.SPA.x86_64.iso", Unknown),
+            ("Macrium_5860_v2.iso", Unknown),
+            // "search" must not trip the "arch" token.
+            ("research-data.iso", Unknown),
+        ];
+        for (name, want) in cases {
+            assert_eq!(family_from_filename(name), want, "{name}");
+        }
+    }
+
+    #[test]
+    fn provisional_report_keeps_rev_zero() {
+        let r = provisional_report("rhel-9.0-x86_64-dvd.iso");
+        assert_eq!(r.family, DistroFamily::RhelFedora);
+        assert_eq!(
+            r.introspect_rev, 0,
+            "provisional must keep optimistic sanboot"
+        );
+        assert!(!r.el_torito);
+    }
+
     #[test]
     fn utf16le_marker_matches_case_insensitively() {
-        // "boot.wim" encoded UTF-16LE, mixed case — UDF stores Windows
-        // filenames this way, which the ASCII scan can't see.
         let s = "BOOT.WIM";
         let utf16: Vec<u8> = s.bytes().flat_map(|b| [b, 0]).collect();
         let mut hay = vec![0u8; 8];
@@ -328,59 +685,41 @@ mod tests {
         assert!(contains_utf16le_ci(&hay, "boot.wim"));
         assert!(contains_utf16le_ci(&hay, "Boot.Wim"));
         assert!(!contains_utf16le_ci(&hay, "install.wim"));
-        // An ASCII (not UTF-16) occurrence must NOT match the UTF-16 scan.
         assert!(!contains_utf16le_ci(b"boot.wim plain ascii", "boot.wim"));
     }
 
     #[test]
-    fn el_torito_boot_catalog_detected() {
-        let dir = tempfile::tempdir().unwrap();
-        // Helper: stamp a 2048-byte descriptor at `lba` with type + magic.
-        let stamp = |img: &mut [u8], lba: usize, ty: u8| {
-            let off = lba * 2048;
-            img[off] = ty;
-            img[off + 1..off + 6].copy_from_slice(b"CD001");
-        };
-
-        // Bootable image: PVD @16, El Torito Boot Record @17, terminator @18.
-        let mut boot = vec![0u8; 2048 * 19];
-        stamp(&mut boot, 16, 0x01);
-        stamp(&mut boot, 17, 0x00);
-        boot[17 * 2048 + 7..17 * 2048 + 7 + EL_TORITO_ID.len()].copy_from_slice(EL_TORITO_ID);
-        stamp(&mut boot, 18, 0xFF);
-        let bp = dir.path().join("boot.iso");
-        std::fs::write(&bp, &boot).unwrap();
-        let mut f = std::fs::File::open(&bp).unwrap();
-        assert!(
-            detect_el_torito(&mut f),
-            "El Torito boot record should match"
-        );
-
-        // Data/appliance image: PVD @16, terminator @17, no boot record.
-        let mut data = vec![0u8; 2048 * 18];
-        stamp(&mut data, 16, 0x01);
-        stamp(&mut data, 17, 0xFF);
-        let dp = dir.path().join("data.iso");
-        std::fs::write(&dp, &data).unwrap();
-        let mut f2 = std::fs::File::open(&dp).unwrap();
-        assert!(!detect_el_torito(&mut f2), "data ISO has no boot catalog");
+    fn el_torito_detected_through_reader() {
+        let with = TestIsoBuilder::new("BOOTABLE").el_torito(true).build();
+        let without = TestIsoBuilder::new("DATA").build();
+        assert!(futures::executor::block_on(detect_el_torito(
+            &mut MemReadAt(with)
+        )));
+        assert!(!futures::executor::block_on(detect_el_torito(
+            &mut MemReadAt(without)
+        )));
     }
 
     #[test]
     fn filename_hint_catches_windows_isos() {
-        use std::path::Path;
-        assert!(filename_looks_windows(Path::new(
+        assert!(filename_looks_windows(
             "en-us_windows_11_iot_enterprise_ltsc_2024_x64_dvd.iso"
-        )));
-        assert!(filename_looks_windows(Path::new(
-            "Win10_22H2_English_x64.iso"
-        )));
-        assert!(filename_looks_windows(Path::new("winserver2022.iso")));
-        assert!(!filename_looks_windows(Path::new(
-            "ubuntu-24.04-desktop.iso"
-        )));
-        assert!(!filename_looks_windows(Path::new(
-            "Rocky-9.4-x86_64-dvd.iso"
-        )));
+        ));
+        assert!(filename_looks_windows("Win10_22H2_English_x64.iso"));
+        assert!(filename_looks_windows("winserver2022.iso"));
+        assert!(!filename_looks_windows("ubuntu-24.04-desktop.iso"));
+        assert!(!filename_looks_windows("Rocky-9.4-x86_64-dvd.iso"));
+    }
+
+    #[test]
+    fn bulk_scan_catches_udf_windows_markers() {
+        // A blob with no ISO9660 tree but a UTF-16 "install.wim" — the
+        // UDF Windows shape after every probe missed.
+        let mut img = vec![0u8; 256 * 1024];
+        let marker: Vec<u8> = "install.wim".bytes().flat_map(|b| [b, 0]).collect();
+        img[100_000..100_000 + marker.len()].copy_from_slice(&marker);
+        let r = introspect_mem(img, "renamed.iso", true);
+        assert_eq!(r.family, DistroFamily::WindowsPe);
+        assert!(!r.has_boot_wim);
     }
 }

