Merge perf/first-frame-latency: driver proto v4 + first-frame/resize latency (P0-P2)

Brings the first-frame-latency branch (P0.1 transition tracing, P1.1/P1.2
Welcome-time display prep, P2 in-place resize; pf-driver-proto v3 -> v4 with
IOCTL_UPDATE_MODES) onto current main. The branch predates the W6.2/W7 splits,
so git's rename detection carried most of it into the moved crates
(pf-capture idd_push, pf-vdisplay manager/pf_vdisplay, pf-win-display,
pf-driver-proto, the driver workspace) and the punktfunk1.rs remainder was
re-homed by hand:

- native/handshake.rs: welcome/start trace marks + the Welcome-time display
  prep spawn (the prep thread BECOMES the stream thread; hand-off via a
  SyncSender<SessionContext>). negotiate() gains bringup/quit/stop and returns
  the PrepHandle.
- native.rs: bringup/resize_ms creation + the stop/quit flags hoisted BEFORE
  the handshake (the close watcher splits: flags pre-handshake, lifecycle
  events post-handshake where `hello` exists); punch_done stamp; the data
  plane adopts the prep thread's result or builds inline.
- native/stream.rs: SessionContext/SendStats carry the trace; send_loop
  finishes it on the first video packet; the resize path gains the in-place
  fast path (try_inplace_resize) with the full rebuild as fallback, restructured
  so both share the post-rebuild bookkeeping; prepare_display/PreparedDisplay/
  PrepHandle; build_pipeline(+retry) thread the stage marks.
- session_status/mgmt: ttff_ms + last_resize_ms per session (union with the
  lifecycle-events fields main added to the same spots).
- pf-capture: Capturer gains capture_target_id() + resize_output() defaults.
- pf-vdisplay manager: perf's faster activation poll (60x50ms) + the settle
  floor before the PnP sweep, on main's knobs/no-trait shape.

Also: packaging/windows/build-gamepad-drivers.ps1 is ASCII again (an em-dash
from the pf-mouse work tripped windows-host.yml's locale-safety gate on main).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

api/openapi.json

@@ -4617,6 +4617,15 @@
             "format": "int32",
             "minimum": 0
           },
+          "last_resize_ms": {
+            "type": [
+              "integer",
+              "null"
+            ],
+            "format": "int32",
+            "description": "Most recent mid-stream resize total, reconfigure → pipeline rebuilt, in ms (native sessions;\n`null` when no resize happened / GameStream).",
+            "minimum": 0
+          },
           "min_fec": {
             "type": "integer",
             "format": "int32",
@@ -4629,6 +4638,15 @@
             "description": "Video payload size per packet (bytes).",
             "minimum": 0
           },
+          "time_to_first_frame_ms": {
+            "type": [
+              "integer",
+              "null"
+            ],
+            "format": "int32",
+            "description": "Session bring-up total, hello → first video packet, in ms (native sessions; `null` on the\nGameStream plane or while the session is still bringing up).",
+            "minimum": 0
+          },
           "width": {
             "type": "integer",
             "format": "int32",

crates/pf-capture/src/lib.rs

@@ -56,6 +56,24 @@ pub trait Capturer: Send {
     fn pipeline_depth(&self) -> usize {
         1
     }
+
+    /// The OS display-target id this capturer is bound to (Windows IDD-push), so the resize path
+    /// can verify the display it just reconfigured is STILL the one this capturer serves (an
+    /// in-place resize keeps the target; a re-arrival fallback mints a new one, which needs a
+    /// fresh capturer). `None` = the backend has no such identity (every non-IDD backend).
+    fn capture_target_id(&self) -> Option<u32> {
+        None
+    }
+
+    /// HOST-INITIATED output resize (latency plan P2.3): the session's resize handler has ALREADY
+    /// committed the display's new mode (the manager's in-place mode set), so a capable capturer
+    /// re-sizes its capture surface NOW — no descriptor-poll debounce (that machinery stays, for
+    /// EXTERNAL changes only) and no teardown: the capture pipeline and its send thread survive;
+    /// only the encoder is swapped by the caller once the first new-size frame arrives. Returns
+    /// `true` when handled; `false` (the default) routes the caller to the full-rebuild path.
+    fn resize_output(&mut self, _width: u32, _height: u32) -> bool {
+        false
+    }
 }
 
 /// A deterministic moving test pattern (BGRx). Lets the spike exercise the encode → file →

crates/pf-capture/src/windows/idd_push.rs

@@ -698,8 +698,24 @@ impl IddPushCapturer {
             let enabled_hdr = client_10bit
                 && pf_win_display::win_display::set_advanced_color(target.target_id, true);
             if enabled_hdr {
-                // Let the colorspace change settle before the driver composes + we size the ring.
+                // Let the colorspace change settle before the driver composes + we size the ring:
-                std::thread::sleep(Duration::from_millis(250));
+                // poll the CCD advanced-color state instead of a fixed sleep (latency plan P0.4),
+                // ceiling = the old 250 ms. A read that never flips within the ceiling proceeds
+                // exactly like the fixed sleep did — the ring is sized FP16 from `enabled_hdr`
+                // either way (the set succeeded; only the driver's compose flip may lag, which the
+                // stash/format-guard machinery absorbs).
+                let hdr_settle = Instant::now();
+                while hdr_settle.elapsed() < Duration::from_millis(250) {
+                    if crate::win_display::advanced_color_enabled(target.target_id) == Some(true) {
+                        break;
+                    }
+                    std::thread::sleep(Duration::from_millis(25));
+                }
+                tracing::debug!(
+                    target_id = target.target_id,
+                    settle_ms = hdr_settle.elapsed().as_millis() as u64,
+                    "IDD push: advanced-color (HDR) enable settle"
+                );
             }
             // A failed open-time read defaults to SDR (unless the 10-bit path enabled HDR above) —
             // there is no "last known" yet; the descriptor poller corrects a wrong guess mid-session.
@@ -983,7 +999,14 @@ impl IddPushCapturer {
                     self.no_first_frame_diagnosis(st)
                 );
             }
-            std::thread::sleep(Duration::from_millis(20));
+            // Event-driven wait (latency plan P0.6): the driver signals the frame-ready event on
+            // every publish, so wake on it instead of a blind sleep — the 20 ms timeout keeps the
+            // driver_status polls above live (status writes don't signal the event). Consuming a
+            // signal here is fine: `next_frame` re-checks the atomic `latest` token, never the
+            // event, for truth.
+            // SAFETY: `self.event` is this capturer's owned, live auto-reset event handle;
+            // `WaitForSingleObject` only reads the handle and the 20 ms timeout bounds the wait.
+            let _ = unsafe { WaitForSingleObject(HANDLE(self.event.as_raw_handle()), 20) };
         }
     }
 
@@ -1615,6 +1638,39 @@ impl Capturer for IddPushCapturer {
         // always has its own texture).
         pf_host_config::config().idd_depth.clamp(1, OUT_RING)
     }
+
+    fn capture_target_id(&self) -> Option<u32> {
+        Some(self.target_id)
+    }
+
+    fn resize_output(&mut self, width: u32, height: u32) -> bool {
+        // Host-initiated resize (latency plan P2.3): the session's resize handler has already
+        // committed the display's new mode (the manager's in-place mode set), so recreate the ring
+        // at the new size NOW — no DescriptorPoller two-strike debounce (that stays, unchanged,
+        // for EXTERNAL changes: HDR flips, game mode-sets). The driver re-attaches to the fresh
+        // ring and republishes; on an in-place mode set the OS's mode-set full redraw gives the
+        // stash/first frame within the recover window. Same recover-or-drop arming as the
+        // poller-driven recreate, so a ring that can't re-attach still fails the session cleanly
+        // instead of freezing.
+        if (width, height) == (self.width, self.height) {
+            return true; // already at the requested size (refresh-only change) — nothing to do
+        }
+        tracing::info!(
+            target_id = self.target_id,
+            from = format!("{}x{}", self.width, self.height),
+            to = format!("{width}x{height}"),
+            "IDD push: host-initiated resize — recreating the ring at the new mode"
+        );
+        self.recovering_since.get_or_insert_with(Instant::now);
+        if let Err(e) = self.recreate_ring(self.display_hdr, width, height) {
+            tracing::warn!(
+                error = %format!("{e:#}"),
+                "IDD push: host-initiated ring recreate failed — falling back to a full rebuild"
+            );
+            return false;
+        }
+        true
+    }
 }
 
 /// A 4:4:4 session while the display is HDR: there is no 10-bit full-chroma source (the FP16

crates/pf-driver-proto/src/lib.rs

@@ -59,7 +59,19 @@ pub const fn interface_guid_fields() -> (u32, u16, u16, [u8; 8]) {
 /// attach a ring naming a different monitor ([`frame::DRV_STATUS_BIND_FAIL`], the gamepad channel's
 /// `pad_index` validation applied to frames). A v2 host never stamps the field, so a v3 driver
 /// against a v2 host would refuse every attach — lockstep by the handshake, as ever.
-pub const PROTOCOL_VERSION: u32 = 3;
+/// v4: ADDITIVE — [`control::IOCTL_UPDATE_MODES`] (the in-place mid-stream resize,
+/// `design/first-frame-and-resize-latency.md` P2): the driver refreshes a LIVE monitor's advertised
+/// target-mode list (`IddCxMonitorUpdateModes2`) so the OS can mode-set to an arbitrary new mode
+/// without a REMOVE→ADD monitor hotplug. Nothing existing changed, so the host accepts a v3 driver
+/// too ([`MIN_DRIVER_PROTOCOL_VERSION`]) and simply falls back to the re-arrival resize against it;
+/// a v4 driver serving an older (v3-asserting) host fails that host's strict handshake — ship
+/// driver+host together, as ever.
+pub const PROTOCOL_VERSION: u32 = 4;
+
+/// The OLDEST driver protocol this host still drives (v4 is additive over v3 — see the v4 note on
+/// [`PROTOCOL_VERSION`]): a v3 driver lacks only `IOCTL_UPDATE_MODES`, which the host gates on the
+/// handshake-reported version and covers with the re-arrival fallback.
+pub const MIN_DRIVER_PROTOCOL_VERSION: u32 = 3;
 
 /// `CTL_CODE(FILE_DEVICE_UNKNOWN = 0x22, func, METHOD_BUFFERED = 0, FILE_ANY_ACCESS = 0)`.
 pub const fn ctl_code(func: u32) -> u32 {
@@ -91,6 +103,13 @@ pub mod control {
     /// host duplicated into the driver's WUDFHost process. Input [`SetFrameChannelRequest`]. Sent once
     /// after the ring is created and again on every mid-session ring recreate (HDR-mode flip).
     pub const IOCTL_SET_FRAME_CHANNEL: u32 = ctl_code(0x906);
+    /// Refresh a LIVE monitor's advertised target-mode list to a new preferred mode (+ the built-in
+    /// fallbacks) via `IddCxMonitorUpdateModes2` — the in-place mid-stream resize (v4,
+    /// `design/first-frame-and-resize-latency.md` P2). Input [`UpdateModesRequest`]. The host then
+    /// CCD-forces the new mode active on the SAME monitor: no REMOVE→ADD hotplug, the monitor's OS
+    /// identity (saved per-monitor DPI) and the driver's swap-chain/stash machinery survive. A v3
+    /// driver fails this unknown IOCTL → the host falls back to the re-arrival resize.
+    pub const IOCTL_UPDATE_MODES: u32 = ctl_code(0x907);
 
     /// `IOCTL_ADD` input. A monotonic `session_id` keys the monitor (the host's refcount manager owns
     /// collision safety — no more SudoVDA's 16-byte GUID + pid-mangling). The driver advertises this
@@ -164,6 +183,22 @@ pub mod control {
         pub session_id: u64,
     }
 
+    /// `IOCTL_UPDATE_MODES` input (v4): the live monitor (by its ADD `session_id`) and the new
+    /// preferred mode its target-mode list should lead with. The driver replaces the stored list
+    /// (new mode first, then its built-in fallbacks — the same shape ADD produces) and pushes it to
+    /// the OS via `IddCxMonitorUpdateModes2`; success means the OS accepted the new list, after
+    /// which the host force-sets the mode via CCD/GDI as usual.
+    #[repr(C)]
+    #[derive(Clone, Copy, Pod, Zeroable, Debug, PartialEq, Eq)]
+    pub struct UpdateModesRequest {
+        pub session_id: u64,
+        pub width: u32,
+        pub height: u32,
+        pub refresh_hz: u32,
+        /// Pads the `u64`-aligned struct to a multiple of 8 (Pod forbids implicit tail padding).
+        pub _reserved: u32,
+    }
+
     /// `IOCTL_SET_RENDER_ADAPTER` input (the GPU the IddCx swap-chain should render on).
     #[repr(C)]
     #[derive(Clone, Copy, Pod, Zeroable, Debug, PartialEq, Eq)]
@@ -253,6 +288,12 @@ pub mod control {
         assert!(size_of::<RemoveRequest>() == 8);
         assert!(offset_of!(RemoveRequest, session_id) == 0);
 
+        assert!(size_of::<UpdateModesRequest>() == 24);
+        assert!(offset_of!(UpdateModesRequest, session_id) == 0);
+        assert!(offset_of!(UpdateModesRequest, width) == 8);
+        assert!(offset_of!(UpdateModesRequest, height) == 12);
+        assert!(offset_of!(UpdateModesRequest, refresh_hz) == 16);
+
         assert!(size_of::<SetRenderAdapterRequest>() == 8);
         assert!(offset_of!(SetRenderAdapterRequest, luid_low) == 0);
         assert!(offset_of!(SetRenderAdapterRequest, luid_high) == 4);
@@ -1104,6 +1145,27 @@ mod tests {
         assert_eq!(bytes[32..40], 0x2000u64.to_le_bytes());
     }
 
+    #[test]
+    fn update_modes_request_roundtrips_and_versions_cohere() {
+        let req = control::UpdateModesRequest {
+            session_id: 42,
+            width: 2560,
+            height: 1409, // deliberately arbitrary — the in-place path serves window-drag modes
+            refresh_hz: 120,
+            _reserved: 0,
+        };
+        let bytes = bytemuck::bytes_of(&req);
+        assert_eq!(bytes.len(), 24);
+        assert_eq!(
+            *bytemuck::from_bytes::<control::UpdateModesRequest>(bytes),
+            req
+        );
+        assert_eq!(bytes[8..12], 2560u32.to_le_bytes());
+        // The compat window: v4 is additive over v3, so the host floor stays one below.
+        assert_eq!(PROTOCOL_VERSION, 4);
+        assert_eq!(MIN_DRIVER_PROTOCOL_VERSION, 3);
+    }
+
     #[test]
     fn gamepad_names_and_magics_are_stable() {
         assert_eq!(gamepad::xusb_boot_name(0), "Global\\pfxusb-boot-0");
@@ -1145,6 +1207,7 @@ mod tests {
             control::IOCTL_GET_INFO,
             control::IOCTL_CLEAR_ALL,
             control::IOCTL_SET_FRAME_CHANNEL,
+            control::IOCTL_UPDATE_MODES,
         ];
         for (i, a) in all.iter().enumerate() {
             for b in &all[i + 1..] {

crates/pf-vdisplay/src/vdisplay/windows/manager.rs

@@ -35,7 +35,8 @@ use windows::Win32::System::Threading::{
 use super::{DisplayOwnership, Mode, VirtualOutput};
 use pf_win_display::win_display::{
     count_other_active, force_extend_topology, isolate_displays_ccd, resolve_gdi_name,
-    restore_displays_ccd, set_active_mode, set_virtual_primary_ccd, SavedConfig,
+    restore_displays_ccd, set_active_mode, set_virtual_primary_ccd, wait_mode_settled,
+    wait_target_departed, SavedConfig,
 };
 