crates/iso-store/src/iso_fs.rs

@@ -0,0 +1,568 @@
+//! Read-only ISO9660 lookup over any random-access byte source.
+//!
+//! v0.7.4: generalized from the http-api crate's local-file-only walker so
+//! the same directory walk drives three consumers:
+//!
+//!   1. `iso_file` HTTP serving — locate `/casper/vmlinuz` inside a local
+//!      *or remote* (NFS/SFTP) ISO and stream just that byte range.
+//!   2. Introspection — probe for well-known kernel/initrd/boot.wim paths
+//!      instead of grepping raw sectors for filename strings (which
+//!      false-positived: any Linux ISO shipping GRUB/syslinux chainload
+//!      modules contains the literal "bootmgr" and used to classify as
+//!      Windows).
+//!   3. Remote introspection — the same probes over an NFSv3 READ3 /
+//!      SFTP seek-read connection, which is what finally classifies
+//!      share-sourced ISOs instead of registering them all as `Unknown`.
+//!
+//! We parse only the Primary Volume Descriptor namespace. Joliet and Rock
+//! Ridge are deliberately ignored — matching is case-insensitive against
+//! plain ISO9660 identifiers (`;1` version suffix and the trailing dot of
+//! extension-less strict-mastered names stripped), which is how the local
+//! serving path has always behaved in production.
+
+use std::collections::HashMap;
+use std::future::Future;
+use std::io::{Read, Seek, SeekFrom};
+use std::path::Path;
+
+pub const SECTOR: u64 = 2048;
+
+/// Upper bound on a single directory extent we'll buffer. Real distro ISO
+/// directories are a handful of KiB; the cap keeps a malformed or hostile
+/// image from asking us to allocate gigabytes.
+const MAX_DIR_BYTES: u64 = 4 * 1024 * 1024;
+
+/// Byte range of one file inside the ISO image.
+#[derive(Debug, Clone)]
+pub struct FileLocation {
+    pub offset: u64,
+    pub length: u64,
+}
+
+/// Random-access reads into an ISO image. Implemented by a local
+/// `std::fs::File`, the NFS and SFTP share readers, and the in-memory
+/// test image.
+///
+/// The contract is `read_exact`-like: the returned buffer is exactly
+/// `len` bytes or the call errors. The future must be `Send` because
+/// remote introspection runs inside spawned tokio tasks.
+pub trait IsoReadAt {
+    fn read_at(
+        &mut self,
+        offset: u64,
+        len: u32,
+    ) -> impl Future<Output = std::io::Result<Vec<u8>>> + Send;
+}
+
+/// Local-file reader. The reads are synchronous inside an async fn —
+/// callers run it either on the blocking pool (introspection at upload)
+/// or through [`lookup_local`]'s `block_on`, never on a hot runtime
+/// worker with real awaits pending.
+pub struct FileReadAt(std::fs::File);
+
+impl FileReadAt {
+    #[must_use]
+    pub fn new(f: std::fs::File) -> Self {
+        Self(f)
+    }
+}
+
+impl IsoReadAt for FileReadAt {
+    async fn read_at(&mut self, offset: u64, len: u32) -> std::io::Result<Vec<u8>> {
+        self.0.seek(SeekFrom::Start(offset))?;
+        let mut buf = vec![0u8; len as usize];
+        self.0.read_exact(&mut buf)?;
+        Ok(buf)
+    }
+}
+
+/// Exact-key read cache for the probe phase of introspection. The probe
+/// table looks up ~20 paths and every one of them re-reads the root
+/// directory (and usually one shared subdirectory); over NFS/SFTP that
+/// would be 20 identical round-trips. Directory reads repeat with the
+/// exact same `(offset, len)`, so a plain map keyed on the pair hits
+/// every time. Large data reads bypass the cache.
+pub struct CachingReadAt<'a, R: IsoReadAt + Send> {
+    inner: &'a mut R,
+    cache: HashMap<(u64, u32), Vec<u8>>,
+}
+
+/// Don't cache reads bigger than this (file payloads, bulk scans).
+const CACHE_MAX_READ: u32 = 256 * 1024;
+/// Bound the cache so a pathological image can't grow it unbounded.
+const CACHE_MAX_ENTRIES: usize = 256;
+
+impl<'a, R: IsoReadAt + Send> CachingReadAt<'a, R> {
+    pub fn new(inner: &'a mut R) -> Self {
+        Self {
+            inner,
+            cache: HashMap::new(),
+        }
+    }
+}
+
+impl<R: IsoReadAt + Send> IsoReadAt for CachingReadAt<'_, R> {
+    async fn read_at(&mut self, offset: u64, len: u32) -> std::io::Result<Vec<u8>> {
+        let key = (offset, len);
+        if let Some(hit) = self.cache.get(&key) {
+            return Ok(hit.clone());
+        }
+        let buf = self.inner.read_at(offset, len).await?;
+        if len <= CACHE_MAX_READ && self.cache.len() < CACHE_MAX_ENTRIES {
+            self.cache.insert(key, buf.clone());
+        }
+        Ok(buf)
+    }
+}
+
+/// Look up `in_iso_path` (leading slash optional, case-insensitive) in
+/// the image behind `r`. Returns `None` on any parsing or IO failure —
+/// "not found" and "couldn't read" are the same answer to a prober.
+pub async fn lookup<R: IsoReadAt + Send>(r: &mut R, in_iso_path: &str) -> Option<FileLocation> {
+    let pvd = r.read_at(16 * SECTOR, 2048).await.ok()?;
+    if pvd[0] != 0x01 || &pvd[1..6] != b"CD001" {
+        return None;
+    }
+    // Root directory record at PVD offset 156, 34 bytes.
+    let (mut lba, mut len) = parse_dir_record_ext(&pvd[156..156 + 34])?;
+
+    let components: Vec<&str> = in_iso_path
+        .trim_start_matches('/')
+        .split('/')
+        .filter(|c| !c.is_empty())
+        .collect();
+    if components.is_empty() {
+        return None;
+    }
+
+    // The original walk was tail-recursive; iterate instead so the future
+    // stays a plain (non-boxed) state machine.
+    for (idx, comp) in components.iter().enumerate() {
+        if len == 0 || len > MAX_DIR_BYTES {
+            return None;
+        }
+        let dir = r.read_at(lba * SECTOR, len as u32).await.ok()?;
+        let hit = scan_dir(&dir, comp)?;
+        let last = idx + 1 == components.len();
+        match (last, hit.is_dir) {
+            (true, false) => {
+                return Some(FileLocation {
+                    offset: hit.lba * SECTOR,
+                    length: hit.len,
+                })
+            }
+            (false, true) => {
+                lba = hit.lba;
+                len = hit.len;
+            }
+            _ => return None,
+        }
+    }
+    None
+}
+
+/// Convenience probe: does `in_iso_path` exist as a file?
+pub async fn exists<R: IsoReadAt + Send>(r: &mut R, in_iso_path: &str) -> bool {
+    lookup(r, in_iso_path).await.is_some()
+}
+
+/// Synchronous wrapper for local files — the shape the HTTP handler's
+/// `spawn_blocking` call site wants. `block_on` is safe here because
+/// `FileReadAt`'s reads never actually await (they complete inline), so
+/// the executor never parks.
+#[must_use]
+pub fn lookup_local(iso_path: &Path, in_iso_path: &str) -> Option<FileLocation> {
+    let f = std::fs::File::open(iso_path).ok()?;
+    futures::executor::block_on(lookup(&mut FileReadAt::new(f), in_iso_path))
+}
+
+struct DirHit {
+    lba: u64,
+    len: u64,
+    is_dir: bool,
+}
+
+/// Scan one directory extent for an identifier. Pure function over the
+/// buffered extent — all protocol/IO concerns live in the caller.
+fn scan_dir(dir: &[u8], target: &str) -> Option<DirHit> {
+    let mut i = 0;
+    while i < dir.len() {
+        let len = dir[i] as usize;
+        if len == 0 {
+            // Records never span sectors; a zero length byte means the
+            // rest of this sector is padding. Hop to the next one.
+            let next = (i / SECTOR as usize + 1) * SECTOR as usize;
+            if next <= i {
+                break;
+            }
+            i = next;
+            continue;
+        }
+        if i + len > dir.len() {
+            break;
+        }
+        let rec = &dir[i..i + len];
+        let name = dir_record_name(rec);
+        let is_dir = (rec.get(25).copied().unwrap_or(0) & 0x02) != 0;
+        // Skip "." (0x00) and ".." (0x01) pseudo-entries.
+        let is_pseudo =
+            rec.get(32).copied() == Some(1) && matches!(rec.get(33).copied(), Some(0x00 | 0x01));
+        if !is_pseudo && name.eq_ignore_ascii_case(target) {
+            let (lba, dlen) = parse_dir_record_ext(rec)?;
+            return Some(DirHit {
+                lba,
+                len: dlen,
+                is_dir,
+            });
+        }
+        i += len;
+    }
+    None
+}
+
+/// Extract (extent LBA, data length in bytes) from a directory record.
+/// Layout per ISO9660: bytes 2..10 extent LBA (LE+BE duplicate), 10..18
+/// data length (LE+BE duplicate). We trust the little-endian copy.
+fn parse_dir_record_ext(rec: &[u8]) -> Option<(u64, u64)> {
+    if rec.len() < 34 {
+        return None;
+    }
+    let lba = u64::from(u32::from_le_bytes(rec[2..6].try_into().ok()?));
+    let len = u64::from(u32::from_le_bytes(rec[10..14].try_into().ok()?));
+    Some((lba, len))
+}
+
+/// Extract the identifier from a directory record, normalizing ISO9660
+/// quirks: the `;N` version suffix and the trailing dot that strict
+/// mastering appends to extension-less names (`VMLINUZ.;1`). Without the
+/// dot strip, level-1 images' kernels never matched `/casper/vmlinuz`.
+fn dir_record_name(rec: &[u8]) -> String {
+    let name_len = *rec.get(32).unwrap_or(&0) as usize;
+    if name_len == 0 || rec.len() < 33 + name_len {
+        return String::new();
+    }
+    let raw = &rec[33..33 + name_len];
+    let s = String::from_utf8_lossy(raw);
+    let s = s.rfind(';').map_or_else(|| s.as_ref(), |i| &s[..i]);
+    s.strip_suffix('.').unwrap_or(s).to_string()
+}
+
+// ── test support ─────────────────────────────────────────────────────
+//
+// A tiny ISO9660 image builder used by this module's tests, the
+// introspection tests, and (behind the `test-image` feature) other
+// crates' integration tests. Lays out: PVD @ LBA 16, optional El Torito
+// boot record @ 17, set terminator @ 18, directories from LBA 20, file
+// data after. Only what `lookup`/introspection read is populated.
+
+#[cfg(any(test, feature = "test-image"))]
+#[doc(hidden)]
+pub mod testiso {
+    use super::SECTOR;
+    use std::collections::BTreeMap;
+
+    #[derive(Default)]
+    struct Node {
+        children: BTreeMap<String, Node>,
+        content: Option<Vec<u8>>,
+    }
+
+    pub struct TestIsoBuilder {
+        root: Node,
+        volume_label: String,
+        el_torito: bool,
+    }
+
+    impl TestIsoBuilder {
+        pub fn new(volume_label: &str) -> Self {
+            Self {
+                root: Node::default(),
+                volume_label: volume_label.to_string(),
+                el_torito: false,
+            }
+        }
+
+        #[must_use]
+        pub fn el_torito(mut self, on: bool) -> Self {
+            self.el_torito = on;
+            self
+        }
+
+        /// Add a file at `path` (e.g. "/casper/vmlinuz") with `content`.
+        #[must_use]
+        pub fn file(mut self, path: &str, content: &[u8]) -> Self {
+            let mut node = &mut self.root;
+            let comps: Vec<&str> = path
+                .trim_start_matches('/')
+                .split('/')
+                .filter(|c| !c.is_empty())
+                .collect();
+            for (i, c) in comps.iter().enumerate() {
+                node = node.children.entry((*c).to_string()).or_default();
+                if i + 1 == comps.len() {
+                    node.content = Some(content.to_vec());
+                }
+            }
+            self
+        }
+
+        pub fn build(self) -> Vec<u8> {
+            // Pass 1: allocate extents. Directories first (1 sector each),
+            // then file contents.
+            let mut next_lba: u64 = 20;
+            let mut dirs: Vec<(*const Node, u64)> = Vec::new();
+            fn alloc_dirs(n: &Node, next: &mut u64, out: &mut Vec<(*const Node, u64)>) {
+                out.push((std::ptr::from_ref(n), *next));
+                *next += 1;
+                for child in n.children.values() {
+                    if child.content.is_none() {
+                        alloc_dirs(child, next, out);
+                    }
+                }
+            }
+            alloc_dirs(&self.root, &mut next_lba, &mut dirs);
+            let lba_of = |n: &Node| -> u64 {
+                dirs.iter()
+                    .find(|(p, _)| std::ptr::eq(*p, n))
+                    .map(|(_, l)| *l)
+                    .expect("dir allocated")
+            };
+            let mut file_lbas: Vec<(*const Node, u64, usize)> = Vec::new();
+            fn alloc_files(n: &Node, next: &mut u64, out: &mut Vec<(*const Node, u64, usize)>) {
+                for child in n.children.values() {
+                    if let Some(c) = &child.content {
+                        out.push((std::ptr::from_ref(child), *next, c.len()));
+                        *next += c.len().div_ceil(SECTOR as usize).max(1) as u64;
+                    } else {
+                        alloc_files(child, next, out);
+                    }
+                }
+            }
+            alloc_files(&self.root, &mut next_lba, &mut file_lbas);
+            let file_lba_of = |n: &Node| -> u64 {
+                file_lbas
+                    .iter()
+                    .find(|(p, _, _)| std::ptr::eq(*p, n))
+                    .map(|(_, l, _)| *l)
+                    .expect("file allocated")
+            };
+
+            let total = next_lba as usize * SECTOR as usize;
+            let mut img = vec![0u8; total];
+
+            // Directory record encoder.
+            fn record(name_bytes: &[u8], lba: u64, len: u64, is_dir: bool) -> Vec<u8> {
+                let mut rec_len = 33 + name_bytes.len();
+                if rec_len % 2 == 1 {
+                    rec_len += 1; // pad to even
+                }
+                let rec_len = rec_len.max(34);
+                let mut r = vec![0u8; rec_len];
+                r[0] = rec_len as u8;
+                r[2..6].copy_from_slice(&(lba as u32).to_le_bytes());
+                r[6..10].copy_from_slice(&(lba as u32).to_be_bytes());
+                r[10..14].copy_from_slice(&(len as u32).to_le_bytes());
+                r[14..18].copy_from_slice(&(len as u32).to_be_bytes());
+                if is_dir {
+                    r[25] = 0x02;
+                }
+                r[32] = name_bytes.len() as u8;
+                r[33..33 + name_bytes.len()].copy_from_slice(name_bytes);
+                r
+            }
+
+            // Pass 2: write each directory extent.
+            fn write_dir(
+                img: &mut [u8],
+                n: &Node,
+                self_lba: u64,
+                parent_lba: u64,
+                lba_of: &dyn Fn(&Node) -> u64,
+                file_lba_of: &dyn Fn(&Node) -> u64,
+            ) {
+                let base = self_lba as usize * SECTOR as usize;
+                let mut off = 0usize;
+                let mut put = |rec: Vec<u8>, off: &mut usize| {
+                    img[base + *off..base + *off + rec.len()].copy_from_slice(&rec);
+                    *off += rec.len();
+                };
+                put(record(&[0x00], self_lba, SECTOR, true), &mut off);
+                put(record(&[0x01], parent_lba, SECTOR, true), &mut off);
+                for (name, child) in &n.children {
+                    if let Some(c) = &child.content {
+                        // Files get the ISO9660 uppercase `;1` treatment so
+                        // the case-insensitive + version-strip matching is
+                        // what the tests actually exercise.
+                        let stored = format!("{};1", name.to_ascii_uppercase());
+                        put(
+                            record(stored.as_bytes(), file_lba_of(child), c.len() as u64, false),
+                            &mut off,
+                        );
+                    } else {
+                        let stored = name.to_ascii_uppercase();
+                        put(
+                            record(stored.as_bytes(), lba_of(child), SECTOR, true),
+                            &mut off,
+                        );
+                    }
+                }
+                for (name, child) in &n.children {
+                    if child.content.is_none() {
+                        write_dir(img, child, lba_of(child), self_lba, lba_of, file_lba_of);
+                    } else if let Some(c) = &child.content {
+                        let b = file_lba_of(child) as usize * SECTOR as usize;
+                        img[b..b + c.len()].copy_from_slice(c);
+                    }
+                    let _ = name;
+                }
+            }
+            let root_lba = lba_of(&self.root);
+            write_dir(
+                &mut img,
+                &self.root,
+                root_lba,
+                root_lba,
+                &lba_of,
+                &file_lba_of,
+            );
+
+            // PVD @ 16.
+            let pvd = 16 * SECTOR as usize;
+            img[pvd] = 0x01;
+            img[pvd + 1..pvd + 6].copy_from_slice(b"CD001");
+            let label = self.volume_label.as_bytes();
+            let label_field = &mut img[pvd + 40..pvd + 72];
+            label_field.fill(b' ');
+            label_field[..label.len().min(32)].copy_from_slice(&label[..label.len().min(32)]);
+            let root_rec = record(&[0x00], root_lba, SECTOR, true);
+            img[pvd + 156..pvd + 156 + 34].copy_from_slice(&root_rec[..34]);
+
+            // Optional El Torito boot record @ 17, terminator after.
+            let mut vd = 17 * SECTOR as usize;
+            if self.el_torito {
+                img[vd] = 0x00;
+                img[vd + 1..vd + 6].copy_from_slice(b"CD001");
+                let id = b"EL TORITO SPECIFICATION";
+                img[vd + 7..vd + 7 + id.len()].copy_from_slice(id);
+                vd += SECTOR as usize;
+            }
+            img[vd] = 0xFF;
+            img[vd + 1..vd + 6].copy_from_slice(b"CD001");
+
+            img
+        }
+    }
+
+    /// In-memory `IsoReadAt` over a built test image.
+    pub struct MemReadAt(pub Vec<u8>);
+
+    impl super::IsoReadAt for MemReadAt {
+        async fn read_at(&mut self, offset: u64, len: u32) -> std::io::Result<Vec<u8>> {
+            let start = usize::try_from(offset).unwrap_or(usize::MAX);
+            let end = start.saturating_add(len as usize);
+            if end > self.0.len() {
+                return Err(std::io::Error::new(
+                    std::io::ErrorKind::UnexpectedEof,
+                    "read past end of test image",
+                ));
+            }
+            Ok(self.0[start..end].to_vec())
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::testiso::{MemReadAt, TestIsoBuilder};
+    use super::*;
+
+    fn block_on<T>(f: impl Future<Output = T>) -> T {
+        futures::executor::block_on(f)
+    }
+
+    #[test]
+    fn lookup_finds_nested_file_case_insensitively() {
+        let img = TestIsoBuilder::new("UBUNTU 24.04")
+            .file("/casper/vmlinuz", b"KERNELDATA")
+            .file("/casper/initrd", b"INITRDDATA")
+            .build();
+        let mut r = MemReadAt(img);
+        let loc = block_on(lookup(&mut r, "/CASPER/VMLINUZ")).expect("found");
+        assert_eq!(loc.length, 10);
+        let bytes = block_on(r.read_at(loc.offset, 10)).unwrap();
+        assert_eq!(&bytes, b"KERNELDATA");
+        // Missing file and missing dir both miss cleanly.
+        assert!(block_on(lookup(&mut r, "/casper/missing")).is_none());
+        assert!(block_on(lookup(&mut r, "/nodir/vmlinuz")).is_none());
+        // A directory path that resolves to a directory is not a file hit.
+        assert!(block_on(lookup(&mut r, "/casper")).is_none());
+    }
+
+    #[test]
+    fn strict_mastered_extensionless_names_match() {
+        // Strict level-1 mastering stores "VMLINUZ" as "VMLINUZ.;1" — the
+        // trailing dot must be normalized away or kernels never match.
+        let img = TestIsoBuilder::new("STRICT")
+            .file("/boot/vmlinuz.", b"K") // builder stores "VMLINUZ.;1"
+            .build();
+        let mut r = MemReadAt(img);
+        assert!(
+            block_on(lookup(&mut r, "/boot/vmlinuz")).is_some(),
+            "trailing-dot ISO9660 name must match the dotless path"
+        );
+    }
+
+    #[test]
+    fn lookup_three_levels_deep() {
+        let img = TestIsoBuilder::new("DEEP")
+            .file("/images/pxeboot/vmlinuz", b"ANACONDA")
+            .build();
+        let mut r = MemReadAt(img);
+        let loc = block_on(lookup(&mut r, "images/pxeboot/vmlinuz")).expect("no leading slash ok");
+        assert_eq!(loc.length, 8);
+    }
+
+    #[test]
+    fn caching_reader_dedupes_repeated_directory_reads() {
+        struct Counting<'a> {
+            inner: &'a mut MemReadAt,
+            calls: usize,
+        }
+        impl IsoReadAt for Counting<'_> {
+            async fn read_at(&mut self, offset: u64, len: u32) -> std::io::Result<Vec<u8>> {
+                self.calls += 1;
+                self.inner.read_at(offset, len).await
+            }
+        }
+        let img = TestIsoBuilder::new("CACHE")
+            .file("/a/one", b"1")
+            .file("/a/two", b"2")
+            .build();
+        let mut mem = MemReadAt(img);
+        let mut counting = Counting {
+            inner: &mut mem,
+            calls: 0,
+        };
+        let mut cr = CachingReadAt::new(&mut counting);
+        assert!(block_on(exists(&mut cr, "/a/one")));
+        assert!(block_on(exists(&mut cr, "/a/two")));
+        assert!(!block_on(exists(&mut cr, "/a/three")));
+        drop(cr);
+        // 3 probes × (PVD + root dir + subdir) = 9 uncached; the cache
+        // collapses the repeats to the 3 distinct extents.
+        assert_eq!(counting.calls, 3, "all repeat reads must hit the cache");
+    }
+
+    #[test]
+    fn lookup_local_reads_a_real_file() {
+        let dir = tempfile::tempdir().unwrap();
+        let p = dir.path().join("t.iso");
+        let img = TestIsoBuilder::new("LOCAL")
+            .file("/sources/boot.wim", b"WIMWIM")
+            .build();
+        std::fs::write(&p, &img).unwrap();
+        let loc = lookup_local(&p, "/sources/boot.wim").expect("found");
+        assert_eq!(loc.length, 6);
+        assert!(lookup_local(&p, "/sources/none").is_none());
+    }
+}