 #[path = "manager/driver.rs"]
@@ -204,6 +205,16 @@ pub struct VirtualDisplayManager {
     /// `&'static` singleton with no raw-handle smuggling.
     device: Mutex<DeviceSlot>,
     watchdog_s: AtomicU32,
+    /// The driver's handshake-reported protocol version (0 until the first open). The in-place
+    /// resize (latency plan P2) gates its UPDATE_MODES attempt on `>= 4`; a v3 driver keeps the
+    /// already-advertised fast path + the re-arrival fallback.
+    driver_proto: AtomicU32,
+    /// Latched `true` after an UPDATE_MODES round-trip failed to make the new mode settable —
+    /// on-glass (build 26200) the OS pins a monitor's settable set at ARRIVAL (it re-parses our
+    /// description + re-queries target modes, then ignores both), so every further attempt for an
+    /// out-of-arrival-list mode would only waste ~1 s per resize before the same re-arrival
+    /// fallback. One attempt per process, in case a future OS build honors the refresh.
+    update_modes_futile: AtomicBool,
     /// Monotonic lease-generation counter (was the `MON_GEN` global).
     gen: AtomicU64,
     state: Mutex<MgrInner>,
@@ -234,6 +245,8 @@ pub(crate) fn init(driver: Box<dyn VdisplayDriver>) -> &'static VirtualDisplayMa
         driver,
         device: Mutex::new(DeviceSlot::default()),
         watchdog_s: AtomicU32::new(3),
+        driver_proto: AtomicU32::new(0),
+        update_modes_futile: AtomicBool::new(false),
         gen: AtomicU64::new(1),
         state: Mutex::new(MgrInner::default()),
         setup_lock: Mutex::new(()),
@@ -321,9 +334,10 @@ impl VirtualDisplayManager {
         // FFI in the caller's apartment; the `device` mutex (held here) serializes it, so there is no
         // concurrent open. `open` has no handle precondition to uphold, and the `OwnedHandle` it
         // returns is the sole owner of the device.
-        let (handle, watchdog_s) = unsafe { self.driver.open(reap)? };
+        let (handle, watchdog_s, driver_proto) = unsafe { self.driver.open(reap)? };
         slot.opened_once = true;
         self.watchdog_s.store(watchdog_s, Ordering::Relaxed);
+        self.driver_proto.store(driver_proto, Ordering::Relaxed);
         let raw = HANDLE(handle.as_raw_handle());
         slot.current = Some(Arc::new(handle));
         if !reap {
@@ -418,9 +432,10 @@ impl VirtualDisplayManager {
             if let Some(SlotState::Lingering { mon, .. } | SlotState::Pinned { mon }) =
                 inner.slots.remove(&slot)
             {
+                let old_target = mon.target_id;
                 tracing::info!(
                     slot,
-                    old_target = mon.target_id,
+                    old_target,
                     "IDD-push reconnect — preempting the kept (lingering/pinned) monitor, recreating a fresh one"
                 );
                 // SAFETY: `teardown_removed` requires `dev` to be a valid control handle; `dev` is the
@@ -430,7 +445,16 @@ impl VirtualDisplayManager {
                 unsafe { self.teardown_removed(dev, &mut inner, mon) };
                 // Let the OS finish the ASYNC monitor departure before the next ADD; a back-to-back
                 // REMOVE→ADD races the teardown and the ADD IOCTL is rejected under reconnect churn.
-                thread::sleep(Duration::from_millis(400));
+                // Verified-state wait, ceiling = the old fixed 400 ms settle (latency plan P0.3).
+                // SAFETY: CCD query FFI over a `Copy` target id, under the held `state` lock.
+                let departed =
+                    unsafe { wait_target_departed(old_target, Duration::from_millis(400)) };
+                if !departed {
+                    tracing::debug!(
+                        old_target,
+                        "preempted monitor still in the active CCD set after the departure ceiling"
+                    );
+                }
             }
         }
 
@@ -446,9 +470,10 @@ impl VirtualDisplayManager {
         if matches!(inner.slots.get(&slot), Some(SlotState::Active { mon, .. }) if !wudf_alive(mon.wudf_pid))
         {
             if let Some(SlotState::Active { mon, .. }) = inner.slots.remove(&slot) {
+                let old_target = mon.target_id;
                 tracing::warn!(
                     slot,
-                    old_target = mon.target_id,
+                    old_target,
                     wudf_pid = mon.wudf_pid,
                     "virtual monitor's WUDFHost is gone — preempting the dead monitor, recreating"
                 );
@@ -457,8 +482,9 @@ impl VirtualDisplayManager {
                 // retired, kept alive; see `DeviceSlot`). `mon` was just removed from the map, so it
                 // is exclusively owned here — no aliasing.
                 unsafe { self.teardown_removed(dev, &mut inner, mon) };
-                // Same async-departure settle as the reconnect preempt above.
+                // Same async-departure settle as the reconnect preempt above (verified wait, P0.3).
-                thread::sleep(Duration::from_millis(400));
+                // SAFETY: CCD query FFI over a `Copy` target id, under the held `state` lock.
+                let _ = unsafe { wait_target_departed(old_target, Duration::from_millis(400)) };
             }
         }
 
@@ -475,6 +501,57 @@ impl VirtualDisplayManager {
                 _ => unreachable!("just matched Active"),
             };
             if cur_mode != mode {
+                // IN-PLACE mode set first (latency plan P2): an already-advertised resolution
+                // (arrival list + the driver's same-id mode history) is CCD-forced on the SAME
+                // monitor — no REMOVE→ADD, so the monitor's OS identity (saved per-monitor DPI),
+                // the driver-side swap-chain machinery and the retained frame stash all survive,
+                // and the whole hotplug cost (departure settle + activation ladder + re-isolate)
+                // disappears. An out-of-list mode fails FAST (see `resize_in_place`) and falls
+                // through to the proven re-arrival below.
+                {
+                    let in_place = {
+                        let Some(SlotState::Active { mon, refs }) = inner.slots.get_mut(&slot)
+                        else {
+                            unreachable!("just matched Active");
+                        };
+                        // SAFETY: `dev` is the handle `ensure_device()` returned above; the CCD
+                        // waits inside run under the held `state` lock (this fn's discipline).
+                        match unsafe { self.resize_in_place(dev, mon, mode) } {
+                            Ok(()) => {
+                                // Same join semantics as the re-arrival: +1 ref for the new
+                                // (build-then-drop overlap) lease; `gen` untouched, so the old
+                                // session's lease stays valid.
+                                *refs += 1;
+                                let refs = *refs;
+                                let out = self.output_for(slot, mon, quit.clone());
+                                tracing::info!(
+                                    slot,
+                                    refs,
+                                    backend = self.driver.name(),
+                                    "virtual monitor resized IN PLACE (identity + swap-chain kept)"
+                                );
+                                Some(out)
+                            }
+                            Err(e) => {
+                                // Expected-normal for a first-seen arbitrary size (the OS pins
+                                // settable modes at arrival; the re-arrival teaches it) — info,
+                                // not warn.
+                                tracing::info!(
+                                    slot,
+                                    reason = %format!("{e:#}"),
+                                    "in-place resize not possible — monitor re-arrival"
+                                );
+                                None
+                            }
+                        }
+                    };
+                    if let Some(out) = in_place {
+                        // The width changed — re-arrange the group so auto-row siblings don't
+                        // overlap the resized display (no-op for a single member).
+                        self.apply_group_layout(&mut inner);
+                        return Ok(out);
+                    }
+                }
                 let Some(SlotState::Active { mon, refs }) = inner.slots.remove(&slot) else {
                     unreachable!("just matched Active");
                 };
@@ -736,8 +813,12 @@ impl VirtualDisplayManager {
     /// # Safety
     /// Runs the CCD (QueryDisplayConfig / SetDisplayConfig) FFI; call under the `state` lock.
     unsafe fn resolve_target_gdi(&self, target_id: u32) -> Option<String> {
-        for _ in 0..15 {
+        // 50 ms sampling (latency plan P0.5): the SAME 3 s per-stage ceilings — the 3-stage ladder
-            thread::sleep(Duration::from_millis(200));
+        // structure encodes real failure modes (headless auto-activate, integrated-panel clone,
+        // lid-closed path activation) and is untouched — but a typical activation resolves on an
+        // early poll, so finer sampling shaves ~150 ms off every stage crossing.
+        for _ in 0..60 {
+            thread::sleep(Duration::from_millis(50));
             // SAFETY: `resolve_gdi_name` is `unsafe` for its CCD FFI; it takes a plain `Copy` `u32`
             // target id by value and returns an owned `String`, so no caller memory is borrowed.
             if let Some(n) = unsafe { resolve_gdi_name(target_id) } {
@@ -746,8 +827,8 @@ impl VirtualDisplayManager {
         }
         // SAFETY: `force_extend_topology` only calls `SetDisplayConfig` (CCD) with no borrowed memory.
         unsafe { force_extend_topology() };
-        for _ in 0..15 {
+        for _ in 0..60 {
-            thread::sleep(Duration::from_millis(200));
+            thread::sleep(Duration::from_millis(50));
             // SAFETY: as the resolve loop above.
             if let Some(n) = unsafe { resolve_gdi_name(target_id) } {
                 return Some(n);
@@ -756,8 +837,8 @@ impl VirtualDisplayManager {
         // SAFETY: `activate_target_path` runs the CCD query/apply FFI with owned local buffers; the
         // `Copy` target id is passed by value, under the `state` lock — the sole topology mutator.
         if unsafe { pf_win_display::win_display::activate_target_path(target_id) } {
-            for _ in 0..15 {
+            for _ in 0..60 {
-                thread::sleep(Duration::from_millis(200));
+                thread::sleep(Duration::from_millis(50));
                 // SAFETY: as the resolve loops above.
                 if let Some(n) = unsafe { resolve_gdi_name(target_id) } {
                     return Some(n);
@@ -919,19 +1000,40 @@ impl VirtualDisplayManager {
                         );
                     }
                 }
-                thread::sleep(Duration::from_millis(1500)); // let the topology settle before capture opens
+                // Topology settle before capture opens: verified-state wait (latency plan P0.2) —
+                // poll until the target's path + active mode are committed, ceiling = the old fixed
+                // 1500 ms sleep (a rejected mode / slow third-party CCD-lock holder burns the
+                // ceiling and proceeds, exactly like the sleep it replaces).
+                let settle_start = std::time::Instant::now();
+                // SAFETY: CCD/GDI query FFI over a `Copy` target id, under the held `state` lock.
+                let settled = unsafe {
+                    wait_mode_settled(added.target_id, mode, Duration::from_millis(1500))
+                };
+                tracing::info!(
+                    settle_ms = settle_start.elapsed().as_millis() as u64,
+                    verified = settled,
+                    "topology settle (verified-state wait)"
+                );
 
                 // EXPERIMENTAL `pnp_disable_monitors`, second selector (ANY topology): monitors
                 // that are connected but NOT part of the desktop — the standby TV/monitor the
                 // deactivated-set selector above structurally misses (it never had an active path
                 // to deactivate), yet whose periodic standby wake events drive the same Windows
                 // reaction cascade (rationale in `windows/monitor_devnode.rs`). Runs AFTER the
-                // settle sleep so the active flags it reads are the committed ones (a display
+                // settle so the active flags it reads are the committed ones (a display still
-                // still mid-activation from the primary topology's force-EXTEND must not read as
+                // mid-activation from the primary topology's force-EXTEND must not read as
-                // inactive and get disabled); in Extend the active physical panels are untouched
+                // inactive and get disabled) — and since the verified wait above only confirms
-                // by construction. First member only — the sweep is group-scoped like the
+                // OUR target (not a physical still lighting up from force-EXTEND), this opt-in
-                // isolate; later members join an already-swept desktop.
+                // sweep keeps the old FULL settle as its floor before reading those flags.
+                // In Extend the active physical panels are untouched by construction. First
+                // member only — the sweep is group-scoped like the isolate; later members join
+                // an already-swept desktop.
                 if first_member && crate::policy::prefs().pnp_disable_monitors() {
+                    if let Some(rest) =
+                        Duration::from_millis(1500).checked_sub(settle_start.elapsed())
+                    {
+                        thread::sleep(rest);
+                    }
                     let mut keep = inner.target_ids();
                     keep.push(added.target_id);
                     for id in pf_win_display::monitor_devnode::disable_connected_inactive(&keep) {
@@ -962,6 +1064,101 @@ impl VirtualDisplayManager {
         })
     }
 
+    /// Mid-stream resize IN PLACE (latency plan P2): the driver refreshes the LIVE monitor's
+    /// advertised target-mode list to lead with `mode` (`IOCTL_UPDATE_MODES` →
+    /// `IddCxMonitorUpdateModes2`, protocol v4), the OS re-enumerates the target's settable modes
+    /// (waited on, bounded), and the usual CCD/GDI force-set + verified settle (P0.2) commit it —
+    /// on the SAME monitor: target id, GDI name, saved per-monitor DPI, the driver's swap-chain
+    /// worker and its retained frame stash all survive (the OS reassigns the swap-chain across a
+    /// mode set; the preserved-publisher/stash hand-off covers that flap — what it was built for).
+    /// On success `mon.mode` is updated in place; any failure leaves `mon` untouched (still at the
+    /// old mode) and the caller falls back to [`re_add`](Self::re_add).
+    ///
+    /// # Safety
+    /// `dev` must be the live control handle; runs the CCD/GDI FFI under the `state` lock.
+    unsafe fn resize_in_place(&self, dev: HANDLE, mon: &mut Monitor, mode: Mode) -> Result<()> {
+        let gdi = mon
+            .gdi_name
+            .clone()
+            .context("in-place resize needs a resolved GDI name")?;
+        let t0 = Instant::now();
+        // FAST PATH (driver-independent): the OS already offers this resolution — the monitor's
+        // arrival list, which since the driver's mode-history union contains every size this
+        // identity ever served — so a plain CCD mode set reaches it with no driver round-trip.
+        let already = crate::win_display::wait_mode_advertised(&gdi, mode, Duration::ZERO);
+        if !already {
+            // Out-of-arrival-list mode. On-glass (build 26200) the OS re-parses our description
+            // AND re-queries target modes after UpdateModes2 — our callbacks served the fresh
+            // list — yet the SETTABLE set stays pruned to the arrival list: the monitor
+            // source-mode set is pinned at arrival. So one bounded UPDATE_MODES attempt per
+            // process (in case a future build honors the refresh), then latch it futile and fail
+            // fast to the re-arrival — whose same-id history union makes THIS size settable in
+            // place from then on.
+            if self.driver_proto.load(Ordering::Relaxed) < 4 {
+                anyhow::bail!(
+                    "{}x{} is not in the advertised mode set (v3 driver: in-place reaches only \
+                     arrival-list modes)",
+                    mode.width,
+                    mode.height
+                );
+            }
+            if self.update_modes_futile.load(Ordering::Relaxed) {
+                anyhow::bail!(
+                    "{}x{} is not in the advertised mode set (UPDATE_MODES latched futile — the \
+                     OS pins settable modes at monitor arrival; the re-arrival teaches this size \
+                     to the identity's history)",
+                    mode.width,
+                    mode.height
+                );
+            }
+            tracing::info!(
+                old = format!(
+                    "{}x{}@{}",
+                    mon.mode.width, mon.mode.height, mon.mode.refresh_hz
+                ),
+                new = format!("{}x{}@{}", mode.width, mode.height, mode.refresh_hz),
+                target = mon.target_id,
+                "virtual-display: updating the live monitor's modes for an in-place resize"
+            );
+            // SAFETY: `dev` is the live control handle (this fn's contract); `update_modes`
+            // forwards it to a synchronous IOCTL with owned/borrowed locals only.
+            unsafe { self.driver.update_modes(dev, &mon.key, mode) }?;
+            // SAFETY: CCD query/apply FFI under the held `state` lock (this fn's contract).
+            unsafe { crate::win_display::force_mode_reenumeration() };
+            if !crate::win_display::wait_mode_advertised(&gdi, mode, Duration::from_millis(800)) {
+                self.update_modes_futile.store(true, Ordering::Relaxed);
+                anyhow::bail!(
+                    "OS did not advertise {}x{} within {}ms of the driver mode-list update \
+                     (offers: {:?}) — latching UPDATE_MODES off for this process",
+                    mode.width,
+                    mode.height,
+                    t0.elapsed().as_millis(),
+                    crate::win_display::advertised_resolutions(&gdi)
+                );
+            }
+        }
+        let advertised_ms = t0.elapsed().as_millis() as u64;
+        set_active_mode(&gdi, mode);
+        // Verified-state settle (P0.2): the same committed-state predicate as the create paths. A
+        // mode set that did not commit within the ceiling routes to the re-arrival fallback.
+        let settle_start = Instant::now();
+        // SAFETY: CCD/GDI query FFI over a `Copy` target id, under the held `state` lock.
+        let settled =
+            unsafe { wait_mode_settled(mon.target_id, mode, Duration::from_millis(1500)) };
+        if !settled {
+            anyhow::bail!(
+                "in-place mode set did not commit within 1.5s (advertised after {advertised_ms} ms)"
+            );
+        }
+        tracing::info!(
+            advertised_ms,
+            settle_ms = settle_start.elapsed().as_millis() as u64,
+            "in-place resize committed (verified-state wait)"
+        );
+        mon.mode = mode;
+        Ok(())
+    }
+
     /// Mid-stream resize by monitor RE-ARRIVAL (`design/midstream-resolution-resize.md` Fix 1).
     ///
     /// The pf-vdisplay driver freezes a monitor's advertised mode list at `IOCTL_ADD` time (the
@@ -1013,9 +1210,17 @@ impl VirtualDisplayManager {
             );
         }
         // Let the OS finish the ASYNC monitor departure before the ADD — a back-to-back REMOVE→ADD
-        // races the teardown and the ADD is rejected under churn (same 400 ms settle as the reconnect
+        // races the teardown and the ADD is rejected under churn. Verified departure wait, ceiling =
-        // preempt path).
+        // the old fixed 400 ms settle (latency plan P0.3); the driver's ghost-reap ADD retry remains
-        thread::sleep(Duration::from_millis(400));
+        // the backstop for a departure the CCD reports early.
+        let depart_start = std::time::Instant::now();
+        // SAFETY: CCD query FFI over a `Copy` target id, under the held `state` lock.
+        let departed = unsafe { wait_target_departed(old.target_id, Duration::from_millis(400)) };
+        tracing::info!(
+            depart_ms = depart_start.elapsed().as_millis() as u64,
+            verified = departed,
+            "re-arrival: old monitor departure settle"
+        );
         // 2. ADD a fresh monitor at the NEW mode, reusing the slot as the preferred (stable) id.
         let render_pin = resolve_render_pin();
         // SAFETY: `dev` is the live control handle; `render_pin`/`client_hdr` are owned `Copy`/`Option`
@@ -1042,7 +1247,18 @@ impl VirtualDisplayManager {
                 //    the group's first-member restore snapshot.
                 // SAFETY: CCD FFI over borrowed Copy target ids, under the `state` lock.
                 unsafe { self.reisolate_after_swap(inner, added.target_id) };
-                thread::sleep(Duration::from_millis(1500)); // let the topology settle before capture reopens
+                // Topology settle before capture reopens: verified-state wait, ceiling = the old
+                // fixed 1500 ms sleep (latency plan P0.2 — the re-arrival twin).
+                let settle_start = std::time::Instant::now();
+                // SAFETY: CCD/GDI query FFI over a `Copy` target id, under the held `state` lock.
+                let settled = unsafe {
+                    wait_mode_settled(added.target_id, mode, Duration::from_millis(1500))
+                };
+                tracing::info!(
+                    settle_ms = settle_start.elapsed().as_millis() as u64,
+                    verified = settled,
+                    "re-arrival topology settle (verified-state wait)"
+                );
             }
             None => tracing::warn!(
                 "re-arrival target {} not yet an active display path (auto-activate, EXTEND preset \

crates/pf-vdisplay/src/vdisplay/windows/pf_vdisplay.rs

@@ -341,7 +341,7 @@ impl VdisplayDriver for PfVdisplayDriver {
         "pf-vdisplay"
     }
 
-    unsafe fn open(&self, reap_orphans: bool) -> Result<(OwnedHandle, u32)> {
+    unsafe fn open(&self, reap_orphans: bool) -> Result<(OwnedHandle, u32, u32)> {
         // SAFETY: `open_device` is `unsafe` only because it issues SetupAPI enumeration + `CreateFileW`
         // FFI; it takes no arguments and returns an owned raw `HANDLE` (or `Err`). Called here on the
         // backend-init thread, with no precondition beyond a valid thread context.
@@ -397,27 +397,43 @@ impl VdisplayDriver for PfVdisplayDriver {
         }
         let info: control::InfoReply =
             bytemuck::pod_read_unaligned(&info_buf[..size_of::<control::InfoReply>()]);
-        if info.protocol_version != pf_driver_proto::PROTOCOL_VERSION {
+        // HARD floor/ceiling instead of strict equality since v4: v4 is ADDITIVE over v3
+        // (IOCTL_UPDATE_MODES — the in-place resize), so this host still drives a v3 driver and
+        // simply gates the in-place path on the reported version (re-arrival fallback). Anything
+        // below the floor or ABOVE this host's own version stays a loud failure.
+        if info.protocol_version < pf_driver_proto::MIN_DRIVER_PROTOCOL_VERSION
+            || info.protocol_version > pf_driver_proto::PROTOCOL_VERSION
+        {
             anyhow::bail!(
-                "pf-vdisplay protocol mismatch: host expects {}, driver reports {} — install matching \
+                "pf-vdisplay protocol mismatch: host drives {}..={}, driver reports {} — install \
-                 host + driver",
+                 matching host + driver",
+                pf_driver_proto::MIN_DRIVER_PROTOCOL_VERSION,
                 pf_driver_proto::PROTOCOL_VERSION,
                 info.protocol_version
             );
         }
         let watchdog_s = info.watchdog_timeout_s.max(1);
-        tracing::info!(
+        if info.protocol_version < pf_driver_proto::PROTOCOL_VERSION {
-            "pf-vdisplay protocol {} (watchdog timeout {}s)",
+            tracing::warn!(
-            info.protocol_version,
+                "pf-vdisplay protocol {} (host supports {}): driver lacks the in-place resize — \
-            watchdog_s
+                 mid-stream resizes use the monitor re-arrival path until the driver is updated",
-        );
+                info.protocol_version,
+                pf_driver_proto::PROTOCOL_VERSION
+            );
+        } else {
+            tracing::info!(
+                "pf-vdisplay protocol {} (watchdog timeout {}s)",
+                info.protocol_version,
+                watchdog_s
+            );
+        }
         // Reap monitors orphaned by a crashed previous host — a FIRST-CLASS op (driver returns
         // SUCCESS). FIRST open of the process only: a REOPEN (the manager retired a dead handle after
         // a driver upgrade / WUDFHost restart) can race sessions that still believe they are live, and
         // an unconditional CLEAR_ALL there would raze them.
         if !reap_orphans {
             reap_ghost_monitors();
-            return Ok((device, watchdog_s));
+            return Ok((device, watchdog_s, info.protocol_version));
         }
         let mut none: [u8; 0] = [];
         // SAFETY: `raw` borrows the live `OwnedHandle` above. `IOCTL_CLEAR_ALL` has no input and no
@@ -434,7 +450,7 @@ impl VdisplayDriver for PfVdisplayDriver {
         // monitor-slot budget — prevents the 0x80070490 slot-exhaustion wedge from carrying across
         // restarts (the reason a restart's CLEAR_ALL alone never recovered it before).
         reap_ghost_monitors();
-        Ok((device, watchdog_s))
+        Ok((device, watchdog_s, info.protocol_version))
     }
 
     unsafe fn add_monitor(
@@ -592,6 +608,38 @@ impl VdisplayDriver for PfVdisplayDriver {
         })
     }
 
+    unsafe fn update_modes(&self, dev: HANDLE, key: &MonitorKey, mode: Mode) -> Result<()> {
+        let MonitorKey::Session(session_id) = key else {
+            anyhow::bail!("pf-vdisplay: unexpected monitor key kind");
+        };
+        let req = control::UpdateModesRequest {
+            session_id: *session_id,
+            width: mode.width,
+            height: mode.height,
+            refresh_hz: mode.refresh_hz,
+            _reserved: 0,
+        };
+        let mut none: [u8; 0] = [];
+        // SAFETY: per `update_modes`'s contract `dev` is the live control handle. `bytes_of(&req)`
+        // borrows the local `UpdateModesRequest` for the duration of this synchronous call as the
+        // input bytes; `none` is empty, so there is no output buffer.
+        unsafe {
+            ioctl(
+                dev,
+                control::IOCTL_UPDATE_MODES,
+                bytemuck::bytes_of(&req),
+                &mut none,
+            )
+        }
+        .map(|_| ())
+        .with_context(|| {
+            format!(
+                "pf-vdisplay UPDATE_MODES {}x{}@{}",
+                mode.width, mode.height, mode.refresh_hz
+            )
+        })
+    }
+
     unsafe fn remove_monitor(&self, dev: HANDLE, key: &MonitorKey) -> Result<()> {
         let MonitorKey::Session(session_id) = key else {
             anyhow::bail!("pf-vdisplay: unexpected monitor key kind");
@@ -740,4 +788,80 @@ mod tests {
         thread::sleep(Duration::from_secs(3));
         drop(vout); // triggers REMOVE + stops the pinger
     }
+
+    /// Live in-place resize spike — skipped unless `PUNKTFUNK_PF_VDISPLAY_LIVE=1` (needs a v4
+    /// pf-vdisplay driver installed + the host service STOPPED, single-instance guard). Answers the
+    /// P2 open questions on real glass with no streaming client: create at one mode, then acquire
+    /// the SAME session's slot at a DIFFERENT mode — the manager's resize branch runs UPDATE_MODES
+    /// → mode-advertised wait → set_active_mode → verified settle. In-place success is visible as
+    /// the SAME OS target id on the second output (a re-arrival fallback mints a new one) plus the
+    /// committed active resolution; the test reports which path ran and asserts the mode landed.
+    #[test]
+    fn live_inplace_resize() {
+        if std::env::var("PUNKTFUNK_PF_VDISPLAY_LIVE").is_err() {
+            return;
+        }
+        // Live-run diagnostics: surface the manager/backend tracing (activation ladder, settle
+        // waits, UPDATE_MODES) on stdout — a bare test harness has no subscriber, which made the
+        // first on-glass run blind.
+        let _ = tracing_subscriber::fmt()
+            .with_env_filter(
+                tracing_subscriber::EnvFilter::try_from_default_env()
+                    .unwrap_or_else(|_| "debug".into()),
+            )
+            .try_init();
+        // Context probe: can this process see the CCD active-path set at all? (`None` = the query
+        // itself fails in this session/window-station — the whole ladder would be blind, and a
+        // "monitor never activated" verdict would be an artifact of the test context.)
+        // SAFETY: CCD query over an owned empty slice (test-only diagnostics).
+        let active0 = unsafe { crate::win_display::count_other_active(&[]) };
+        println!("spike: CCD active paths visible before create: {active0:?}");
+        let mut vd = PfVdisplayDisplay::new().expect("open pf-vdisplay");
+        let first = vd
+            .create(Mode {
+                width: 1920,
+                height: 1080,
+                refresh_hz: 60,
+            })
+            .expect("create virtual display");
+        let t1 = first
+            .win_capture
+            .as_ref()
+            .expect("no capture target")
+            .target_id;
+        thread::sleep(Duration::from_secs(2)); // let the activation/settle fully quiesce
+                                               // A deliberately arbitrary (window-drag-shaped) mode the ADD never advertised.
+        let t0 = std::time::Instant::now();
+        let second = vd
+            .create(Mode {
+                width: 2356,
+                height: 1332,
+                refresh_hz: 60,
+            })
+            .expect("in-place resize acquire");
+        let resize_ms = t0.elapsed().as_millis();
+        let t2 = second
+            .win_capture
+            .as_ref()
+            .expect("no capture target")
+            .target_id;
+        let in_place = t1 == t2;
+        // SAFETY: CCD query over a Copy target id (test-only diagnostics).
+        let active = unsafe { crate::win_display::active_resolution(t2) };
+        println!(
+            "in-place resize spike: in_place={in_place} (target {t1} -> {t2}) took {resize_ms} ms, \
+             active resolution now {active:?}"
+        );
+        assert_eq!(
+            active,
+            Some((2356, 1332)),
+            "the new mode did not become the active resolution"
+        );
+        assert!(
+            in_place,
+            "the resize fell back to re-arrival (target id changed) — UPDATE_MODES path not taken"
+        );
+        drop(second);
+        drop(first);
+    }
 }

crates/pf-win-display/src/win_display.rs

@@ -232,6 +232,169 @@ pub unsafe fn active_resolution(target_id: u32) -> Option<(u32, u32)> {
     Some((dm.dmPelsWidth, dm.dmPelsHeight))
 }
 