crates/iso-store/src/lib.rs

@@ -18,8 +18,15 @@
 
 pub mod entry;
 pub mod introspect;
+// v0.7.4: read-only ISO9660 walker generic over any random-access byte
+// source (local file, NFS READ3, SFTP seek-read). Powers both in-ISO
+// HTTP serving and the probe-based introspection.
+pub mod iso_fs;
 pub mod nfs_share;
 pub mod pxe_logo;
+// v0.7.4: persisted cache of remote-share introspection results so a
+// container restart doesn't re-probe an unchanged 40-ISO library.
+pub mod remote_cache;
 pub mod sftp_share;
 pub mod smb;
 pub mod smb_share;
@@ -29,6 +36,7 @@ pub mod windows;
 
 pub use entry::{BootEntry, BootKind, KernelArgs};
 pub use introspect::{DistroFamily, IntrospectionReport};
+pub use iso_fs::FileLocation;
 // v0.4.65: kernel-mount NFS is gone. SMB shares via Samba's userspace
 // `smbclient` CLI replaced it — works in any container (no
 // CAP_SYS_ADMIN, no host kernel modules), matching how Bootimus and

crates/iso-store/src/nfs_share.rs

@@ -57,7 +57,9 @@
 //! UI to ask for. (If a future server needs Kerberos or non-default
 //! uid mapping we can add those, but for ISO read access nobody does.)
 
-use crate::introspect::IntrospectionReport;
+use crate::introspect::{introspect_reader, provisional_report};
+use crate::iso_fs::{self, FileLocation, IsoReadAt};
+use crate::remote_cache::RemoteIntrospectCache;
 use crate::store::{generate_boot_entries_for, slugify_str, IsoSource, IsoStore};
 use bytes::Bytes;
 use nfs3_client::tokio::TokioConnector;
@@ -100,6 +102,18 @@ const READ_CHUNK_BYTES: u32 = 64 * 1024;
 /// client park gigabytes of decoded ISO in RAM.
 const STREAM_BUFFER_DEPTH: usize = 16;
 
+/// v0.7.4: per-ISO budget for a background introspection probe. A probe
+/// is one connection plus a few dozen KiB-sized reads — sub-second on a
+/// LAN — so anything past this is a wedged server, not a slow one.
+const INTROSPECT_TIMEOUT: Duration = Duration::from_secs(30);
+
+/// One queued background-introspection unit (v0.7.4).
+struct ProbeJob {
+    iso_id: String,
+    filename: String,
+    size: u64,
+}
+
 /// One configured NFS share. The id is derived from server+export so
 /// re-adding the same coordinates is idempotent.
 #[derive(Debug, Clone, Serialize, Deserialize)]
@@ -177,6 +191,9 @@ pub struct NfsShareManager {
     /// opens its own NFS connection so concurrency isn't a hard
     /// requirement, but serializing keeps log output predictable.
     op_lock: Arc<tokio::sync::Mutex<()>>,
+    /// v0.7.4: persisted introspection results keyed `share/path@size`,
+    /// so a restart re-probes only new or replaced ISOs.
+    introspect_cache: RemoteIntrospectCache,
 }
 
 impl NfsShareManager {
@@ -190,6 +207,7 @@ impl NfsShareManager {
             inner: Arc::new(Mutex::new(Inner::default())),
             iso_store,
             op_lock: Arc::new(tokio::sync::Mutex::new(())),
+            introspect_cache: RemoteIntrospectCache::open(work_dir, "nfs_introspect_cache.json"),
         }
     }
 
@@ -387,12 +405,26 @@ impl NfsShareManager {
         };
 
         let mut count = 0u32;
+        let mut to_probe: Vec<ProbeJob> = Vec::new();
         for entry in listing {
             let iso_id = format!("nfs-{}-{}", share.id, slugify_str(&entry.filename));
-            // Same approach as SMB: no real introspection over the
-            // network in v0.4.67. The boot-entry generator falls back
-            // to filename-based sanboot detection.
-            let report = IntrospectionReport::default();
+            // v0.7.4: real introspection over the share — NFSv3 READ3
+            // takes an offset, so the ISO9660 probes work remotely. A
+            // cache hit registers the full report immediately; a miss
+            // registers a provisional filename-based report (so the scan
+            // returns fast) and queues a background probe that upgrades
+            // the entry in place.
+            let cached = self
+                .introspect_cache
+                .get(&share.id, &entry.filename, entry.size);
+            let report = cached.unwrap_or_else(|| {
+                to_probe.push(ProbeJob {
+                    iso_id: iso_id.clone(),
+                    filename: entry.filename.clone(),
+                    size: entry.size,
+                });
+                provisional_report(&entry.filename)
+            });
             let boot_entries = generate_boot_entries_for(&iso_id, &entry.filename, &report);
             let source = IsoSource::Nfs {
                 share_id: share.id.clone(),
@@ -413,11 +445,88 @@ impl NfsShareManager {
             target: "openpxe::nfs",
             id = %id, server = %share.server, export = %share.export,
             iso_count = count,
+            pending_introspection = to_probe.len(),
             "NFS share scanned"
         );
+        if !to_probe.is_empty() {
+            self.spawn_introspection_pass(&share, to_probe);
+        }
         Ok(count)
     }
 