+/// Verified-state topology-settle wait (latency plan P0.2): poll the CCD state until the target is
+/// actually COMMITTED — an active path exists (the GDI name resolves) and the active resolution
+/// equals the requested one — instead of sleeping a fixed interval. The conditions are exactly what
+/// `resolve_gdi_name`/`set_active_mode` already established once; this waits until the OS reports
+/// them stable. `ceiling` (the old fixed sleep) is the worst-case bound: a mode the driver rejected
+/// (`set_active_mode` left the OS default) or a slow third-party CCD-lock holder (SteelSeries
+/// class) burns the ceiling and proceeds — behavior identical to the fixed sleep it replaces.
+/// Returns `true` when the state verified (typical: one or two 25 ms polls), `false` on ceiling.
+///
+/// # Safety
+/// Runs the CCD/GDI query FFI; call under the manager `state` lock like the callers it serves.
+pub(crate) unsafe fn wait_mode_settled(
+    target_id: u32,
+    mode: Mode,
+    ceiling: std::time::Duration,
+) -> bool {
+    let deadline = std::time::Instant::now() + ceiling;
+    loop {
+        // SAFETY (both calls): CCD/GDI FFI over a `Copy` target id, owned returns — the callers'
+        // own safety contract (under the `state` lock) covers them.
+        if resolve_gdi_name(target_id).is_some()
+            && active_resolution(target_id) == Some((mode.width, mode.height))
+        {
+            return true;
+        }
+        if std::time::Instant::now() >= deadline {
+            return false;
+        }
+        std::thread::sleep(std::time::Duration::from_millis(25));
+    }
+}
+
+/// Re-commit the CURRENT active config with `SDC_FORCE_MODE_ENUMERATION` — the nudge that makes
+/// the OS re-query an indirect display's target modes. Observed on-glass (P2): after
+/// `IddCxMonitorUpdateModes2` the OS did NOT re-enumerate on its own within 2 s, so a freshly
+/// advertised mode never became settable; the isolate/layout paths already re-commit with this
+/// flag for the same "the OS won't re-evaluate unless told" class. Best-effort.
+///
+/// # Safety
+/// Runs the CCD query/apply FFI; call under the manager `state` lock (sole topology mutator).
+pub(crate) unsafe fn force_mode_reenumeration() -> bool {
+    let Some((paths, modes)) = query_active_config() else {
+        return false;
+    };
+    let rc = SetDisplayConfig(
+        Some(paths.as_slice()),
+        Some(modes.as_slice()),
+        SDC_APPLY
+            | SDC_USE_SUPPLIED_DISPLAY_CONFIG
+            | SDC_ALLOW_CHANGES
+            | SDC_FORCE_MODE_ENUMERATION,
+    );
+    if rc != 0 {
+        tracing::debug!("force mode re-enumeration: SetDisplayConfig rc={rc:#x}");
+    }
+    rc == 0
+}
+
+/// The distinct resolutions `gdi_name` currently advertises (diagnostics for the in-place-resize
+/// path: what the OS actually offers when a requested mode never shows up).
+pub(crate) fn advertised_resolutions(gdi_name: &str) -> Vec<(u32, u32)> {
+    let wname: Vec<u16> = gdi_name.encode_utf16().chain(std::iter::once(0)).collect();
+    let mut set = std::collections::BTreeSet::new();
+    let mut i = 0u32;
+    loop {
+        let mut dm = DEVMODEW {
+            dmSize: size_of::<DEVMODEW>() as u16,
+            ..Default::default()
+        };
+        // SAFETY: `wname` is a live NUL-terminated UTF-16 device name; `&mut dm` is a live,
+        // size-stamped DEVMODEW the API fills for mode index `i`. Both outlive the call.
+        let ok = unsafe {
+            EnumDisplaySettingsW(
+                PCWSTR(wname.as_ptr()),
+                ENUM_DISPLAY_SETTINGS_MODE(i),
+                &mut dm,
+            )
+        }
+        .as_bool();
+        if !ok {
+            break;
+        }
+        set.insert((dm.dmPelsWidth, dm.dmPelsHeight));
+        i += 1;
+    }
+    set.into_iter().collect()
+}
+
+/// Wait (bounded) until `gdi_name` ADVERTISES `mode`'s resolution in its display-mode list — the
+/// gate between a driver-side mode-list refresh (`IOCTL_UPDATE_MODES`, latency plan P2) and the
+/// CCD/GDI force-set: the OS re-evaluates an indirect display's settable modes asynchronously after
+/// `IddCxMonitorUpdateModes2`, so an immediate `set_active_mode` could race the re-enumeration and
+/// silently leave the old mode. Returns `true` once any refresh at the requested WxH is enumerable.
+pub(crate) fn wait_mode_advertised(
+    gdi_name: &str,
+    mode: Mode,
+    ceiling: std::time::Duration,
+) -> bool {
+    let wname: Vec<u16> = gdi_name.encode_utf16().chain(std::iter::once(0)).collect();
+    let deadline = std::time::Instant::now() + ceiling;
+    loop {
+        let mut i = 0u32;
+        loop {
+            let mut dm = DEVMODEW {
+                dmSize: size_of::<DEVMODEW>() as u16,
+                ..Default::default()
+            };
+            // SAFETY: `wname` is a live NUL-terminated UTF-16 device name whose pointer stays valid
+            // for the call; `&mut dm` is a live, size-stamped DEVMODEW the API fills for mode index
+            // `i`. Both outlive this synchronous call.
+            let ok = unsafe {
+                EnumDisplaySettingsW(
+                    PCWSTR(wname.as_ptr()),
+                    ENUM_DISPLAY_SETTINGS_MODE(i),
+                    &mut dm,
+                )
+            }
+            .as_bool();
+            if !ok {
+                break;
+            }
+            if dm.dmPelsWidth == mode.width && dm.dmPelsHeight == mode.height {
+                return true;
+            }
+            i += 1;
+        }
+        if std::time::Instant::now() >= deadline {
+            return false;
+        }
+        std::thread::sleep(std::time::Duration::from_millis(25));
+    }
+}
+
+/// Monitor-departure wait (latency plan P0.3): after a REMOVE, poll until the target has left the
+/// ACTIVE CCD set — two consecutive absent samples, so one transient query failure mid-teardown
+/// can't read as "gone" — instead of sleeping the fixed departure settle. `ceiling` (the old fixed
+/// sleep) bounds the worst case. The OS-side departure may still be finishing driver-side when the
+/// CCD stops listing the target; the ADD path's ghost-reap retry (pf_vdisplay) remains the backstop
+/// for that rare race, exactly as it was for a settle that expired. Returns `true` when departure
+/// was observed, `false` on ceiling.
+///
+/// # Safety
+/// Runs the CCD query FFI; call under the manager `state` lock like the callers it serves.
+pub(crate) unsafe fn wait_target_departed(target_id: u32, ceiling: std::time::Duration) -> bool {
+    let deadline = std::time::Instant::now() + ceiling;
+    let mut absent_streak = 0u32;
+    loop {
+        // SAFETY: CCD FFI over a `Copy` target id, owned return, under the caller's `state` lock.
+        if resolve_gdi_name(target_id).is_none() {
+            absent_streak += 1;
+            if absent_streak >= 2 {
+                return true;
+            }
+        } else {
+            absent_streak = 0;
+        }
+        if std::time::Instant::now() >= deadline {
+            return false;
+        }
+        std::thread::sleep(std::time::Duration::from_millis(25));
+    }
+}
+
 /// Toggle the virtual-display target's advanced-color (HDR) state via the CCD API. Disabling HDR while on the
 /// secure (Winlogon) desktop makes it render SDR/composed so DXGI Desktop Duplication can capture it
 /// (the HDR fullscreen independent-flip otherwise storms `ACCESS_LOST` → black); re-enable on return so

crates/punktfunk-host/src/bringup.rs

@@ -0,0 +1,89 @@
+//! Session-transition latency trace (design/first-frame-and-resize-latency.md P0.1).
+//!
+//! One [`Trace`] per transition — session bring-up (`hello → … → first_packet`) or a mid-stream
+//! resize (`reconfigure_received → … → pipeline_rebuilt`) — collects millisecond stage stamps
+//! across the threads a transition crosses (handshake task, display-prep/encode thread, send
+//! thread) and emits ONE summary `info!` line when the transition completes, so every landed
+//! latency change is measured against a number instead of vibes. The completed total also lands
+//! in a shared slot [`crate::session_status`] exposes (`time_to_first_frame_ms` /
+//! `last_resize_ms`), so the web-console Dashboard and future regressions can read it per session.
+//!
+//! Deliberately coarse: stages are stamped where the session layer can see them; layers the trace
+//! doesn't reach (the Windows display manager's activation ladder / settle waits) log their own
+//! per-stage deltas and correlate by wall clock.
+
+use std::sync::atomic::{AtomicBool, AtomicU32, Ordering};
+use std::sync::{Arc, Mutex};
+use std::time::Instant;
+
+/// A single transition's stage trace. Cheap and thread-safe: `mark` is a mutex push, `finish`
+/// emits the one summary line (exactly once — later calls no-op, so an abandoned trace stays
+/// silent).
+pub(crate) struct Trace {
+    /// Which transition this traces (`"bringup"` / `"resize"`) — the summary line's `kind`.
+    kind: &'static str,
+    origin: Instant,
+    /// `(stage, ms since origin)` in stamp order.
+    stages: Mutex<Vec<(&'static str, u32)>>,
+    finished: AtomicBool,
+    /// Where the completed total lands — shared with [`crate::session_status`].
+    total_ms: Arc<AtomicU32>,
+}
+
+impl Trace {
+    /// Start a trace at "now" (= the first stage's zero point). `total_ms` is the shared slot the
+    /// completed total is stored into (0 until the transition finishes).
+    pub(crate) fn start(kind: &'static str, total_ms: Arc<AtomicU32>) -> Arc<Self> {
+        Arc::new(Self {
+            kind,
+            origin: Instant::now(),
+            stages: Mutex::new(Vec::new()),
+            finished: AtomicBool::new(false),
+            total_ms,
+        })
+    }
+
+    /// The shared slot the completed total is stored into (for `session_status::register`).
+    pub(crate) fn total_slot(&self) -> Arc<AtomicU32> {
+        self.total_ms.clone()
+    }
+
+    /// Stamp a stage at "now" — first occurrence only (a retried build re-crosses its stamp
+    /// points; the first crossing is the one the transition timeline wants). No-op after
+    /// [`finish`](Self::finish), so steady-state paths that also cross a stamped point stay free.
+    pub(crate) fn mark(&self, stage: &'static str) {
+        if self.finished.load(Ordering::Relaxed) {
+            return;
+        }
+        let ms = self.origin.elapsed().as_millis().min(u32::MAX as u128) as u32;
+        let mut stages = self.stages.lock().unwrap();
+        if stages.iter().any(|(s, _)| *s == stage) {
+            return;
+        }
+        stages.push((stage, ms));
+    }
+
+    /// Stamp the final stage and emit the one-line summary (first call only). The final stage's
+    /// offset is the transition total, stored into the shared slot.
+    pub(crate) fn finish(&self, stage: &'static str) {
+        if self.finished.swap(true, Ordering::Relaxed) {
+            return;
+        }
+        let total = self.origin.elapsed().as_millis().min(u32::MAX as u128) as u32;
+        let mut stages = self.stages.lock().unwrap();
+        stages.push((stage, total));
+        let line = stages
+            .iter()
+            .map(|(s, ms)| format!("{s}+{ms}"))
+            .collect::<Vec<_>>()
+            .join(" ");
+        drop(stages);
+        self.total_ms.store(total.max(1), Ordering::Relaxed);
+        tracing::info!(
+            kind = self.kind,
+            total_ms = total,
+            stages = %line,
+            "session-transition trace"
+        );
+    }
+}

crates/punktfunk-host/src/main.rs

@@ -19,6 +19,7 @@
 #![deny(clippy::undocumented_unsafe_blocks)]
 
 mod audio;
+mod bringup;
 mod capture;
 mod detect;
 mod devtest;

crates/punktfunk-host/src/mgmt/host.rs

@@ -134,6 +134,12 @@ pub(crate) struct StreamInfo {
     /// Client's parity floor per FEC block (`minRequiredFecPackets`).
     min_fec: u8,
     codec: ApiCodec,
+    /// Session bring-up total, hello → first video packet, in ms (native sessions; `null` on the
+    /// GameStream plane or while the session is still bringing up).
+    time_to_first_frame_ms: Option<u32>,
+    /// Most recent mid-stream resize total, reconfigure → pipeline rebuilt, in ms (native sessions;
+    /// `null` when no resize happened / GameStream).
+    last_resize_ms: Option<u32>,
 }
 
 /// Non-sensitive host status for the local tray icon: counts and booleans only — no PIN values,
@@ -332,6 +338,9 @@ pub(crate) async fn get_status(State(st): State<Arc<MgmtState>>) -> Json<Runtime
             packet_size: c.packet_size as u32,
             min_fec: c.min_fec,
             codec: c.codec.into(),
+            // Transition latencies are traced on the native plane only (latency plan P0.1).
+            time_to_first_frame_ms: None,
+            last_resize_ms: None,
         })
         .or_else(|| {
             native.first().map(|s| StreamInfo {
@@ -344,6 +353,9 @@ pub(crate) async fn get_status(State(st): State<Arc<MgmtState>>) -> Json<Runtime
                 packet_size: 0,
                 min_fec: 0,
                 codec: s.codec.into(),
+                time_to_first_frame_ms: (s.time_to_first_frame_ms > 0)
+                    .then_some(s.time_to_first_frame_ms),
+                last_resize_ms: (s.last_resize_ms > 0).then_some(s.last_resize_ms),
             })
         });
     Json(RuntimeStatus {

crates/punktfunk-host/src/native.rs

@@ -777,14 +777,56 @@ async fn serve_session(
     let source = opts.source;
     let frames = opts.frames;
     let data_port = opts.data_port;
-    let (hello, welcome, udp_port, data_sock, direct, start, compositor) = tokio::time::timeout(
+    // Session-transition trace (latency plan P0.1): zeroed here — the Hello is in hand, pairing
-        HANDSHAKE_TIMEOUT,
+    // gates are behind us — and finished by the send thread when the FIRST video packet leaves.
-        handshake::negotiate(
+    // The completed totals surface per session in `session_status` (→ mgmt `/status`).
-            &conn, &mut send, &mut recv, &first, source, frames, data_port,
+    let bringup = crate::bringup::Trace::start("bringup", Arc::new(AtomicU32::new(0)));
-        ),
+    // The mid-stream resize counterpart: each accepted Reconfigure runs its own trace into this
-    )
+    // shared slot (latest wins), registered alongside the bring-up total.
-    .await
+    let resize_ms: Arc<AtomicU32> = Arc::new(AtomicU32::new(0));
-    .map_err(|_| anyhow!("handshake timed out after {HANDSHAKE_TIMEOUT:?}"))??;
+
+    // Stop signal: stream duration elapsed or the client went away. Created (with its watcher)
+    // BEFORE the handshake so the Welcome-time display prep can already observe a client that
+    // vanished mid-handshake (its build-retry loop aborts on `stop`).
+    let stop = Arc::new(AtomicBool::new(false));
+    // Deliberate-quit signal: set (before `stop`, so the display lease reads it on teardown) when
+    // the client closed the connection with `QUIT_CODE` — a user "stop", which skips the
+    // keep-alive linger. A bare disconnect / idle timeout leaves it false → the display lingers
+    // for a reconnect.
+    let quit = Arc::new(AtomicBool::new(false));
+    {
+        let stop = stop.clone();
+        let quit = quit.clone();
+        let conn = conn.clone();
+        tokio::spawn(async move {
+            let reason = conn.closed().await;
+            if matches!(&reason, quinn::ConnectionError::ApplicationClosed(ac)
+                if ac.error_code == quinn::VarInt::from_u32(QUIT_CODE))
+            {
+                quit.store(true, Ordering::SeqCst);
+            }
+            stop.store(true, Ordering::SeqCst);
+        });
+    }
+
+    let (hello, welcome, udp_port, data_sock, direct, start, compositor, prep) =
+        tokio::time::timeout(
+            HANDSHAKE_TIMEOUT,
+            handshake::negotiate(
+                &conn,
+                &mut send,
+                &mut recv,
+                &first,
+                source,
+                frames,
+                data_port,
+                &bringup,
+                quit.clone(),
+                stop.clone(),
+            ),
+        )
+        .await
+        .map_err(|_| anyhow!("handshake timed out after {HANDSHAKE_TIMEOUT:?}"))??;
     let (ctrl_send, ctrl_recv) = (send, recv);
     // Can this session's backend live-reconfigure (mid-stream Reconfigure)? Gated OFF for:
     //   * gamescope (all sub-modes): a spawn respawn restarts the game, managed restarts the box's
@@ -949,12 +991,8 @@ async fn serve_session(
         );
     });
 
-    // Stop signal: stream duration elapsed or the client went away.
+    // (The stop/quit flags + their disconnect watcher are created above, before the handshake, so
-    let stop = Arc::new(AtomicBool::new(false));
+    // the Welcome-time display prep can observe a mid-handshake disconnect.)
-    // Deliberate-quit signal: set (before `stop`, so the display lease reads it on teardown) when the
-    // client closed the connection with `QUIT_CODE` — a user "stop", which skips the keep-alive linger.
-    // A bare disconnect / idle timeout leaves it false → the display lingers for a reconnect.
-    let quit = Arc::new(AtomicBool::new(false));
     // Lifecycle events (RFC §4): this point — handshake complete, pairing/admission passed — is
     // where the client counts as CONNECTED; the close watcher below pairs it with the
     // disconnect + its decoded reason. A client rejected earlier never emits either.
@@ -967,17 +1005,9 @@ async fn serve_session(
         client: event_client.clone(),
     });
     {
-        let stop = stop.clone();
-        let quit = quit.clone();
         let conn = conn.clone();
         tokio::spawn(async move {
             let reason = conn.closed().await;
-            if matches!(&reason, quinn::ConnectionError::ApplicationClosed(ac)
-                if ac.error_code == quinn::VarInt::from_u32(QUIT_CODE))
-            {
-                quit.store(true, Ordering::SeqCst);
-            }
-            stop.store(true, Ordering::SeqCst);
             let why = match &reason {
                 quinn::ConnectionError::ApplicationClosed(ac)
                     if ac.error_code == quinn::VarInt::from_u32(QUIT_CODE) =>
@@ -1175,6 +1205,10 @@ async fn serve_session(
     let client_label = endpoint::peer_fingerprint(&conn)
         .map(|fp| fingerprint_hex(&fp)[..12].to_string())
         .unwrap_or_else(|| conn.remote_address().ip().to_string());
+    // Transition-trace handles for the data plane (P0.1): the punch stamp + the virtual-stream
+    // stages ride the same per-session trace; resizes write their totals into the shared slot.
+    let bringup_dp = bringup.clone();
+    let resize_ms_dp = resize_ms.clone();
     let result: Result<()> = async {
         tokio::task::spawn_blocking(move || -> Result<()> {
             // Bring up the (already-bound) data-plane socket. Default: hole-punch — wait briefly
@@ -1204,6 +1238,7 @@ async fn serve_session(
                     return Err(anyhow::Error::new(e)).context("bind data plane");
                 }
             };
+            bringup_dp.mark("punch_done");
             tracing::info!(
                 %client_udp,
                 udp_port,
@@ -1229,7 +1264,7 @@ async fn serve_session(
                 Punktfunk1Source::Virtual => {
                     let compositor = compositor
                         .expect("the Virtual source resolves a compositor during the handshake");
-                    virtual_stream(SessionContext {
+                    let ctx = SessionContext {
                         session,
                         mode,
                         seconds,
@@ -1256,7 +1291,29 @@ async fn serve_session(
                         client_label,
                         launch: launch_for_dp,
                         client_hdr,
-                    })
+                        bringup: bringup_dp,
+                        resize_ms: resize_ms_dp,
+                    };
+                    match prep {
+                        // P1.1: the display prep started at Welcome on its own thread — hand it
+                        // the post-punch context and adopt its result as the stream result (that
+                        // thread runs `virtual_stream` on the pipeline it already built).
+                        Some((ctx_tx, prep_thread)) => match ctx_tx.send(ctx) {
+                            Ok(()) => match prep_thread.join() {
+                                Ok(r) => r,
+                                Err(_) => Err(anyhow!("prepared stream thread panicked")),
+                            },
+                            // The prep thread died before the hand-off (panicked during prep —
+                            // its guard/lease unwound): run the stream inline instead.
+                            Err(std::sync::mpsc::SendError(ctx)) => {
+                                tracing::warn!(
+                                    "display-prep thread gone before hand-off — building inline"
+                                );
+                                virtual_stream(ctx, None)
+                            }
+                        },
+                        None => virtual_stream(ctx, None),
+                    }
                 }
             }
         })

crates/punktfunk-host/src/native/handshake.rs

@@ -19,6 +19,14 @@ pub(super) async fn negotiate(
     source: Punktfunk1Source,
     frames: u32,
     data_port: Option<u16>,
+    // Session bring-up trace (latency plan P0.1): `welcome`/`start` stamps land here, and the
+    // Welcome-time display prep threads it into the pipeline-build stages.
+    bringup: &Arc<crate::bringup::Trace>,
+    // The session's quit/stop flags — created BEFORE the handshake so the Welcome-time display
+    // prep below can observe a client that vanished mid-handshake (its build retry aborts on
+    // `stop`; `quit` rides into the display lease).
+    quit: Arc<AtomicBool>,
+    stop: Arc<AtomicBool>,
 ) -> Result<(
     Hello,
     Welcome,
@@ -27,6 +35,7 @@ pub(super) async fn negotiate(
     bool,
     Start,
     Option<crate::vdisplay::Compositor>,
+    Option<super::stream::PrepHandle>,
 )> {
     let peer = conn.remote_address();
     let mut hello = Hello::decode(first).map_err(|e| anyhow!("Hello decode: {e:?}"))?;
@@ -377,10 +386,74 @@ pub(super) async fn negotiate(
             },
     };
     io::write_msg(send, &welcome.encode()).await?;
+    bringup.mark("welcome");
+
+    // P1.1/P1.2 (latency plan): kick the display prep NOW — the negotiated mode is final in
+    // the Welcome just sent, and nothing in monitor create → activation → settle → capture
+    // attach → encoder open needs the client's Start or the punched socket. The prep thread
+    // BECOMES the stream thread: the data plane hands it the post-punch SessionContext and it
+    // runs `virtual_stream` on the warm pipeline, so the whole display bring-up hides behind
+    // the Start RTT + the (up to 2.5 s) hole-punch wait. If the session dies before its data
+    // plane comes up (handshake timeout, client vanished), the channel drops and the prep
+    // result is released — the monitor lands in the keep-alive machinery exactly like a
+    // normal session end (and `stop`, watched by the caller, aborts a still-running build
+    // retry). Windows native path only: the Linux backends bind launch semantics before create
+    // (gamescope nests the launch command), which must not run for a client that never sends
+    // Start; GameStream has neither a Start gate nor a punch.
+    #[cfg(target_os = "windows")]
+    let prep: Option<super::stream::PrepHandle> = match (source, compositor) {
+        (Punktfunk1Source::Virtual, Some(comp)) => {
+            let (ctx_tx, ctx_rx) = std::sync::mpsc::sync_channel::<SessionContext>(1);
+            let client_identity = endpoint::peer_fingerprint(conn);
+            let client_hdr = hello.display_hdr;
+            let (mode, shard_payload) = (hello.mode, welcome.shard_payload);
+            let trace = bringup.clone();
+            std::thread::Builder::new()
+                .name("punktfunk1-stream".into())
+                .spawn(move || -> Result<()> {
+                    let prepared = super::stream::prepare_display(
+                        comp,
+                        mode,
+                        client_identity,
+                        client_hdr,
+                        bitrate_kbps,
+                        bit_depth,
+                        chroma,
+                        codec,
+                        shard_payload,
+                        &quit,
+                        &stop,
+                        &trace,
+                    );
+                    let Ok(ctx) = ctx_rx.recv() else {
+                        // No data plane ever came (handshake abort / punch failure): drop
+                        // `prepared` — its lease release hands the monitor to keep-alive
+                        // policy, exactly like a normal session end.
+                        return Ok(());
+                    };
+                    match prepared {
+                        Ok(p) => virtual_stream(ctx, Some(p)),
+                        Err(e) => Err(e),
+                    }
+                })
+                .map(|handle| (ctx_tx, handle))
+                .map_err(|e| {
+                    tracing::warn!(error = %e,
+                        "display-prep thread spawn failed — falling back to inline bring-up")
+                })
+                .ok()
+        }
+        _ => None,
+    };
+    #[cfg(not(target_os = "windows"))]
+    let prep: Option<super::stream::PrepHandle> = None;
+    #[cfg(not(target_os = "windows"))]
+    let _ = (quit, stop);
 
     let start =
         Start::decode(&io::read_msg(recv).await?).map_err(|e| anyhow!("Start decode: {e:?}"))?;
+    bringup.mark("start");
     Ok::<_, anyhow::Error>((
-        hello, welcome, udp_port, data_sock, direct, start, compositor,
+        hello, welcome, udp_port, data_sock, direct, start, compositor, prep,
     ))
 }

crates/punktfunk-host/src/native/stream.rs

@@ -310,6 +310,10 @@ struct SendStats {
     /// Live encoder bitrate (kbps) — the capture thread updates it on a mid-stream adaptive
     /// bitrate change, so the web-console sample reports what the encoder is ACTUALLY targeting.
     bitrate_kbps: Arc<AtomicU32>,
+    /// The session's bring-up trace (P0.1): the send thread FINISHES it — `first_packet` — the
+    /// moment the first video AU's packets have fully left the socket (finish is once-only, so
+    /// the per-frame call is a cheap no-op afterwards).
+    bringup: Arc<crate::bringup::Trace>,
 }
 
 /// Whether a session on `compositor` (`None` = the synthetic source) with a `per_client_mode`
@@ -389,6 +393,11 @@ fn send_loop(
                 burst_cap,
             ) {
                 Ok(stat) => {
+                    // First VIDEO packets are on the wire — complete the bring-up trace (P0.1;
+                    // once-only, no-op on every later frame). Speed-test filler isn't video.
+                    if msg.flags & FLAG_PROBE as u32 == 0 {
+                        stats.bringup.finish("first_packet");
+                    }
                     // Host timing (0xCF): stamped now — the AU's packets have fully left the
                     // socket — against the same capture anchor the wire pts carries, so the
                     // client's per-frame math tiles exactly (network = its host+network − this).
@@ -743,9 +752,15 @@ pub(super) struct SessionContext {
     /// so host apps tone-map to the client's real panel) and preferred over the generic baseline
     /// for the 0xCE mastering metadata.
     pub(super) client_hdr: Option<punktfunk_core::quic::HdrMeta>,
+    /// The session's bring-up trace (latency plan P0.1): the pipeline-build stages stamp into it
+    /// and the send thread finishes it when the first video packet leaves.
+    pub(super) bringup: Arc<crate::bringup::Trace>,
+    /// Shared slot the latest completed mid-stream resize total (ms) lands in — registered with
+    /// `session_status` so the Dashboard shows it.
+    pub(super) resize_ms: Arc<AtomicU32>,
 }
 