+    /// v0.7.4: probe each queued ISO over its own NFS connection and swap
+    /// the full introspection into the store as results land. Runs
+    /// detached so neither startup nor the share-add API call waits on a
+    /// 40-ISO library; per-ISO failures (or a share removed mid-pass)
+    /// leave the provisional entry in place, which still sanboots.
+    fn spawn_introspection_pass(&self, share: &NfsShare, work: Vec<ProbeJob>) {
+        let store = self.iso_store.clone();
+        let cache = self.introspect_cache.clone();
+        let share_id = share.id.clone();
+        let server = share.server.clone();
+        let export = share.export.clone();
+        let port = share.port;
+        let queued = work.len();
+        tokio::spawn(async move {
+            let mut upgraded = 0usize;
+            for job in work {
+                let probe = async {
+                    let mut reader = NfsReadAt::open(&server, &export, port, &job.filename).await?;
+                    let report =
+                        introspect_reader(&mut reader, job.size, &job.filename, false).await;
+                    reader.finish().await;
+                    Ok::<_, NfsClientError>(report)
+                };
+                match tokio::time::timeout(INTROSPECT_TIMEOUT, probe).await {
+                    Ok(Ok(report)) => {
+                        cache.put(&share_id, &job.filename, job.size, report.clone());
+                        if store.update_external_introspection(&job.iso_id, report) {
+                            upgraded += 1;
+                        }
+                    }
+                    Ok(Err(e)) => tracing::warn!(
+                        target: "openpxe::nfs",
+                        share = %share_id, iso = %job.filename,
+                        "introspection failed: {e}"
+                    ),
+                    Err(_) => tracing::warn!(
+                        target: "openpxe::nfs",
+                        share = %share_id, iso = %job.filename,
+                        "introspection timed out after {}s", INTROSPECT_TIMEOUT.as_secs()
+                    ),
+                }
+            }
+            tracing::info!(
+                target: "openpxe::nfs",
+                share = %share_id, queued, upgraded,
+                "remote introspection pass complete"
+            );
+        });
+    }
+
+    /// v0.7.4: locate `in_iso_path` inside a share-hosted ISO. Returns the
+    /// byte range so the HTTP layer can serve kernel/initrd files out of
+    /// remote ISOs with a follow-up ranged [`Self::stream_iso`].
+    pub async fn locate_in_iso(
+        &self,
+        share_id: &str,
+        filename: &str,
+        in_iso_path: &str,
+    ) -> Result<Option<FileLocation>> {
+        let share = self
+            .get(share_id)
+            .ok_or_else(|| Error::Invalid(format!("no such NFS share '{share_id}'")))?;
+        if filename.contains('/') || filename.contains('\\') || filename.contains("..") {
+            return Err(Error::Invalid(format!("invalid filename '{filename}'")));
+        }
+        let mut reader = NfsReadAt::open(&share.server, &share.export, share.port, filename)
+            .await
+            .map_err(|e| Error::Invalid(format!("nfs open '{filename}': {e}")))?;
+        let loc = iso_fs::lookup(&mut reader, in_iso_path).await;
+        reader.finish().await;
+        Ok(loc)
+    }
+
     fn update_status(
         &self,
         id: &str,
@@ -490,6 +599,96 @@ struct NfsListEntry {
     size: u64,
 }
 
+/// The connection type [`build_connection`] yields.
+type NfsConn = nfs3_client::Nfs3Connection<nfs3_client::tokio::TokioIo<tokio::net::TcpStream>>;
+
+/// v0.7.4: random-access reader over one NFS connection + file handle —
+/// the [`IsoReadAt`] impl that lets the ISO9660 walker and introspection
+/// probes run against share-hosted images.
+struct NfsReadAt {
+    conn: NfsConn,
+    fh: nfs_fh3,
+}
+
+impl NfsReadAt {
+    /// Connect, mount, and LOOKUP `filename` at the export root.
+    async fn open(
+        server: &str,
+        export: &str,
+        port: u16,
+        filename: &str,
+    ) -> std::result::Result<Self, NfsClientError> {
+        let mut conn = build_connection(server, export, port).await?;
+        let root = conn.root_nfs_fh3();
+        let lookup = conn
+            .lookup(&LOOKUP3args {
+                what: diropargs3 {
+                    dir: root,
+                    name: filename3(Opaque::borrowed(filename.as_bytes())),
+                },
+            })
+            .await
+            .map_err(NfsClientError::Rpc)?;
+        let fh = match lookup {
+            Nfs3Result::Ok(o) => o.object,
+            Nfs3Result::Err((status, _)) => {
+                return Err(NfsClientError::Nfsstat(status_label(status)));
+            }
+        };
+        Ok(Self { conn, fh })
+    }
+
+    /// Best-effort unmount. Consumes the reader — it's done.
+    async fn finish(self) {
+        let _ = self.conn.unmount().await;
+    }
+}
+
+impl IsoReadAt for NfsReadAt {
+    async fn read_at(&mut self, offset: u64, len: u32) -> std::io::Result<Vec<u8>> {
+        let mut out: Vec<u8> = Vec::with_capacity(len as usize);
+        let mut off = offset;
+        // READ3 may legally return fewer bytes than asked (server cap);
+        // loop until the exact-read contract is satisfied or the file
+        // genuinely ends short.
+        while (out.len() as u32) < len {
+            let want = (len - out.len() as u32).min(READ_CHUNK_BYTES);
+            let res = self
+                .conn
+                .read(&READ3args {
+                    file: self.fh.clone(),
+                    offset: off,
+                    count: want,
+                })
+                .await
+                .map_err(|e| std::io::Error::other(e.to_string()))?;
+            let ok = match res {
+                Nfs3Result::Ok(o) => o,
+                Nfs3Result::Err((status, _)) => {
+                    return Err(std::io::Error::other(status_label(status)));
+                }
+            };
+            let data = ok.data.as_ref();
+            if data.is_empty() {
+                return Err(std::io::Error::new(
+                    std::io::ErrorKind::UnexpectedEof,
+                    "NFS read past end of file",
+                ));
+            }
+            out.extend_from_slice(data);
+            off += data.len() as u64;
+            if ok.eof && (out.len() as u32) < len {
+                return Err(std::io::Error::new(
+                    std::io::ErrorKind::UnexpectedEof,
+                    "NFS read past end of file",
+                ));
+            }
+        }
+        out.truncate(len as usize);
+        Ok(out)
+    }
+}
+
 /// Connect, READDIR the export root, look up each `*.iso` to get its
 /// size + file handle. Returns a flat list. Errors are returned with
 /// a human-readable message; the caller decides how to surface them.

crates/iso-store/src/remote_cache.rs

@@ -0,0 +1,142 @@
+//! Persisted cache of remote-share introspection results.
+//!
+//! NFS/SFTP introspection costs a connection plus a few dozen small
+//! reads per ISO. Shares are rescanned on every startup and share-add,
+//! so without a cache a 40-ISO library would re-probe 40 ISOs on every
+//! container restart. The cache keys on `share/path@size` — a replaced
+//! file (new size) re-probes, an untouched one is free — and entries
+//! only count as hits when their `introspect_rev` matches the current
+//! logic, so an upgrade that changes detection re-probes everything
+//! exactly once.
+//!
+//! One file per protocol (`nfs_introspect_cache.json`,
+//! `sftp_introspect_cache.json`) so the two managers never contend over
+//! one writer.
+
+use crate::introspect::{IntrospectionReport, INTROSPECT_REV};
+use parking_lot::Mutex;
+use std::collections::HashMap;
+use std::path::{Path, PathBuf};
+use std::sync::Arc;
+
+/// Hard cap on cached entries; beyond it the cache resets rather than
+/// growing unbounded (a cache wipe only costs one re-probe pass).
+const MAX_ENTRIES: usize = 4096;
+
+#[derive(Clone, Debug)]
+pub struct RemoteIntrospectCache {
+    path: Arc<PathBuf>,
+    map: Arc<Mutex<HashMap<String, IntrospectionReport>>>,
+}
+
+impl RemoteIntrospectCache {
+    /// Open (or start empty) the cache at `<work_dir>/<file_name>`.
+    /// A corrupt or missing file is an empty cache, never an error.
+    #[must_use]
+    pub fn open(work_dir: &Path, file_name: &str) -> Self {
+        let path = work_dir.join(file_name);
+        let map = std::fs::read_to_string(&path)
+            .ok()
+            .and_then(|text| {
+                serde_json::from_str::<HashMap<String, IntrospectionReport>>(&text).ok()
+            })
+            .unwrap_or_default();
+        Self {
+            path: Arc::new(path),
+            map: Arc::new(Mutex::new(map)),
+        }
+    }
+
+    fn key(share_id: &str, relative_path: &str, size: u64) -> String {
+        format!("{share_id}/{relative_path}@{size}")
+    }
+
+    /// A hit requires the entry to have been produced by the *current*
+    /// introspection logic — stale-rev entries are misses, which is how
+    /// the cache self-invalidates across upgrades.
+    #[must_use]
+    pub fn get(
+        &self,
+        share_id: &str,
+        relative_path: &str,
+        size: u64,
+    ) -> Option<IntrospectionReport> {
+        self.map
+            .lock()
+            .get(&Self::key(share_id, relative_path, size))
+            .filter(|r| r.introspect_rev == INTROSPECT_REV)
+            .cloned()
+    }
+
+    pub fn put(&self, share_id: &str, relative_path: &str, size: u64, report: IntrospectionReport) {
+        let snapshot = {
+            let mut g = self.map.lock();
+            if g.len() >= MAX_ENTRIES {
+                g.clear();
+            }
+            g.insert(Self::key(share_id, relative_path, size), report);
+            g.clone()
+        };
+        // Persist outside the lock; tmp+rename so a crash mid-write
+        // leaves the previous cache intact.
+        let path = self.path.as_path();
+        let tmp = path.with_extension("json.tmp");
+        let Ok(body) = serde_json::to_vec_pretty(&snapshot) else {
+            return;
+        };
+        if let Some(parent) = path.parent() {
+            let _ = std::fs::create_dir_all(parent);
+        }
+        if std::fs::write(&tmp, body).is_ok() {
+            let _ = std::fs::rename(&tmp, path);
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::introspect::DistroFamily;
+
+    #[test]
+    fn round_trips_across_reopen_and_rev_gates() {
+        let dir = tempfile::tempdir().unwrap();
+        let cache = RemoteIntrospectCache::open(dir.path(), "t.json");
+        assert!(cache.get("s1", "a.iso", 100).is_none());
+
+        let fresh = IntrospectionReport {
+            family: DistroFamily::RhelFedora,
+            introspect_rev: INTROSPECT_REV,
+            el_torito: true,
+            ..Default::default()
+        };
+        cache.put("s1", "a.iso", 100, fresh.clone());
+        assert_eq!(
+            cache.get("s1", "a.iso", 100).unwrap().family,
+            DistroFamily::RhelFedora
+        );
+        // Different size = different file = miss.
+        assert!(cache.get("s1", "a.iso", 101).is_none());
+
+        // Survives a reopen.
+        let cache2 = RemoteIntrospectCache::open(dir.path(), "t.json");
+        assert!(cache2.get("s1", "a.iso", 100).is_some());
+
+        // Stale-rev entries never hit.
+        let stale = IntrospectionReport {
+            family: DistroFamily::Arch,
+            introspect_rev: INTROSPECT_REV - 1,
+            ..Default::default()
+        };
+        cache2.put("s1", "b.iso", 7, stale);
+        assert!(cache2.get("s1", "b.iso", 7).is_none());
+    }
+
+    #[test]
+    fn corrupt_cache_file_starts_empty() {
+        let dir = tempfile::tempdir().unwrap();
+        std::fs::write(dir.path().join("t.json"), b"{nope").unwrap();
+        let cache = RemoteIntrospectCache::open(dir.path(), "t.json");
+        assert!(cache.get("s", "x.iso", 1).is_none());
+    }
+}

crates/iso-store/src/sftp_share.rs

@@ -56,7 +56,9 @@
 //! hint}` error shape is shared so the storage tab renders all three
 //! protocols through one code path.
 
-use crate::introspect::IntrospectionReport;
+use crate::introspect::{introspect_reader, provisional_report};
+use crate::iso_fs::{self, FileLocation, IsoReadAt};
+use crate::remote_cache::RemoteIntrospectCache;
 use crate::store::{generate_boot_entries_for, slugify_str, IsoSource, IsoStore};
 use bytes::Bytes;
 use openpxe_core::{Error, Result};
@@ -96,6 +98,18 @@ const READ_CHUNK_BYTES: usize = 64 * 1024;
 /// body stream. 16 * 64 KiB ≈ 1 MiB max buffer per stream.
 const STREAM_BUFFER_DEPTH: usize = 16;
 
+/// v0.7.4: per-ISO budget for a background introspection probe — one SSH
+/// connection plus a few dozen KiB-sized reads. SSH handshakes cost more
+/// than NFS mounts, but 30s still only trips on a wedged server.
+const INTROSPECT_TIMEOUT: Duration = Duration::from_secs(30);
+
+/// One queued background-introspection unit (v0.7.4).
+struct ProbeJob {
+    iso_id: String,
+    filename: String,
+    size: u64,
+}
+
 /// Which credential the share authenticates with. The secret itself
 /// lives in the 0600 creds file, never here.
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
@@ -209,6 +223,9 @@ pub struct SftpShareManager {
     /// Serializes scan operations on the same manager for predictable
     /// log output; each scan opens its own SSH connection.
     op_lock: Arc<tokio::sync::Mutex<()>>,
+    /// v0.7.4: persisted introspection results keyed `share/path@size`,
+    /// so a restart re-probes only new or replaced ISOs.
+    introspect_cache: RemoteIntrospectCache,
 }
 
 impl SftpShareManager {
@@ -222,6 +239,7 @@ impl SftpShareManager {
             inner: Arc::new(Mutex::new(Inner::default())),
             iso_store,
             op_lock: Arc::new(tokio::sync::Mutex::new(())),
+            introspect_cache: RemoteIntrospectCache::open(work_dir, "sftp_introspect_cache.json"),
         }
     }
 
@@ -474,13 +492,25 @@ impl SftpShareManager {
         };
 
         let mut count = 0u32;
+        let mut to_probe: Vec<ProbeJob> = Vec::new();
         for entry in listing {
             let iso_id = format!("sftp-{}-{}", share.id, slugify_str(&entry.filename));
-            // Same as NFS/SMB: no over-the-network introspection yet, so
-            // register `Unknown` and let the boot-entry generator fall
-            // back to filename-based detection. SFTP *could* do bounded
-            // PVD reads (it has random access) — a follow-up can add it.
-            let report = IntrospectionReport::default();
+            // v0.7.4: real introspection over the share — SFTP file
+            // handles are seekable, so the ISO9660 probes work remotely.
+            // Cache hit → full report now; miss → provisional filename-
+            // based report (scan returns fast) + a queued background
+            // probe that upgrades the entry in place.
+            let cached = self
+                .introspect_cache
+                .get(&share.id, &entry.filename, entry.size);
+            let report = cached.unwrap_or_else(|| {
+                to_probe.push(ProbeJob {
+                    iso_id: iso_id.clone(),
+                    filename: entry.filename.clone(),
+                    size: entry.size,
+                });
+                provisional_report(&entry.filename)
+            });
             let boot_entries = generate_boot_entries_for(&iso_id, &entry.filename, &report);
             let source = IsoSource::Sftp {
                 share_id: share.id.clone(),
@@ -506,11 +536,88 @@ impl SftpShareManager {
             target: "openpxe::sftp",
             id = %id, server = %share.server, export = %share.export,
             iso_count = count,
+            pending_introspection = to_probe.len(),
             "SFTP share scanned"
         );
+        if !to_probe.is_empty() {
+            self.spawn_introspection_pass(&share, &creds, to_probe);
+        }
         Ok(count)
     }
 
+    /// v0.7.4: probe each queued ISO over its own SSH connection and swap
+    /// the full introspection into the store as results land. Detached so
+    /// neither startup nor the share-add API call waits on a big library;
+    /// per-ISO failures leave the provisional entry, which still sanboots.
+    fn spawn_introspection_pass(&self, share: &SftpShare, creds: &SftpCreds, work: Vec<ProbeJob>) {
+        let store = self.iso_store.clone();
+        let cache = self.introspect_cache.clone();
+        let share_id = share.id.clone();
+        let params = ConnParams::from_share(share);
+        let creds = creds.clone();
+        let queued = work.len();
+        tokio::spawn(async move {
+            let mut upgraded = 0usize;
+            for job in work {
+                let probe = async {
+                    let mut reader = SftpReadAt::open(&params, &creds, &job.filename).await?;
+                    let report =
+                        introspect_reader(&mut reader, job.size, &job.filename, false).await;
+                    Ok::<_, SftpClientError>(report)
+                };
+                match tokio::time::timeout(INTROSPECT_TIMEOUT, probe).await {
+                    Ok(Ok(report)) => {
+                        cache.put(&share_id, &job.filename, job.size, report.clone());
+                        if store.update_external_introspection(&job.iso_id, report) {
+                            upgraded += 1;
+                        }
+                    }
+                    Ok(Err(e)) => tracing::warn!(
+                        target: "openpxe::sftp",
+                        share = %share_id, iso = %job.filename,
+                        "introspection failed: {e}"
+                    ),
+                    Err(_) => tracing::warn!(
+                        target: "openpxe::sftp",
+                        share = %share_id, iso = %job.filename,
+                        "introspection timed out after {}s", INTROSPECT_TIMEOUT.as_secs()
+                    ),
+                }
+            }
+            tracing::info!(
+                target: "openpxe::sftp",
+                share = %share_id, queued, upgraded,
+                "remote introspection pass complete"
+            );
+        });
+    }
+
+    /// v0.7.4: locate `in_iso_path` inside a share-hosted ISO. Returns the
+    /// byte range so the HTTP layer can serve kernel/initrd files out of
+    /// remote ISOs with a follow-up ranged [`Self::stream_iso`].
+    pub async fn locate_in_iso(
+        &self,
+        share_id: &str,
+        filename: &str,
+        in_iso_path: &str,
+    ) -> Result<Option<FileLocation>> {
+        let share = self
+            .get(share_id)
+            .ok_or_else(|| Error::Invalid(format!("no such SFTP share '{share_id}'")))?;
+        if filename.contains('/') || filename.contains('\\') || filename.contains("..") {
+            return Err(Error::Invalid(format!("invalid filename '{filename}'")));
+        }
+        let creds = self
+            .read_creds(share_id)
+            .await
+            .map_err(|e| Error::Invalid(format!("could not read credentials: {e}")))?;
+        let params = ConnParams::from_share(&share);
+        let mut reader = SftpReadAt::open(&params, &creds, filename)
+            .await
+            .map_err(|e| Error::Invalid(format!("sftp open '{filename}': {e}")))?;
+        Ok(iso_fs::lookup(&mut reader, in_iso_path).await)
+    }
+
     fn pin_fingerprint(&self, id: &str, fingerprint: String) {
         if fingerprint.is_empty() {
             return;
@@ -652,6 +759,43 @@ struct SftpConn {
     sftp: SftpSession,
 }
 
+/// v0.7.4: random-access reader over one SSH connection + open file
+/// handle — the [`IsoReadAt`] impl that lets the ISO9660 walker and
+/// introspection probes run against share-hosted images. Holds the
+/// `SftpConn` so the SSH session outlives every read.
+struct SftpReadAt {
+    _conn: SftpConn,
+    file: russh_sftp::client::fs::File,
+}
+
+impl SftpReadAt {
+    async fn open(
+        p: &ConnParams,
+        creds: &SftpCreds,
+        filename: &str,
+    ) -> std::result::Result<Self, SftpClientError> {
+        let (conn, _fp) = connect(p, creds).await?;
+        let full = format!("{}/{}", p.export.trim_end_matches('/'), filename);
+        let file = conn
+            .sftp
+            .open(full)
+            .await
+            .map_err(|e| SftpClientError::Sftp(e.to_string()))?;
+        Ok(Self { _conn: conn, file })
+    }
+}
+
+impl IsoReadAt for SftpReadAt {
+    async fn read_at(&mut self, offset: u64, len: u32) -> std::io::Result<Vec<u8>> {
+        self.file.seek(SeekFrom::Start(offset)).await?;
+        let mut buf = vec![0u8; len as usize];
+        // read_exact loops over the transport's short reads and fails
+        // with UnexpectedEof past end-of-file — exactly the contract.
+        self.file.read_exact(&mut buf).await?;
+        Ok(buf)
+    }
+}
+
 /// russh client handler implementing trust-on-first-use host-key
 /// verification. We never construct an `Err` from `check_server_key`;
 /// returning `Ok(false)` makes russh abort the handshake, and the

crates/iso-store/src/smb_share.rs

@@ -56,7 +56,7 @@
 //! streaming. A follow-up release can add libsmbclient-based seek if
 //! a real workload needs it.
 