-pub(super) fn virtual_stream(ctx: SessionContext) -> Result<()> {
+pub(super) fn virtual_stream(ctx: SessionContext, prepared: Option<PreparedDisplay>) -> Result<()> {
     // This thread runs the capture+encode loop (single-process — the only topology: Linux portal /
     // synthetic, Windows in-process IDD-push). Elevate it so a CPU-heavy game can't deschedule our GPU
     // submission.
@@ -792,6 +807,8 @@ pub(super) fn virtual_stream(ctx: SessionContext) -> Result<()> {
         client_label,
         launch,
         client_hdr,
+        bringup,
+        resize_ms,
     } = ctx;
     tracing::info!(
         compositor = compositor.id(),
@@ -800,54 +817,79 @@ pub(super) fn virtual_stream(ctx: SessionContext) -> Result<()> {
         bit_depth,
         "punktfunk/1 virtual display"
     );
-    // Open the backend FIRST — on Windows this constructs the vdisplay backend, which initialises the
+    // The vdisplay backend + built pipeline: either PREPARED at Welcome time on this very thread
-    // host-lifetime VirtualDisplayManager (§2.5). It does NO monitor work, so it must precede the IDD-push
+    // (P1.1/P1.2 — the display bring-up already overlapped the Start RTT + hole-punch), or built
-    // preempt below (which reaches the manager) — otherwise `vdm()` is called before init and panics.
+    // inline now (Linux, synthetic-adjacent paths, prep fallback).
-    let mut vd = crate::vdisplay::open(compositor)?;
+    let (mut vd, pipe) = match prepared {
-    // Per-client STABLE monitor identity (Phase 2): hand the backend the connecting client's cert
+        Some(p) => (p.vd, p.pipeline),
-    // fingerprint so a freshly CREATED virtual monitor gets this client's persistent id — Windows then
+        None => {
-    // reapplies the client's saved per-monitor config (DPI scaling) on reconnect. No-op on Linux backends
+            // Open the backend FIRST — on Windows this constructs the vdisplay backend, which
-    // and for anonymous/GameStream clients (no fingerprint → the driver auto-allocates).
+            // initialises the host-lifetime VirtualDisplayManager (§2.5). It does NO monitor work,
-    vd.set_client_identity(endpoint::peer_fingerprint(&conn));
+            // so it must precede the IDD-push preempt below (which reaches the manager) —
-    // The client display's HDR volume (Hello) → a freshly created virtual monitor's EDID CTA HDR
+            // otherwise `vdm()` is called before init and panics.
-    // block (pf-vdisplay), so host apps + the OS tone-map to the client's real panel instead of the
+            let mut vd = crate::vdisplay::open(compositor)?;
-    // driver's built-in ~1000-nit placeholder. No-op on Linux backends and for older/SDR clients.
+            // Per-client STABLE monitor identity (Phase 2): hand the backend the connecting
-    vd.set_client_hdr(client_hdr);
+            // client's cert fingerprint so a freshly CREATED virtual monitor gets this client's
-    // Deliberate-quit wiring (Windows pf-vdisplay; no-op elsewhere): every lease the backend mints —
+            // persistent id — Windows then reapplies the client's saved per-monitor config (DPI
-    // the retry-hold below AND the capturer's — carries the session's quit flag, so a user "stop"
+            // scaling) on reconnect. No-op on Linux backends and for anonymous/GameStream clients
-    // (⌘D → the QUIT close code) tears the virtual monitor down the moment the pipeline drops instead
+            // (no fingerprint → the driver auto-allocates).
-    // of lingering 10 s. The reconnect then finds the manager Idle and does a clean fresh ADD (with
+            vd.set_client_identity(endpoint::peer_fingerprint(&conn));
-    // the user's think-time as driver settle) rather than the Lingering-preempt's REMOVE→ADD churn.
+            // The client display's HDR volume (Hello) → a freshly created virtual monitor's EDID
-    // `keep_alive = forever` (gaming-rig) outranks the quit — the monitor pins as before.
+            // CTA HDR block (pf-vdisplay), so host apps + the OS tone-map to the client's real
-    vd.set_quit_flag(quit.clone());
+            // panel instead of the driver's built-in ~1000-nit placeholder. No-op on Linux
-    // Per-session launch (non-Windows): hand the resolved command to the backend instance so
+            // backends and for older/SDR clients.
-    // gamescope's bare spawn nests it — per-instance, no process-global env, so concurrent sessions
+            vd.set_client_hdr(client_hdr);
-    // can't stomp each other's launch target. The other backends' default `set_launch_command` is a
+            // Deliberate-quit wiring (Windows pf-vdisplay; no-op elsewhere): every lease the
-    // no-op; they get the command spawned into the live session after capture is up (below).
+            // backend mints — the retry-hold below AND the capturer's — carries the session's quit
-    #[cfg(not(target_os = "windows"))]
+            // flag, so a user "stop" (⌘D → the QUIT close code) tears the virtual monitor down the
-    vd.set_launch_command(launch.clone());
+            // moment the pipeline drops instead of lingering 10 s. The reconnect then finds the
-    // IDD-push reconnect preempt (the dance now lives in the manager, Goal-1 §2.5): serialize setup so a
+            // manager Idle and does a clean fresh ADD (with the user's think-time as driver
-    // reconnect FLOOD can't run concurrent monitor create/teardown, STOP the prior session + WAIT for it
+            // settle) rather than the Lingering-preempt's REMOVE→ADD churn. `keep_alive = forever`
-    // to release its monitor (instead of tearing a monitor out from under a still-live session), and
+            // (gaming-rig) outranks the quit — the monitor pins as before.
-    // register THIS session's stop. The returned guard holds the setup lock across the pipeline build;
+            vd.set_quit_flag(quit.clone());
-    // dropping it lets the next reconnect begin (and preempt us). Held BEFORE the monitor is created
+            // Per-session launch (non-Windows): hand the resolved command to the backend instance
-    // (build_pipeline → vd.create), so the preempt still precedes this session's monitor creation.
+            // so gamescope's bare spawn nests it — per-instance, no process-global env, so
-    // SLOT-scoped (Stage W1): the preempt targets only a prior session holding THIS client's slot —
+            // concurrent sessions can't stomp each other's launch target. The other backends'
-    // a different identity's session is an admission question, never a preempt.
+            // default `set_launch_command` is a no-op; they get the command spawned into the live
-    #[cfg(target_os = "windows")]
+            // session after capture is up (below).
-    let _idd_setup_guard =
+            #[cfg(not(target_os = "windows"))]
-        (plan.capture == crate::session_plan::CaptureBackend::IddPush).then(|| {
+            vd.set_launch_command(launch.clone());
-            let slot = crate::vdisplay::manager::slot_id_for(
+            // IDD-push reconnect preempt (the dance now lives in the manager, Goal-1 §2.5):
-                endpoint::peer_fingerprint(&conn),
+            // serialize setup so a reconnect FLOOD can't run concurrent monitor create/teardown,
-                (mode.width, mode.height),
+            // STOP the prior session + WAIT for it to release its monitor (instead of tearing a
-            );
+            // monitor out from under a still-live session), and register THIS session's stop. The
-            crate::vdisplay::manager::vdm().begin_idd_setup(slot, stop.clone())
+            // returned guard holds the setup lock across the pipeline build; dropping it (end of
-        });
+            // this arm) lets the next reconnect begin (and preempt us). Held BEFORE the monitor is
+            // created (build_pipeline → vd.create), so the preempt still precedes this session's
+            // monitor creation. SLOT-scoped (Stage W1): the preempt targets only a prior session
+            // holding THIS client's slot — a different identity's session is an admission
+            // question, never a preempt.
+            #[cfg(target_os = "windows")]
+            let _idd_setup_guard = (plan.capture == crate::session_plan::CaptureBackend::IddPush)
+                .then(|| {
+                    let slot = crate::vdisplay::manager::slot_id_for(
+                        endpoint::peer_fingerprint(&conn),
+                        (mode.width, mode.height),
+                    );
+                    crate::vdisplay::manager::vdm().begin_idd_setup(slot, stop.clone())
+                });
+            let pipe = build_pipeline_with_retry(
+                &mut vd,
+                mode,
+                bitrate_kbps,
+                bit_depth,
+                plan,
+                &quit,
+                &stop,
+                Some(bringup.as_ref()),
+            )?;
+            // Setup done — the IDD-push setup lock releases as the guard leaves this arm's scope,
+            // so the next reconnect can begin (and preempt us).
+            (vd, pipe)
+        }
+    };
     let (mut capturer, mut enc, mut frame, mut interval, mut cur_node_id, mut cur_display_gen) =
-        build_pipeline_with_retry(&mut vd, mode, bitrate_kbps, bit_depth, plan, &quit, &stop)?;
+        pipe;
-    // Setup done — release the IDD-push setup lock so the next reconnect can begin (and preempt us).
-    #[cfg(target_os = "windows")]
-    drop(_idd_setup_guard);
 
     // Capture is live — launch the requested title so it renders onto the streamed output and
     // grabs focus. Windows spawns the library id into the interactive user session; Linux spawns
@@ -914,6 +956,7 @@ pub(super) fn virtual_stream(ctx: SessionContext) -> Result<()> {
         codec: plan.codec.label(),
         client: client_label.clone(),
         bitrate_kbps: live_bitrate.clone(),
+        bringup: bringup.clone(),
     };
     let send_thread = std::thread::Builder::new()
         .name("punktfunk-send".into())
@@ -949,6 +992,8 @@ pub(super) fn virtual_stream(ctx: SessionContext) -> Result<()> {
         force_idr.clone(),
         client_label,
         plan.hdr,
+        bringup.total_slot(),
+        resize_ms.clone(),
     );
 
     // Mid-stream session-switch watcher (opt-in via PUNKTFUNK_SESSION_WATCH; never under an explicit
@@ -1081,6 +1126,7 @@ pub(super) fn virtual_stream(ctx: SessionContext) -> Result<()> {
                             plan,
                             &quit,
                             &stop,
+                            None,
                         )?;
                         Ok((new_vd, pipe))
                     })();
@@ -1131,6 +1177,10 @@ pub(super) fn virtual_stream(ctx: SessionContext) -> Result<()> {
         }
         if let Some(new_mode) = want {
             tracing::info!(?new_mode, "rebuilding pipeline for mode switch");
+            // Resize trace (P0.1): reconfigure-received → pipeline rebuilt (incl. the first
+            // new-mode frame — `build_pipeline` waits for it). Total lands in the shared
+            // `resize_ms` slot (→ `session_status`); a failed rebuild abandons it silently.
+            let resize_trace = crate::bringup::Trace::start("resize", resize_ms.clone());
             // PyroWave's Automatic bitrate is a per-mode ~1.6 bpp pin (resolve_bitrate_kbps_for) —
             // a resolution change moves the operating point (1080p→4K quadruples the pixel rate),
             // so re-resolve it for the new mode. Explicit client rates stay put (the operator knows
@@ -1140,72 +1190,111 @@ pub(super) fn virtual_stream(ctx: SessionContext) -> Result<()> {
             } else {
                 bitrate_kbps
             };
-            // Build the new pipeline BEFORE dropping the old one: the host already acked
+            // IN-PLACE fast path first (latency plan P2.3, Windows IDD-push): keep the capturer +
-            // the switch as accepted, so a rebuild failure must not kill an otherwise
+            // send thread, mode-set the SAME monitor in place (P2.1/P2.2), resize the ring, swap
+            // only the encoder. Any decline (v3 driver → the manager re-arrived, ring recreate
+            // failed, no new-size frame) falls through to the full rebuild below.
+            #[cfg(target_os = "windows")]
+            let fast_done = plan.capture == crate::session_plan::CaptureBackend::IddPush
+                && try_inplace_resize(
+                    &mut vd,
+                    &mut capturer,
+                    &mut enc,
+                    &mut frame,
+                    &mut interval,
+                    new_mode,
+                    mode_bitrate,
+                    bit_depth,
+                    plan,
+                    &quit,
+                    resize_trace.as_ref(),
+                );
+            #[cfg(not(target_os = "windows"))]
+            let fast_done = false;
+            // Full rebuild — build the new pipeline BEFORE dropping the old one: the host already
+            // acked the switch as accepted, so a rebuild failure must not kill an otherwise
             // healthy session — keep streaming the current mode and log instead.
-            match build_pipeline(&mut vd, new_mode, mode_bitrate, bit_depth, plan, &quit) {
+            let rebuilt = fast_done
-                Ok(next_pipe) => {
+                || match build_pipeline(
-                    if mode_bitrate != bitrate_kbps {
+                    &mut vd,
-                        tracing::info!(
+                    new_mode,
-                            from_kbps = bitrate_kbps,
+                    mode_bitrate,
-                            to_kbps = mode_bitrate,
+                    bit_depth,
-                            "pinned PyroWave bitrate re-resolved for the new mode"
+                    plan,
-                        );
+                    &quit,
-                        bitrate_kbps = mode_bitrate;
+                    Some(resize_trace.as_ref()),
-                        live_bitrate.store(mode_bitrate, Ordering::Relaxed);
+                ) {
+                    Ok(next_pipe) => {
+                        let old_display_gen = cur_display_gen;
+                        // The destructuring assignment drops the OLD capturer (→ its display lease)
+                        // as each binding is replaced — the new pipeline is already up
+                        // (create-before-drop).
+                        (capturer, enc, frame, interval, cur_node_id, cur_display_gen) = next_pipe;
+                        // H4: the old display's lease drop above is indistinguishable from a
+                        // disconnect to the keep-alive machinery — under linger/forever policies
+                        // every resize would ACCUMULATE kept monitors at stale modes. Retire the
+                        // superseded entry now (a no-op when it was already torn down under
+                        // `immediate`, or off Linux; the in-place fast path keeps the SAME display,
+                        // so it has nothing to retire).
+                        if let Some(g) = old_display_gen.filter(|g| cur_display_gen != Some(*g)) {
+                            crate::vdisplay::registry::retire(g);
+                        }
+                        true
                     }
-                    let old_display_gen = cur_display_gen;
+                    Err(e) => {
-                    // The destructuring assignment drops the OLD capturer (→ its display lease) as
+                        tracing::warn!(error = %format!("{e:#}"), ?new_mode,
-                    // each binding is replaced — the new pipeline is already up (create-before-drop).
+                            "mode-switch rebuild failed — staying on the current mode");
-                    (capturer, enc, frame, interval, cur_node_id, cur_display_gen) = next_pipe;
+                        // H2 rollback: the control task acked the switch BEFORE this rebuild, so the
-                    cur_mode = new_mode;
+                        // client's mode slot already flipped to `new_mode`. A second accepted ack
-                    next = std::time::Instant::now();
+                        // carrying the still-live mode corrects it (any accepted ack means "the
-                    // H4: the old display's lease drop above is indistinguishable from a disconnect
+                        // active mode is now X" client-side; old clients just log it). `frame` is
-                    // to the keep-alive machinery — under linger/forever policies every resize would
+                        // untouched here (the fast path returned false before swapping anything and
-                    // ACCUMULATE kept monitors at stale modes. Retire the superseded entry now (a
+                        // the destructure only runs on the Ok arm), so it's still the OLD
-                    // no-op when it was already torn down under `immediate`, or off Linux).
+                        // pipeline's frame — its real dims + interval are what's still on glass.
-                    if let Some(g) = old_display_gen.filter(|g| cur_display_gen != Some(*g)) {
-                        crate::vdisplay::registry::retire(g);
-                    }
-                    // H2/H3: the backend may have honored a different mode than requested — KWin
-                    // caps a virtual output's refresh, or Windows pf-vdisplay rejects an in-place
-                    // SetMode to a resolution its running monitor doesn't advertise and the host
-                    // falls back to the actual display mode. `frame` is the NEW pipeline's first
-                    // frame (just rebound above), so its dims are what the client actually decodes.
-                    // Publish that ACTUAL mode to the live stats slot, and correct the client's mode
-                    // slot when it differs from the accept ack it already got.
-                    let actual = delivered_mode(frame.width, frame.height, interval);
-                    live_mode.store(
-                        pack_mode(actual.width, actual.height, actual.refresh_hz),
-                        Ordering::Relaxed,
-                    );
-                    if actual != new_mode {
                         let _ = reconfig_result_tx.send(Reconfigured {
                             accepted: true,
-                            mode: actual,
+                            mode: delivered_mode(frame.width, frame.height, interval),
                         });
+                        false
                     }
-                    // The owed AUs died with the old encoder — drop their in-flight records
+                };
-                    // and restart the encode-stall clock for the fresh one.
+            if rebuilt {
-                    inflight.clear();
+                if mode_bitrate != bitrate_kbps {
-                    last_au_at = std::time::Instant::now();
+                    tracing::info!(
-                    encoder_resets = 0;
+                        from_kbps = bitrate_kbps,
-                    last_forced_idr = Some(std::time::Instant::now()); // fresh encoder opens on an IDR — anchor the cooldown
+                        to_kbps = mode_bitrate,
+                        "pinned PyroWave bitrate re-resolved for the new mode"
+                    );
+                    bitrate_kbps = mode_bitrate;
+                    live_bitrate.store(mode_bitrate, Ordering::Relaxed);
                 }
-                Err(e) => {
+                cur_mode = new_mode;
-                    tracing::warn!(error = %format!("{e:#}"), ?new_mode,
+                next = std::time::Instant::now();
-                        "mode-switch rebuild failed — staying on the current mode");
+                // H2/H3: the backend may have honored a different mode than requested — KWin caps
-                    // H2 rollback: the control task acked the switch BEFORE this rebuild, so the
+                // a virtual output's refresh, or Windows pf-vdisplay rejects a resolution its
-                    // client's mode slot already flipped to `new_mode`. A second accepted ack
+                // running monitor doesn't advertise and the host falls back to the actual display
-                    // carrying the still-live mode corrects it (any accepted ack means "the active
+                // mode. `frame` is the NEW pipeline's first frame (just rebound above), so its
-                    // mode is now X" client-side; old clients just log it). `frame` is untouched
+                // dims are what the client actually decodes. Publish that ACTUAL mode to the live
-                    // here (the destructure only runs on the Ok arm), so it's still the OLD
+                // stats slot, and correct the client's mode slot when it differs from the accept
-                    // pipeline's frame — its real dims + interval are exactly what's still on glass.
+                // ack it already got.
+                let actual = delivered_mode(frame.width, frame.height, interval);
+                live_mode.store(
+                    pack_mode(actual.width, actual.height, actual.refresh_hz),
+                    Ordering::Relaxed,
+                );
+                if actual != new_mode {
                     let _ = reconfig_result_tx.send(Reconfigured {
                         accepted: true,
-                        mode: delivered_mode(frame.width, frame.height, interval),
+                        mode: actual,
                     });
                 }
+                // The owed AUs died with the old encoder — drop their in-flight records
+                // and restart the encode-stall clock for the fresh one.
+                inflight.clear();
+                last_au_at = std::time::Instant::now();
+                encoder_resets = 0;
+                last_forced_idr = Some(std::time::Instant::now()); // fresh encoder opens on an IDR — anchor the cooldown
+                resize_trace.finish("pipeline_rebuilt");
             }
         }
         // Adaptive bitrate: drain to the NEWEST requested rate (the client's controller may step
@@ -1499,6 +1588,7 @@ pub(super) fn virtual_stream(ctx: SessionContext) -> Result<()> {
                         plan,
                         &quit,
                         &stop,
+                        None,
                     ) {
                         Ok(p) => break p,
                         Err(e2) => {
@@ -1743,6 +1833,7 @@ pub(super) fn virtual_stream(ctx: SessionContext) -> Result<()> {
             };
             // Hand to the send thread; this blocks (backpressure) if it's behind. An Err means it
             // exited (send failure / stop) — end the encode loop too.
+            bringup.mark("first_au"); // P0.1 (first-crossing only; free afterwards)
             if frame_tx.send(msg).is_err() {
                 send_gone = true;
                 break;
@@ -1849,6 +1940,191 @@ type Pipeline = (
     Option<u64>,
 );
 
+/// The in-place resize fast path (latency plan P2.3, Windows IDD-push): the manager mode-sets the
+/// SAME monitor in place (driver protocol v4 — `IOCTL_UPDATE_MODES`; internally falls back to
+/// re-arrival against an older driver), then the existing capturer re-sizes its ring immediately
+/// (no descriptor-poll debounce) and only the ENCODER is swapped once the first new-size frame
+/// arrives — the capture pipeline, its send thread and the whole session transport survive.
+/// Returns `true` when the stream is now delivering the new mode on the same capturer; `false`
+/// routes the caller to the full rebuild (which is also the correct path when the manager had to
+/// re-arrive a fresh monitor — this capturer's ring/broker are bound to the departed target).
+#[cfg(target_os = "windows")]
+#[allow(clippy::too_many_arguments)]
+fn try_inplace_resize(
+    vd: &mut Box<dyn crate::vdisplay::VirtualDisplay>,
+    capturer: &mut Box<dyn crate::capture::Capturer>,
+    enc: &mut Box<dyn crate::encode::Encoder>,
+    frame: &mut crate::capture::CapturedFrame,
+    interval: &mut std::time::Duration,
+    new_mode: punktfunk_core::Mode,
+    bitrate_kbps: u32,
+    bit_depth: u8,
+    plan: crate::session_plan::SessionPlan,
+    quit: &Arc<AtomicBool>,
+    trace: &crate::bringup::Trace,
+) -> bool {
+    let Some(cur_target) = capturer.capture_target_id() else {
+        return false; // not an IDD-push capturer — nothing to reuse
+    };
+    // Acquire at the new mode: the manager's resize branch runs the in-place mode set (or its
+    // re-arrival fallback) and returns a +1-ref lease, released again when `vout` drops below —
+    // the capturer keeps holding its own original lease (`gen` is preserved by both paths).
+    let vout = match crate::vdisplay::registry::acquire(vd, new_mode, quit.clone()) {
+        Ok(v) => v,
+        Err(e) => {
+            tracing::warn!(error = %format!("{e:#}"), "in-place resize: acquire failed");
+            return false;
+        }
+    };
+    trace.mark("display_resized");
+    let effective_hz = vout
+        .preferred_mode
+        .map(|(_, _, hz)| hz)
+        .filter(|&hz| hz > 0)
+        .unwrap_or(new_mode.refresh_hz);
+    if vout.win_capture.as_ref().map(|t| t.target_id) != Some(cur_target) {
+        // The manager re-arrived a fresh monitor (old driver / in-place failure): this capturer is
+        // bound to the departed target. The full rebuild re-acquires (JOINing the already-resized
+        // monitor) with a fresh capturer.
+        tracing::info!(
+            "resize: monitor re-arrived (no in-place support) — running the full pipeline rebuild"
+        );
+        return false;
+    }
+    if !capturer.resize_output(new_mode.width, new_mode.height) {
+        return false;
+    }
+    trace.mark("ring_recreated");
+    // Bounded wait for the first frame at the new size (the driver re-attaches to the fresh ring;
+    // the mode-set full redraw composes promptly). Mirrors the capturer's own 3 s recover-or-drop.
+    let deadline = std::time::Instant::now() + std::time::Duration::from_secs(3);
+    let new_frame = loop {
+        match capturer.try_latest() {
+            Ok(Some(f)) if (f.width, f.height) == (new_mode.width, new_mode.height) => break f,
+            Ok(_) => {
+                if std::time::Instant::now() >= deadline {
+                    tracing::warn!(
+                        "resize: no new-size frame within 3s of the in-place mode set — running \
+                         the full pipeline rebuild"
+                    );
+                    return false;
+                }
+                std::thread::sleep(std::time::Duration::from_millis(5));
+            }
+            Err(e) => {
+                tracing::warn!(error = %format!("{e:#}"),
+                    "resize: capture failed after the in-place mode set — running the full rebuild");
+                return false;
+            }
+        }
+    };
+    trace.mark("first_new_frame");
+    // Fresh encoder at the delivered size — the one component that can't follow a resolution
+    // change in place today (P2.4 stays unimplemented: `open_video` is ms-scale, measured).
+    let mut new_enc = match crate::encode::open_video(
+        plan.codec,
+        new_frame.format,
+        new_frame.width,
+        new_frame.height,
+        effective_hz,
+        bitrate_kbps as u64 * 1000,
+        new_frame.is_cuda(),
+        bit_depth,
+        plan.chroma,
+    ) {
+        Ok(e) => e,
+        Err(e) => {
+            tracing::warn!(error = %format!("{e:#}"),
+                "resize: encoder open failed after the in-place mode set — running the full rebuild");
+            return false;
+        }
+    };
+    if let Some(c) = plan.wire_chunk {
+        new_enc.set_wire_chunking(c);
+    }
+    *enc = new_enc;
+    *frame = new_frame;
+    *interval = std::time::Duration::from_secs_f64(1.0 / effective_hz.max(1) as f64);
+    trace.mark("encoder_open");
+    true
+}
+
+/// The Welcome-time display-prep hand-off (latency plan P1.1/P1.2): the opened vdisplay backend +
+/// the fully built pipeline — monitor create, activation, settle, capture attach, first frame,
+/// encoder open — produced on the prep/stream thread while the client's Start round-trip and the
+/// UDP hole-punch are still in flight, so the entire display bring-up hides behind the network
+/// waits. Constructed on the Windows native path only today: the Linux backends bind launch
+/// semantics before create (gamescope nests the launch command), which must not run for a client
+/// that never sends Start.
+pub(super) struct PreparedDisplay {
+    pub(super) vd: Box<dyn crate::vdisplay::VirtualDisplay>,
+    pub(super) pipeline: Pipeline,
+}
+
+/// The prep thread's hand-off pair: the sender delivers the post-punch [`SessionContext`] to the
+/// thread (which then runs [`virtual_stream`] on its prepared display); the join handle returns
+/// the stream result. Dropping the sender un-received aborts the prep cleanly (the prepared
+/// display's lease releases into keep-alive policy).
+pub(super) type PrepHandle = (
+    std::sync::mpsc::SyncSender<SessionContext>,
+    std::thread::JoinHandle<Result<()>>,
+);
+
+/// Build the session's display + pipeline at Welcome time (latency plan P1.1/P1.2), before the
+/// client's `Start` and the hole-punch — the negotiated mode is final once the Welcome is built,
+/// and nothing in monitor create → activation → settle → capture attach → encoder open needs the
+/// punched socket. Mirrors `virtual_stream`'s inline bring-up exactly: same backend setters, same
+/// slot-scoped `begin_idd_setup` serialization (the guard releases when this returns), same
+/// retry-wrapped build. The caller threads the SAME values the Welcome committed, so the prepared
+/// pipeline and the later `SessionContext` can never disagree.
+#[cfg(target_os = "windows")]
+#[allow(clippy::too_many_arguments)]
+pub(super) fn prepare_display(
+    compositor: crate::vdisplay::Compositor,
+    mode: punktfunk_core::Mode,
+    client_identity: Option<[u8; 32]>,
+    client_hdr: Option<punktfunk_core::quic::HdrMeta>,
+    bitrate_kbps: u32,
+    bit_depth: u8,
+    chroma: crate::encode::ChromaFormat,
+    codec: crate::encode::Codec,
+    shard_payload: u16,
+    quit: &Arc<AtomicBool>,
+    stop: &Arc<AtomicBool>,
+    trace: &crate::bringup::Trace,
+) -> Result<PreparedDisplay> {
+    // Same plan resolution as `virtual_stream` (pure in these inputs + host config), including
+    // PyroWave's datagram-aligned wire mode — `Session::shard_payload()` echoes the negotiated
+    // Welcome value passed here.
+    let mut plan = crate::session_plan::SessionPlan::resolve(bit_depth, chroma, codec);
+    if codec == crate::encode::Codec::PyroWave {
+        plan.wire_chunk = Some(shard_payload as usize);
+    }
+    let mut vd = crate::vdisplay::open(compositor)?;
+    vd.set_client_identity(client_identity);
+    vd.set_client_hdr(client_hdr);
+    vd.set_quit_flag(quit.clone());
+    // Slot-scoped setup serialization + reconnect preempt — see the inline arm in
+    // `virtual_stream` for the full rationale; released when this fn returns.
+    let _idd_setup_guard =
+        (plan.capture == crate::session_plan::CaptureBackend::IddPush).then(|| {
+            let slot =
+                crate::vdisplay::manager::slot_id_for(client_identity, (mode.width, mode.height));
+            crate::vdisplay::manager::vdm().begin_idd_setup(slot, stop.clone())
+        });
+    let pipeline = build_pipeline_with_retry(
+        &mut vd,
+        mode,
+        bitrate_kbps,
+        bit_depth,
+        plan,
+        quit,
+        stop,
+        Some(trace),
+    )?;
+    Ok(PreparedDisplay { vd, pipeline })
+}
+
 /// Build the pipeline, retrying *transient* failures with bounded exponential backoff.
 ///
 /// Bringing a virtual output to first-frame races several async steps — the compositor parenting
@@ -1859,6 +2135,7 @@ type Pipeline = (
 /// error chain is classified and permanent ones short-circuit. Each failed attempt drops its
 /// capturer, which (via `PortalCapturer::Drop`) tears the PipeWire thread + virtual output down
 /// before the next attempt — no leak across retries.
+#[allow(clippy::too_many_arguments)]
 fn build_pipeline_with_retry(
     vd: &mut Box<dyn crate::vdisplay::VirtualDisplay>,
     mode: punktfunk_core::Mode,
@@ -1867,6 +2144,9 @@ fn build_pipeline_with_retry(
     plan: crate::session_plan::SessionPlan,
     quit: &Arc<AtomicBool>,
     stop: &Arc<AtomicBool>,
+    // Transition trace (P0.1): `Some` for the traced builds (bring-up, resize); each stage stamps
+    // once (first crossing) so the retry loop can pass it through unconditionally.
+    trace: Option<&crate::bringup::Trace>,
 ) -> Result<Pipeline> {
     // ~10s first-frame wait per attempt. 8 gives a ~90s budget for the SLOW case: a host-managed
     // gamescope session cold-starting Steam Big Picture (the SteamOS/Bazzite takeover) can take
@@ -1904,7 +2184,7 @@ fn build_pipeline_with_retry(
                 attempt - 1
             );
         }
-        match build_pipeline(vd, mode, bitrate_kbps, bit_depth, plan, quit) {
+        match build_pipeline(vd, mode, bitrate_kbps, bit_depth, plan, quit, trace) {
             Ok(pipe) => {
                 if attempt > 1 {
                     tracing::info!(attempt, "pipeline up after retry");
@@ -1979,6 +2259,7 @@ fn reset_stalled_encoder(
     true
 }
 