-use crate::introspect::IntrospectionReport;
+use crate::introspect::provisional_report;
 use crate::store::{generate_boot_entries_for, slugify_str, IsoSource, IsoStore};
 use openpxe_core::{Error, Result};
 use parking_lot::Mutex;
@@ -455,15 +455,14 @@ impl SmbShareManager {
         let mut count = 0u32;
         for entry in listing {
             let iso_id = format!("smb-{}-{}", share.id, slugify_str(&entry.filename));
-            // SMB sources don't get a real introspection pass — that
-            // would require seeking into the ISO9660 PVD over the
-            // network, and smbclient CLI doesn't seek. We register an
-            // `Unknown` family so the boot-entry generator falls back
-            // to generic sanboot/wimboot detection from the filename
-            // and the operator gets *something* bootable. A follow-up
-            // release can do a bounded `smbclient get` of the first
-            // 64 KiB for real detection.
-            let report = IntrospectionReport::default();
+            // SMB sources can't get the content-probe pass NFS/SFTP got
+            // in v0.7.4 — the ISO9660 probes need seeks, and smbclient's
+            // CLI streaming doesn't seek. The provisional filename-token
+            // report is as far as SMB detection goes: family for the UI
+            // when the name says it ("rhel-9.0…", "Win11_…"), and
+            // `introspect_rev = 0` so the entry generator keeps the
+            // optimistic sanboot entry.
+            let report = provisional_report(&entry.filename);
             let boot_entries = generate_boot_entries_for(&iso_id, &entry.filename, &report);
             let source = IsoSource::Smb {
                 share_id: share.id.clone(),

crates/iso-store/src/store.rs

@@ -415,6 +415,28 @@ impl IsoStore {
         self.inner.write().isos.insert(id, meta);
     }
 
+    /// v0.7.4: swap in a completed introspection for an external ISO and
+    /// regenerate its boot entries. Used by the NFS/SFTP managers'
+    /// background probe pass — the scan registers a provisional
+    /// (filename-only) report immediately so startup and share-add stay
+    /// fast, then this upgrades each entry as its probe finishes.
+    /// Operator-set fields (category, password) are preserved; returns
+    /// `false` when the id is gone (share removed or re-scanned away
+    /// mid-probe), which callers treat as a benign no-op.
+    pub fn update_external_introspection(
+        &self,
+        id: &str,
+        introspection: IntrospectionReport,
+    ) -> bool {
+        let mut g = self.inner.write();
+        let Some(m) = g.isos.get_mut(id) else {
+            return false;
+        };
+        m.boot_entries = generate_boot_entries(&m.id, &m.filename, &introspection);
+        m.introspection = introspection;
+        true
+    }
+
     /// Drop every entry that belongs to `share_id`. Used by the SMB
     /// and NFS share managers when an operator removes a share, or
     /// before re-scanning to clean out stale entries. The same id

crates/openpxe/src/main.rs

@@ -215,6 +215,7 @@ async fn main() -> anyhow::Result<()> {
         started_at: time::OffsetDateTime::now_utc(),
         public_base_url: public_base_url.clone(),
         nic_name: net.nic_name,
+        nic_link: net.nic_link,
         subnet_mask: net.subnet_mask,
         gateway: net.gateway,
     };
@@ -448,6 +449,12 @@ struct NetworkInfo {
     nic_name: String,
     subnet_mask: String,
     gateway: String,
+    /// v0.7.2: physical link summary for the Network tab — operstate,
+    /// negotiated speed/duplex, and the port's own MAC. Helps operators
+    /// in multi-NIC / trunked environments confirm *which* port the PXE
+    /// server actually answers on. Empty when sysfs isn't available
+    /// (non-Linux dev builds) or the NIC wasn't identified.
+    nic_link: String,
 }
 
 /// Best-effort population of the Network tab's read-only fields. We shell
@@ -508,9 +515,44 @@ fn detect_network_info(our_ip: Ipv4Addr) -> NetworkInfo {
         }
     }
 
+    info.nic_link = detect_link_info(&info.nic_name);
+
     info
 }
 
+/// v0.7.2: read the NIC's physical link details from sysfs. Every field
+/// is optional — virtual NICs report no speed (`-1` or absent), and
+/// non-Linux dev machines have no `/sys/class/net` at all — so the
+/// result is whatever could be read, joined human-readably, or empty.
+fn detect_link_info(nic: &str) -> String {
+    if nic.is_empty() {
+        return String::new();
+    }
+    let read = |file: &str| {
+        std::fs::read_to_string(format!("/sys/class/net/{nic}/{file}"))
+            .map(|s| s.trim().to_string())
+            .unwrap_or_default()
+    };
+    let mut parts: Vec<String> = Vec::new();
+    let state = read("operstate");
+    if !state.is_empty() {
+        parts.push(format!("link {state}"));
+    }
+    let speed = read("speed");
+    if !speed.is_empty() && speed != "-1" {
+        parts.push(format!("{speed} Mb/s"));
+    }
+    let duplex = read("duplex");
+    if !duplex.is_empty() && duplex != "unknown" {
+        parts.push(format!("{duplex} duplex"));
+    }
+    let mac = read("address");
+    if !mac.is_empty() {
+        parts.push(format!("port {mac}"));
+    }
+    parts.join(" · ")
+}
+
 fn prefix_to_dotted(prefix: u8) -> String {
     let prefix = prefix.min(32);
     let mask: u32 = if prefix == 0 {

crates/webui/src/app.css

@@ -912,3 +912,28 @@ tr.unbootable td:first-child { border-left: 3px solid var(--warn); }
   display: flex; justify-content: flex-end; gap: 10px; margin-top: 18px;
 }
 .modal-actions .submit { width: auto; padding: 8px 18px; }
+
+/* v0.7.2: a label.field directly followed by the card's action button
+   stacked its own 14px bottom margin onto the button's 16px top margin
+   (30px total) — visible on Queue "Launch for all waiting" and the
+   Network "Save". Collapse the doubled gap so every primary action sits
+   the same 16px below its form. */
+.card .body > label.field:has(+ button) { margin-bottom: 0; }
+
+/* v0.7.2: inline list filter above a table (Available images). The input
+   is wrapped in a label.field so it borrows the standard text-field chrome
+   and matches every other input in the app; this wrapper just insets it
+   from the card edges so it lines up with the header text above. */
+.list-search { padding: 14px 16px; }
+
+/* v0.7.4: pager footer under the Available-images table — quiet status
+   text on the left, ghost Prev/Next on the right. */
+.list-pager {
+  display: flex; align-items: center; gap: 8px;
+  padding: 12px 16px;
+  color: var(--fg-dim); font-size: 12px;
+  font-variant-numeric: tabular-nums;
+}
+.list-pager .spacer { flex: 1; }
+.list-pager button { padding: 4px 12px; font-size: 12px; }
+.list-pager button:disabled { opacity: 0.45; cursor: default; }

crates/webui/src/app.js

@@ -178,12 +178,14 @@
     if (iso.introspection.el_torito) {
       return { ok: true, warn: 'generic bootable ISO — boots via sanboot (emulated CD)' };
     }
-    // Remote-share ISOs aren't introspected (no random access over the
-    // network), so el_torito is unknown — assume bootable and let sanboot
-    // try rather than cry wolf.
+    // v0.7.4: NFS/SFTP ISOs now introspect over the share, so a probed
+    // remote ISO flows through the kernel/el_torito branches above like
+    // a local one. introspect_rev 0 means the probe hasn't landed yet
+    // (it runs in the background right after a scan) or never can (SMB —
+    // smbclient can't seek): stay optimistic and let sanboot try.
     const remote = iso.source && iso.source.kind && iso.source.kind !== 'local';
-    if (remote) {
-      return { ok: true, warn: 'remote ISO — not introspected; sanboot is attempted at boot' };
+    if (remote && (iso.introspection.introspect_rev || 0) === 0) {
+      return { ok: true, warn: 'remote ISO — awaiting introspection; sanboot is attempted at boot' };
     }
     // Local ISO with no Windows/Linux boot files and no El Torito catalog:
     // a data/appliance image (e.g. a VMware vCenter bundle), not a bootable
@@ -199,109 +201,47 @@
     })[k] || (k || 'Unknown');
   }
 
-  // v0.7.0: the Boot rules card — ordered first-match-wins rules
-  // (MAC prefix / architecture → target) plus the optional
-  // boot-decision webhook. Saved as one config because rule order
-  // matters. With no rules and no webhook, behavior is identical to
-  // before the feature existed.
-  function bootRulesCard(cfg, targetOptions) {
-    const archChoices = [
-      ['', 'any arch'], ['bios', 'BIOS'], ['uefi-x64', 'UEFI x64'],
-      ['uefi-ia32', 'UEFI IA32'], ['uefi-arm64', 'UEFI ARM64'],
-    ];
-    // v0.7.1: optional first-boot binary pin. "Auto" lets the
-    // escalation ladder learn per machine; pinning skips the learning
-    // walk entirely (e.g. a rack known to run Secure Boot → shim).
-    const modeChoices = [
-      ['', 'auto (learn)'], ['firmware', 'Firmware NIC'],
-      ['builtin', 'iPXE drivers'], ['shim', 'Secure Boot (shim)'],
-    ];
-    const rules = (cfg.rules || []).map(r => Object.assign({}, r));
-    const tbody = el('tbody', {});
-    const msg = el('div', {class:'msg'});
-    const webhookInput = el('input', {type:'text', spellcheck:'false',
-      placeholder:'http://automation.example/boot-decision (optional)',
-      value: cfg.webhook_url || ''});
-
-    const targetSelect = (val) => el('select', {},
-      [el('option', {value:''}, '— target —')]
-      .concat(targetOptions.map(t =>
-        el('option', Object.assign({value: t.id}, t.id === val ? {selected:''} : {}), t.title))));
-
-    const redraw = () => {
-      tbody.innerHTML = '';
-      if (!rules.length) {
-        tbody.appendChild(el('tr', {}, el('td', {colspan:'7', class:'empty', style:'padding:14px'},
-          'No rules. Add one to route whole groups of machines (an OUI, an architecture) to a target — or to pin a boot binary (e.g. Secure Boot racks → shim, zero failed cycles).')));
-      }
-      rules.forEach((r, i) => {
-        const macIn = el('input', {type:'text', spellcheck:'false', placeholder:'aa:bb:cc (prefix)',
-          value: r.mac_prefix || '', oninput: e => { r.mac_prefix = e.target.value; }});
-        const archSel = el('select', {onchange: e => { r.arch = e.target.value; }},
-          archChoices.map(([v, label]) =>
-            el('option', Object.assign({value: v}, v === (r.arch || '') ? {selected:''} : {}), label)));
-        const tgtSel = targetSelect(r.target || '');
-        tgtSel.onchange = e => { r.target = e.target.value; };
-        const modeSel = el('select', {onchange: e => { r.driver_mode = e.target.value; }},
-          modeChoices.map(([v, label]) =>
-            el('option', Object.assign({value: v}, v === (r.driver_mode || '') ? {selected:''} : {}), label)));
-        const noteIn = el('input', {type:'text', placeholder:'note',
-          value: r.note || '', oninput: e => { r.note = e.target.value; }});
-        const enabled = el('input', {type:'checkbox', onchange: e => { r.enabled = e.target.checked; }});
-        enabled.checked = r.enabled !== false;
-        tbody.appendChild(el('tr', {}, [
-          el('td', {}, macIn),
-          el('td', {}, archSel),
-          el('td', {}, tgtSel),
-          el('td', {}, modeSel),
-          el('td', {}, noteIn),
-          el('td', {style:'text-align:center'}, enabled),
-          el('td', {style:'text-align:right'},
-            el('button', {class:'danger', onclick: () => { rules.splice(i, 1); redraw(); }}, '✕')),
-        ]));
-      });
+  // v0.7.2: compact read-out of saved group rules — created from the
+  // unified "Pin MAC" form on the Hosts tab (a prefix or an architecture
+  // there saves a rule instead of a pin). First match wins, top to
+  // bottom. The boot-decision webhook remains available via the API
+  // (/api/boot-rules `webhook_url`) but no longer has a UI knob.
+  function groupRulesCard(cfg, targetOptions) {
+    const rules = (cfg && cfg.rules) || [];
+    if (!rules.length) return null;
+    const titleFor = id => {
+      const t = targetOptions.find(x => x.id === id);
+      return t ? t.title : id;
     };
-    redraw();
-
-    const addBtn = el('button', {class:'ghost', onclick: () => {
-      rules.push({mac_prefix:'', arch:'', target:'', enabled:true, note:''});
-      redraw();
-    }}, '+ Add rule');
-    const saveBtn = el('button', {onclick: async () => {
-      // A rule needs at least one effect: a target or a boot-binary pin.
-      const bad = rules.find(r => r.enabled !== false && !r.target && !r.driver_mode);
-      if (bad) { msg.textContent = 'Every enabled rule needs a target or a boot-binary pin.'; msg.className = 'msg err'; return; }
-      const r = await putJSON('/api/boot-rules', {rules, webhook_url: webhookInput.value.trim()});
-      if (r.ok) { msg.textContent = 'Saved.'; msg.className = 'msg ok'; }
-      else { msg.textContent = 'Save failed: ' + await r.text(); msg.className = 'msg err'; }
-    }}, 'Save rules');
-
+    const modeLabel = {firmware:'Firmware NIC', builtin:'iPXE drivers', shim:'Secure Boot (shim)'};
+    const rows = rules.map((r, i) => el('tr', r.enabled === false ? {style:'opacity:.5'} : {}, [
+      el('td', {class:'mono'}, r.mac_prefix || el('span', {class:'tag'}, 'any MAC')),
+      el('td', {}, r.arch || el('span', {class:'tag'}, 'any arch')),
+      el('td', {}, r.target ? titleFor(r.target) : el('span', {class:'tag'}, '—')),
+      el('td', {}, r.driver_mode
+        ? el('span', {class:'tag accent'}, modeLabel[r.driver_mode] || r.driver_mode)
+        : el('span', {class:'tag'}, 'auto')),
+      el('td', {}, r.note || ''),
+      el('td', {style:'text-align:right'},
+        el('button', {class:'danger', onclick: async () => {
+          if (!confirm('Remove this group rule?')) return;
+          const fresh = await getJSON('/api/boot-rules').catch(() => ({rules: [], webhook_url: ''}));
+          (fresh.rules = fresh.rules || []).splice(i, 1);
+          await putJSON('/api/boot-rules', fresh);
+          render('hosts');
+        }}, 'Remove')),
+    ]));
     return el('div', {class:'card'}, [
       el('header', {}, [
-        el('h2', {}, 'Boot rules'),
+        el('h2', {}, 'Group rules'),
         el('span', {class:'sub'}, 'first match wins · checked top to bottom'),
       ]),
-      el('div', {class:'body'}, [
-        el('table', {}, [
-          el('thead', {}, el('tr', {}, [
-            el('th',{},'MAC prefix'), el('th',{},'Arch'), el('th',{},'Target'),
-            el('th',{},'Boot binary'), el('th',{},'Note'), el('th',{},'On'), el('th',{},''),
-          ])),
-          tbody,
-        ]),
-        el('div', {style:'margin-top:12px'}, [addBtn, saveBtn]),
-        el('label', {class:'field', style:'margin-top:16px;display:block'}, [
-          el('span', {class:'name'}, 'Boot-decision webhook (optional)'),
-          webhookInput,
-          el('span', {class:'hint'},
-            'When no pin or rule matches, OpenPXE GETs this URL with ?mac=…&arch=… ' +
-            'A 200 reply of {"target": "<entry-id>"} chains to that target; anything ' +
-            'else (404, timeout, error) falls through to the menu — a dead endpoint ' +
-            'can never block PXE.'),
-        ]),
-        msg,
-        el('p', {class:'msg', style:'margin-top:10px'},
-          'Decision order per boot: exact MAC pin → first matching rule → webhook → interactive menu.'),
+      el('table', {}, [
+        el('thead', {}, el('tr', {}, [
+          el('th',{},'MAC prefix'), el('th',{},'Arch'), el('th',{},'Target'),
+          el('th',{},'Boot binary'), el('th',{},'Note'), el('th',{},''),
+        ])),
+        el('tbody', {}, rows),
       ]),
     ]);
   }
@@ -329,7 +269,7 @@
       el('label', {class:'field'}, [
         el('span', {class:'name'}, 'Auto IP address (optional)'), ipInput]),
       el('label', {class:'field'}, [
-        el('span', {class:'name'}, 'Unattended file'), sel]),
+        el('span', {class:'name'}, 'Unattended file (in Storage → Advanced)'), sel]),
     ]);
     return {
       wrap,
@@ -507,6 +447,12 @@
             el('div', {class:'v'}, net.gateway || '?'),
             el('div', {class:'k'}, 'Public base URL'),
             el('div', {class:'v'}, net.public_base_url),
+            // v0.7.2: physical link details (operstate · speed · duplex ·
+            // port MAC) so the operator can confirm WHICH port answers PXE
+            // in multi-NIC / trunked environments. Kept last — the joined
+            // value runs long, so it wraps cleanly at the bottom of the list.
+            el('div', {class:'k'}, 'Link'),
+            el('div', {class:'v'}, net.nic_link || '—'),
           ]),
           el('p', {class:'msg'},
             'Server IP, NIC, mask, and gateway are auto-detected at startup. ' +
@@ -896,6 +842,12 @@
           render('storage');
         };
 
+        // v0.7.2: searchable haystack for the list filter — filename,
+        // detected family, category, and source all match.
+        const searchText = [
+          i.filename, familyLabel(i.introspection.family), i.category || '',
+          isSmb ? 'smb' : isNfs ? 'nfs' : 'local', i.id,
+        ].join(' ').toLowerCase();
         const tr = el('tr', b.ok ? {} : {class: 'unbootable'}, [
           el('td', {}, [
             el('div', {style:'display:flex;align-items:center;gap:8px'}, [
@@ -940,8 +892,42 @@
                 }}, 'Remove'),
           ]),
         ]);
+        tr.dataset.search = searchText;
         rowsAndEditors.push(tr, editorRow);
       });
+
+      // v0.7.2: client-side filter over the image table; v0.7.4: paged
+      // 5 at a time so a 50-image library doesn't become a scroll wall.
+      // Rows travel in (row, password-editor) pairs. One view function
+      // applies filter-then-page; editors close on any view change.
+      const ISO_PAGE_SIZE = 5;
+      let isoPage = 0;
+      const pagerInfo = el('span', {});
+      const prevBtn = el('button', {class:'ghost', onclick: () => { isoPage -= 1; applyIsoListView(); }}, '‹ Prev');
+      const nextBtn = el('button', {class:'ghost', onclick: () => { isoPage += 1; applyIsoListView(); }}, 'Next ›');
+      function applyIsoListView() {
+        const q = isoSearch.value.trim().toLowerCase();
+        const visible = [];
+        for (let k = 0; k + 1 < rowsAndEditors.length; k += 2) {
+          const row = rowsAndEditors[k];
+          rowsAndEditors[k + 1].style.display = 'none';
+          row.style.display = 'none';
+          if (!q || (row.dataset.search || '').includes(q)) visible.push(row);
+        }
+        const pages = Math.max(1, Math.ceil(visible.length / ISO_PAGE_SIZE));
+        if (isoPage >= pages) isoPage = pages - 1;
+        if (isoPage < 0) isoPage = 0;
+        visible.slice(isoPage * ISO_PAGE_SIZE, (isoPage + 1) * ISO_PAGE_SIZE)
+          .forEach(r => { r.style.display = ''; });
+        pagerInfo.textContent = visible.length
+          ? 'Showing ' + (isoPage * ISO_PAGE_SIZE + 1) + '–' +
+            Math.min(visible.length, (isoPage + 1) * ISO_PAGE_SIZE) + ' of ' + visible.length
+          : 'No images match';
+        prevBtn.disabled = isoPage === 0;
+        nextBtn.disabled = isoPage >= pages - 1;
+      }
+      const isoSearch = el('input', {type:'search', placeholder:'Filter images by name, type, or source',
+        spellcheck:'false', oninput: () => { isoPage = 0; applyIsoListView(); }});
       const isoTable = isos.length
         ? el('table', {}, [
             el('thead', {}, el('tr', {}, [
@@ -953,6 +939,13 @@
             el('tbody', {}, rowsAndEditors),
           ])
         : el('div', {class:'empty'}, 'No images yet. Upload an ISO or add an SMB share.');
+      // v0.7.4: pager footer, shown once the library outgrows one page.
+      const isoPager = isos.length > ISO_PAGE_SIZE
+        ? el('div', {class:'list-pager'}, [
+            pagerInfo, el('span', {class:'spacer'}), prevBtn, nextBtn,
+          ])
+        : null;
+      if (isos.length) applyIsoListView();
 
       // ── Remote shares section (v0.5.1) ──
       // SMB + NFS unified into one "Remote shares" card with a protocol
@@ -1279,7 +1272,10 @@
       unattFile.onchange = () => { if (unattFile.files[0]) uploadUnattended(unattFile.files[0]); };
 
       const unattRows = unattendedFiles.length
-        ? unattendedFiles.map(f => el('div', {class:'nfs-row'}, [
+        ? unattendedFiles.map(f => el('div', {
+            class:'nfs-row',
+            'data-search': (f.filename + ' ' + unattendedKindLabel(f.kind) + ' ' + f.id).toLowerCase(),
+          }, [
             el('span', {class:'dot ok'}),
             el('div', {}, [
               el('div', {class:'id'}, [
@@ -1307,6 +1303,17 @@
           ]),
           el('div', {class:'body'}, [
             unattDrop, unattFile, unattMsg,
+            // v0.7.2: filter for big answer-file libraries.
+            unattendedFiles.length > 1 ? (() => {
+              const search = el('input', {type:'search', placeholder:'Filter files by name or kind',
+                spellcheck:'false', oninput: () => {
+                  const q = search.value.trim().toLowerCase();
+                  unattRows.forEach(r => {
+                    r.style.display = (!q || (r.dataset.search || '').includes(q)) ? '' : 'none';
+                  });
+                }});
+              return el('label', {class:'field', style:'margin-top:14px;margin-bottom:0'}, search);
+            })() : null,
             el('div', {style:'margin-top:16px;display:grid;gap:8px'}, unattRows),
             el('p', {class:'msg', style:'margin-top:14px'},
               'These answer files drive unattended installs. Attach one to a ' +
@@ -1353,7 +1360,11 @@
             el('h2', {}, 'Available images'),
             el('span', {class:'sub'}, isos.length + ' image' + (isos.length === 1 ? '' : 's')),
           ]),
+          isos.length > 1
+            ? el('div', {class:'list-search'}, el('label', {class:'field', style:'margin-bottom:0'}, isoSearch))
+            : null,
           isoTable,
+          isoPager,
         ]),
       ]), unattendedAdvanced]);
     },
@@ -1378,8 +1389,22 @@
         {id: '_tools_menu',   title: '↳ Tools menu (built-in)'},
       ];
 
-      const macInput   = el('input', {type:'text', placeholder:'aa:bb:cc:dd:ee:ff', spellcheck:'false'});
+      const macInput   = el('input', {type:'text', placeholder:'aa:bb:cc:dd:ee:ff or aa:bb:cc', spellcheck:'false'});
       const labelInput = el('input', {type:'text', placeholder:'optional, e.g. "rack-3 spine"'});
+      // v0.7.2: the former separate "Boot rules" card folded into this
+      // form. A full MAC with no architecture saves a per-host pin
+      // exactly as before; a MAC *prefix* and/or an architecture saves a
+      // first-match-wins group rule instead. Same form, one mental model.
+      const archSel = el('select', {}, [
+        ['', 'any (this exact MAC)'], ['bios', 'BIOS'], ['uefi-x64', 'UEFI x64'],
+        ['uefi-ia32', 'UEFI IA32'], ['uefi-arm64', 'UEFI ARM64'],
+      ].map(([v, t]) => el('option', {value: v}, t)));
+      // v0.7.1's boot-binary pin keeps its home here too (auto = let the
+      // escalation ladder learn; shim = known Secure Boot fleet).
+      const binSel = el('select', {}, [
+        ['', 'auto (learn per machine)'], ['firmware', 'Firmware NIC'],
+        ['builtin', 'iPXE drivers'], ['shim', 'Secure Boot (shim)'],
+      ].map(([v, t]) => el('option', {value: v}, t)));
       const targetSel = el('select', {},
         [el('option', {value:''}, '— choose a target —')]
         .concat(reserved.map(t => el('option', {value: t.id}, t.title)))
@@ -1392,21 +1417,40 @@
       // hostname/IP templated into the served answer file.
       const profileFields = buildProfileFields({}, unattendedFiles, 'form-row cols-3');
 
+      const FULL_MAC = /^([0-9a-f]{2}[:-]){5}[0-9a-f]{2}$/i;
       const upsertBtn = el('button', {onclick: async () => {
-        if (!macInput.value || !targetSel.value) {
-          msg.textContent = 'MAC and target are required.'; msg.className = 'msg err'; return;
+        const mac = macInput.value.trim();
+        const isGroup = !!archSel.value || !!binSel.value || (mac !== '' && !FULL_MAC.test(mac));
+        if (!isGroup) {
+          // Exact-MAC pin — unchanged behavior.
+          if (!mac || !targetSel.value) {
+            msg.textContent = 'MAC and target are required.'; msg.className = 'msg err'; return;
+          }
+          const r = await postJSON('/api/hosts', Object.assign({
+            mac, target: targetSel.value, label: labelInput.value,
+          }, profileFields.read()));
+          if (r.ok) { msg.textContent = 'Saved.'; msg.className = 'msg ok'; render('hosts'); }
+          else { msg.textContent = 'Save failed: ' + await r.text(); msg.className = 'msg err'; }
+          return;
         }
-        const r = await postJSON('/api/hosts', Object.assign({
-          mac: macInput.value, target: targetSel.value, label: labelInput.value,
-        }, profileFields.read()));
-        if (r.ok) {
-          msg.textContent = 'Saved.'; msg.className = 'msg ok';
-          render('hosts');
-        } else {
-          const t = await r.text();
-          msg.textContent = 'Save failed: ' + t; msg.className = 'msg err';
+        // Group rule (prefix and/or architecture). Per-host profile
+        // fields don't apply to a group — they're per-machine values.
+        if (!targetSel.value && !binSel.value) {
+          msg.textContent = 'A group rule needs a target or a boot binary.'; msg.className = 'msg err'; return;
         }
-      }}, 'Bind MAC to target');
+        const p = profileFields.read();
+        if (p.auto_hostname || p.auto_ip || p.unattended_file) {
+          msg.textContent = 'Auto-deploy fields are per-machine — clear them, or use a full MAC.'; msg.className = 'msg err'; return;
+        }
+        const cfg = await getJSON('/api/boot-rules').catch(() => ({rules: [], webhook_url: ''}));
+        (cfg.rules = cfg.rules || []).push({
+          mac_prefix: mac, arch: archSel.value, target: targetSel.value,
+          driver_mode: binSel.value, enabled: true, note: labelInput.value,
+        });
+        const r = await putJSON('/api/boot-rules', cfg);
+        if (r.ok) { msg.textContent = 'Group rule saved.'; msg.className = 'msg ok'; render('hosts'); }
+        else { msg.textContent = 'Save failed: ' + await r.text(); msg.className = 'msg err'; }
+      }}, 'Bind to target');
 
       const rows = hosts.map(h => {
         // v0.5.0: Wake-on-LAN. Only shown for bound hosts (this whole
@@ -1463,23 +1507,32 @@
         el('div', {class:'card'}, [
           el('header', {}, el('h2', {}, 'Pin MAC to boot target')),
           el('div', {class:'body'}, [
+            // v0.7.3: MAC · Label · Architecture · Boot binary share one
+            // 4-up row so the controls line up across the page; the
+            // per-field guidance that used to sit under them moved into the
+            // note below to keep the inputs flush. Target spans full width
+            // on its own line beneath them.
             el('div', {class:'form-row'}, [
-              el('label', {class:'field'}, [el('span', {class:'name'}, 'MAC address'), macInput]),
+              el('label', {class:'field'}, [el('span', {class:'name'}, 'MAC address or prefix'), macInput]),
               el('label', {class:'field'}, [el('span', {class:'name'}, 'Label (optional)'), labelInput]),
-              el('label', {class:'field', style:'grid-column:1 / -1'}, [
-                el('span', {class:'name'}, 'Target'),
-                targetSel,
-                el('span', {class:'hint'},
-                  'Built-in shortcuts skip the menu entirely. Per-ISO entries chain straight to the boot script.'),
-              ]),
+              el('label', {class:'field'}, [el('span', {class:'name'}, 'Architecture (optional)'), archSel]),
+              el('label', {class:'field'}, [el('span', {class:'name'}, 'Boot binary (optional)'), binSel]),
+            ]),
+            el('label', {class:'field', style:'margin-top:14px'}, [
+              el('span', {class:'name'}, 'Target'),
+              targetSel,
             ]),
             el('div', {style:'margin-top:16px'}, profileFields.wrap),
             upsertBtn, msg,
             el('p', {class:'msg', style:'margin-top:14px'},
-              'When a client with a bound MAC requests boot.ipxe, OpenPXE ' +
-              'short-circuits past the interactive menu and chains directly. ' +
-              'If an unattended file is selected, the matching kernel argument ' +
-              'is injected and the hostname/IP are templated into the answer file.'),
+              'A full MAC pins one machine; a MAC prefix (OUI) or an architecture ' +
+              'saves a first-match group rule. Pin the boot binary to “shim” for ' +
+              'Secure Boot racks — zero failed boot cycles. When a matching client ' +
+              'requests boot.ipxe, OpenPXE short-circuits past the interactive menu ' +
+              'and chains directly; decision order is exact MAC pin → first matching ' +
+              'group rule → menu. If an unattended file is selected on a pin, the ' +
+              'matching kernel argument is injected and the hostname/IP are templated ' +
+              'into the answer file.'),
           ]),
         ]),
         el('div', {class:'card'}, [
@@ -1489,7 +1542,7 @@
           ]),
           table,
         ]),
-        bootRulesCard(rulesCfg, reserved.concat(targets)),
+        groupRulesCard(rulesCfg, reserved.concat(targets)),
         el('div', {class:'card'}, [
           el('header', {}, [
             el('h2', {}, 'Host log'),
@@ -2166,9 +2219,24 @@
         el('div', {class:'about-hero'}, [
           el('h2', {}, 'OpenPXE'),
           el('p', {class:'lead'},
-            'Air-gapped network PXE boot, container-native, that anyone can run. ' +
-            'No CDN calls, no telemetry, no surprise external dependencies — ship ' +
-            'the image once, run it forever.'),
+            'The network-boot platform for modern infrastructure. Drop in an ISO ' +
+            'and every machine on your network — BIOS, UEFI, Secure Boot — can ' +
+            'boot it, image from it, and install unattended. One container, one ' +
+            'static binary, nothing installed on clients, nothing leaving your network.'),
+          el('div', {style:'display:grid;grid-template-columns:repeat(3,1fr);gap:14px;margin:18px 0'}, [
+            ['Boot anything', 'Linux, Windows, hypervisors, rescue tools — uploaded ' +
+              'ISOs become menu entries automatically, served on demand from local ' +
+              'disk or your existing NFS, SMB, or SFTP libraries.'],
+            ['Adapt to every machine', 'Per-machine boot intelligence: firmware quirks, ' +
+              'NIC driver fallback, and a Microsoft-signed Secure Boot chain are ' +
+              'negotiated automatically and remembered — no toggles, no client prep.'],
+            ['Run it in production', 'SAML single sign-on, token-scoped answer files, ' +
+              'fleet routing rules, Wake-on-LAN, queued mass deployment, Prometheus ' +
+              'metrics. Built in Rust for boot infrastructure that cannot flinch.'],
+          ].map(([h, body]) => el('div', {}, [
+            el('h3', {style:'margin:0 0 6px;font-size:13.5px'}, h),
+            el('p', {class:'msg', style:'font-size:12px;margin:0'}, body),
+          ]))),
           el('div', {class:'who'}, [
             el('span', {}, 'Developer: '), el('strong', {}, 'Miles Ward'), el('br'),
             el('span', {}, 'Version: '), el('strong', {}, status.version || '?'), el('br'),
@@ -2179,15 +2247,16 @@
           ]),
           el('div', {style:'margin-top:18px'}, [updBtn, updMsg]),
           el('p', {class:'msg', style:'margin-top:18px'},
-            'iPXE is an internal implementation detail. Everything the firmware ' +
-            'executes is generated from the settings on these tabs — there is no ' +
-            'hand-written .ipxe path anywhere in this product.'),
+            'Private by design: no telemetry, no CDN calls, no runtime ' +
+            'dependencies on the outside world. Air-gapped labs, customer sites ' +
+            'without internet, and locked-down OpenShift clusters run the same ' +
+            'image, the same way, indefinitely.'),
           el('p', {class:'msg'},
-            'Vision: a deployment-grade tool that works on first try in the most ' +
-            'awkward environments — air-gapped labs, customer sites without ' +
-            'internet, OpenShift clusters with strict SCCs — without ever asking ' +
-            'an operator to install drivers signed with test certificates or to ' +
-            'flip "testsigning" on a target machine.'),
+            'Principled by default: OpenPXE never asks an operator to install ' +
+            'test-signed drivers, modify a client’s trust store, or weaken ' +
+            'Secure Boot. Everything the firmware executes is generated from the ' +
+            'settings on these tabs — there are no hand-written boot scripts to ' +
+            'maintain and no internals to learn.'),
         ]),
       ]);