+#[allow(clippy::too_many_arguments)]
 fn build_pipeline(
     vd: &mut Box<dyn crate::vdisplay::VirtualDisplay>,
     mode: punktfunk_core::Mode,
@@ -1986,6 +2267,9 @@ fn build_pipeline(
     bit_depth: u8,
     plan: crate::session_plan::SessionPlan,
     quit: &Arc<AtomicBool>,
+    // Transition trace (P0.1): stamps the build's stages (display acquire, capture attach, first
+    // frame, encoder open) into the bring-up/resize timeline. `None` on untraced rebuilds.
+    trace: Option<&crate::bringup::Trace>,
 ) -> Result<Pipeline> {
     // Acquire through the registry (design/display-management.md): on Linux this pools the display
     // for keep-alive (reuse a kept one, or create + keep the backend's keepalive so it outlives the
@@ -1994,6 +2278,9 @@ fn build_pipeline(
     // `quit` flag rides into the lease so a deliberate-quit teardown skips the keep-alive linger.
     let vout = crate::vdisplay::registry::acquire(vd, mode, quit.clone())
         .context("create virtual output")?;
+    if let Some(t) = trace {
+        t.mark("display_acquired");
+    }
     // A2: if this was a REUSED kept display and its first frame fails, tear the (dead) pool entry down
     // so the retry loop's next acquire creates fresh instead of re-wedging on the same corpse. Read the
     // gen BEFORE `capture_virtual_output` consumes `vout`. (Linux-only — the pool is Linux.)
@@ -2031,6 +2318,9 @@ fn build_pipeline(
     let mut capturer =
         crate::capture::capture_virtual_output(vout, plan.output_format(), plan.capture)
             .context("capture virtual output")?;
+    if let Some(t) = trace {
+        t.mark("capture_attached");
+    }
     capturer.set_active(true);
     let frame = match capturer.next_frame().context("first frame") {
         Ok(f) => f,
@@ -2043,6 +2333,9 @@ fn build_pipeline(
             return Err(e);
         }
     };
+    if let Some(t) = trace {
+        t.mark("first_frame");
+    }
     // `bit_depth` is the handshake-negotiated value (8, or 10 = HEVC Main10 when the client
     // advertised VIDEO_CAP_10BIT and the host opted in). Threaded down from the Welcome.
     let mut enc = crate::encode::open_video(
@@ -2057,6 +2350,9 @@ fn build_pipeline(
         plan.chroma,
     )
     .context("open video encoder")?;
+    if let Some(t) = trace {
+        t.mark("encoder_open");
+    }
     if let Some(c) = plan.wire_chunk {
         enc.set_wire_chunking(c);
     }

crates/punktfunk-host/src/session_status.rs

@@ -40,6 +40,11 @@ struct LiveSession {
     client: String,
     /// Whether the session negotiated HDR — carried on the lifecycle events.
     hdr: bool,
+    /// Completed bring-up total (hello → first packet), ms; 0 until the first packet left. Written
+    /// once by the session's [`crate::bringup::Trace`] (latency plan P0.1).
+    ttff_ms: Arc<AtomicU32>,
+    /// Most recent completed mid-stream resize (reconfigure → pipeline rebuilt), ms; 0 = none yet.
+    last_resize_ms: Arc<AtomicU32>,
 }
 
 /// A resolved read of one live session, for the `/status` view.
@@ -50,6 +55,10 @@ pub struct SessionSnapshot {
     pub fps: u32,
     pub bitrate_kbps: u32,
     pub codec: Codec,
+    /// Bring-up total (hello → first packet), ms; 0 while still bringing up (latency plan P0.1).
+    pub time_to_first_frame_ms: u32,
+    /// Most recent mid-stream resize total, ms; 0 = no resize this session.
+    pub last_resize_ms: u32,
 }
 
 fn registry() -> &'static Mutex<Vec<LiveSession>> {
@@ -84,6 +93,8 @@ pub fn register(
     force_idr: Arc<AtomicBool>,
     client: String,
     hdr: bool,
+    ttff_ms: Arc<AtomicU32>,
+    last_resize_ms: Arc<AtomicU32>,
 ) -> LiveSessionGuard {
     let id = next_id();
     let session = LiveSession {
@@ -95,6 +106,8 @@ pub fn register(
         force_idr,
         client,
         hdr,
+        ttff_ms,
+        last_resize_ms,
     };
     crate::events::emit(crate::events::EventKind::SessionStarted {
         session: session_ref(&session),
@@ -140,6 +153,8 @@ pub fn snapshot() -> Vec<SessionSnapshot> {
                 fps,
                 bitrate_kbps: s.bitrate_kbps.load(Ordering::Relaxed),
                 codec: s.codec,
+                time_to_first_frame_ms: s.ttff_ms.load(Ordering::Relaxed),
+                last_resize_ms: s.last_resize_ms.load(Ordering::Relaxed),
             }
         })
         .collect()

packaging/windows/build-gamepad-drivers.ps1

@@ -14,7 +14,7 @@
   --gamepad` consumes (per-driver .inf/.cat/.dll + one shared punktfunk-driver.cer).
 
   Output (-Out): pf_dualsense.{dll,inf,cat} + pf_xusb.{dll,inf,cat} + pf_mouse.{dll,inf,cat} +
-  punktfunk-driver.cer. (pf_mouse is the resident virtual HID pointer, not a gamepad — it shares
+  punktfunk-driver.cer. (pf_mouse is the resident virtual HID pointer, not a gamepad - it shares
   this pipeline + the --gamepad install path.)
 
 .EXAMPLE

packaging/windows/drivers/pf-vdisplay/src/callbacks.rs

@@ -129,6 +129,16 @@ pub unsafe extern "C" fn parse_monitor_description2(
         return STATUS_NOT_FOUND;
     };
     let count = crate::monitor::flatten(&modes).count() as u32;
+    // Bring-up/diagnostic visibility (P2): does the OS ever RE-parse the description after an
+    // UPDATE_MODES? The head mode names which list generation this call served.
+    if let Some(head) = crate::monitor::flatten(&modes).next() {
+        dbglog!(
+            "[pf-vd] parse_monitor_description2(id={id}): {count} modes, head {}x{}@{}",
+            head.width,
+            head.height,
+            head.refresh_rate
+        );
+    }
     out_args.MonitorModeBufferOutputCount = count;
     if in_args.MonitorModeBufferInputCount < count {
         // A zero input count is a count-only probe (success); a non-zero too-small buffer is an error.
@@ -204,6 +214,17 @@ pub unsafe extern "C" fn monitor_query_modes2(
         return STATUS_NOT_FOUND;
     };
     let count = crate::monitor::flatten(&modes).count() as u32;
+    // Diagnostic visibility (P2): shows whether/when the OS re-queries target modes after an
+    // UPDATE_MODES (the head mode names the list generation this call served).
+    if let Some(head) = crate::monitor::flatten(&modes).next() {
+        dbglog!(
+            "[pf-vd] monitor_query_modes2: {count} modes, head {}x{}@{} (fill={})",
+            head.width,
+            head.height,
+            head.refresh_rate,
+            in_args.TargetModeBufferInputCount >= count
+        );
+    }
     out_args.TargetModeBufferOutputCount = count;
     if in_args.TargetModeBufferInputCount >= count {
         // SAFETY: `pTargetModes` points to >= `count` IDDCX_TARGET_MODE2 entries.

packaging/windows/drivers/pf-vdisplay/src/control.rs

@@ -95,6 +95,8 @@ pub unsafe fn dispatch(request: WDFREQUEST, ioctl_code: u32) {
         control::IOCTL_SET_RENDER_ADAPTER => unsafe { set_render_adapter(request) },
         // SAFETY: `request` is the framework WDFREQUEST.
         control::IOCTL_SET_FRAME_CHANNEL => unsafe { set_frame_channel(request) },
+        // SAFETY: `request` is the framework WDFREQUEST.
+        control::IOCTL_UPDATE_MODES => unsafe { update_modes(request) },
         _ => complete(request, STATUS_NOT_FOUND),
     }
 }
@@ -198,6 +200,28 @@ unsafe fn set_frame_channel(request: WDFREQUEST) {
     }
 }
 
+/// `IOCTL_UPDATE_MODES` (v4): refresh a LIVE monitor's target-mode list to a new preferred mode —
+/// the in-place mid-stream resize (`design/first-frame-and-resize-latency.md` P2). The monitor is
+/// NOT departed: its OS identity, swap-chain machinery and retained frame stash all survive; the
+/// host force-sets the freshly-advertised mode afterwards.
+///
+/// # Safety
+/// `request` is the framework `WDFREQUEST`.
+unsafe fn update_modes(request: WDFREQUEST) {
+    // SAFETY: `request` is the framework WDFREQUEST.
+    let Some(req) = (unsafe { read_input::<control::UpdateModesRequest>(request) }) else {
+        complete(request, STATUS_INVALID_PARAMETER);
+        return;
+    };
+    if !valid_mode(req.width, req.height, req.refresh_hz) {
+        complete(request, STATUS_INVALID_PARAMETER);
+        return;
+    }
+    let st =
+        crate::monitor::update_monitor_modes(req.session_id, req.width, req.height, req.refresh_hz);
+    complete(request, st);
+}
+
 /// `IOCTL_REMOVE`: depart + drop the monitor for the given session id.
 ///
 /// # Safety

packaging/windows/drivers/pf-vdisplay/src/monitor.rs

@@ -7,7 +7,7 @@
 use std::sync::Mutex;
 use std::time::{Duration, Instant};
 
-use wdk_sys::{WDFOBJECT, call_unsafe_wdf_function_binding, iddcx};
+use wdk_sys::{NTSTATUS, WDFOBJECT, call_unsafe_wdf_function_binding, iddcx};
 
 /// One resolution with the refresh rates it supports.
 #[derive(Clone)]
@@ -146,6 +146,41 @@ pub fn reap_orphaned(grace: Duration) -> usize {
     n
 }
 
+/// Append `from`'s modes to `into`, skipping resolutions already present, capped at
+/// [`MODE_LIST_CAP`] — the accumulate half of the union semantics (see [`update_monitor_modes`]).
+fn union_modes(into: &mut Vec<Mode>, from: &[Mode]) {
+    for m in from {
+        if into.len() >= MODE_LIST_CAP {
+            break;
+        }
+        if !into
+            .iter()
+            .any(|e| (e.width, e.height) == (m.width, m.height))
+        {
+            into.push(m.clone());
+        }
+    }
+}
+
+/// The last advertised mode list of a DEPARTED monitor, per monitor id — consumed by the next
+/// same-id [`create_monitor`] so a re-arrived monitor's ARRIVAL list already contains every mode
+/// its predecessor ever served. The OS pins a monitor's settable set at arrival (see
+/// [`update_monitor_modes`]), so this is what makes a windowed↔fullscreen cycle (or any return to
+/// a previously-used size) an IN-PLACE mode set instead of another hotplug. In-process only (a
+/// WUDFHost restart forgets it — harmless, the next resizes re-teach it); bounded: ≤ 16 ids ×
+/// [`MODE_LIST_CAP`] modes.
+static MODE_HISTORY: Mutex<Vec<(u32, Vec<Mode>)>> = Mutex::new(Vec::new());
+
+/// Record a departing monitor's advertised list for its id ([`MODE_HISTORY`]).
+fn remember_modes(id: u32, modes: &[Mode]) {
+    let mut hist = MODE_HISTORY.lock().unwrap_or_else(|e| e.into_inner());
+    if let Some(slot) = hist.iter_mut().find(|(i, _)| *i == id) {
+        slot.1 = modes.to_vec();
+    } else {
+        hist.push((id, modes.to_vec()));
+    }
+}
+
 /// Fallback modes appended after the requested mode, so a topology change still has options.
 fn default_modes() -> Vec<Mode> {
     vec![
@@ -365,9 +400,7 @@ pub fn preserve_publisher(
 /// swap-chain's render adapter matches the publisher's ([`FramePublisher::render_adapter`]) — same
 /// pooled device, so its context + opened ring textures are still valid; on a mismatch the caller drops
 /// it and waits for a fresh channel delivery instead. `None` until a worker has stashed one.
-pub fn take_preserved_publisher(
+pub fn take_preserved_publisher(target_id: u32) -> Option<crate::frame_transport::FramePublisher> {
-    target_id: u32,
-) -> Option<crate::frame_transport::FramePublisher> {
     if target_id == 0 {
         return None;
     }
@@ -496,6 +529,15 @@ pub fn create_monitor(
     let id = {
         let mut lock = lock_monitors();
         let id = resolve_id(&lock, preferred_id);
+        // Same-id mode history (P2 union semantics): a RE-ARRIVED monitor advertises every mode
+        // its departed predecessor served, so the OS's arrival-pinned settable set already
+        // contains them — a return to any previously-used size is then an IN-PLACE mode set.
+        {
+            let hist = MODE_HISTORY.lock().unwrap_or_else(|e| e.into_inner());
+            if let Some((_, prev)) = hist.iter().find(|(i, _)| *i == id) {
+                union_modes(&mut modes, prev);
+            }
+        }
         lock.push(MonitorObject {
             object: None,
             id,
@@ -600,6 +642,77 @@ pub fn create_monitor(
     Some((id, target_id, luid_low, luid_high))
 }
 
+/// How many distinct resolutions a monitor's advertised list may accumulate (the requested head +
+/// history + the built-in fallbacks). Bounds the union growth across many resizes; the OLDEST
+/// history entries fall off first.
+const MODE_LIST_CAP: usize = 12;
+
+/// `IOCTL_UPDATE_MODES` (v4): refresh the LIVE monitor's advertised mode list to lead with a new
+/// preferred mode and push the new TARGET mode list to the OS via `IddCxMonitorUpdateModes2` —
+/// the in-place mid-stream resize (`design/first-frame-and-resize-latency.md` P2). No departure:
+/// the monitor's OS identity, its swap-chain worker and the retained frame stash all survive.
+/// The `*2` (HDR) DDI matches the `*2` mode/buffer family this driver already requires
+/// (IddCx 1.10), so it adds no new OS floor.
+///
+/// UNION semantics (on-glass finding, build 26200): the OS re-parses the description AND
+/// re-queries target modes after `UpdateModes2` — our callbacks served the fresh list — yet the
+/// SETTABLE set stays pruned to the modes known at monitor ARRIVAL (the monitor source-mode set
+/// is pinned then). So replacing the list can only ever LOSE settable modes (v1 of this op
+/// dropped the arrival mode from the target list, breaking even a resize BACK to it); the update
+/// therefore accumulates — new mode first, every previously-advertised mode kept (deduped by
+/// resolution, capped at [`MODE_LIST_CAP`]) — and the real payoff is at the NEXT re-arrival,
+/// where [`create_monitor`]'s same-id history union makes every previously-used mode settable.
+///
+/// The stored list is updated FIRST (under the lock) so any OS re-query through the mode DDIs
+/// ([`modes_for_object`]/[`modes_for_id`]) sees the new list, and REVERTED if the DDI fails — the
+/// OS then still holds the old list and the two stay coherent. The DDI itself is called OUTSIDE
+/// the lock (it may re-enter the mode-query callbacks, which lock [`MONITOR_MODES`]).
+pub fn update_monitor_modes(session_id: u64, width: u32, height: u32, refresh: u32) -> NTSTATUS {
+    // Swap the stored list (union — see above) + grab the live handle under the lock.
+    let (object, old_modes, new_modes) = {
+        let mut lock = lock_monitors();
+        let Some(m) = lock.iter_mut().find(|m| m.session_id == session_id) else {
+            return crate::STATUS_NOT_FOUND;
+        };
+        let Some(object) = m.object else {
+            return crate::STATUS_NOT_FOUND; // created but not yet arrived — nothing to update
+        };
+        let mut new_modes = vec![Mode {
+            width,
+            height,
+            refresh_rates: vec![refresh],
+        }];
+        union_modes(&mut new_modes, &m.modes);
+        let old = core::mem::replace(&mut m.modes, new_modes.clone());
+        (object, old, new_modes)
+    };
+
+    // The OS's target-mode list for this monitor (the `*2`/HDR shape, like `monitor_query_modes2`).
+    let mut targets: Vec<iddcx::IDDCX_TARGET_MODE2> = flatten(&new_modes)
+        .map(|item| target_mode2(item.width, item.height, item.refresh_rate))
+        .collect();
+    let mut in_args = pod_init!(iddcx::IDARG_IN_UPDATEMODES2);
+    in_args.Reason = iddcx::IDDCX_UPDATE_REASON::IDDCX_UPDATE_REASON_OTHER;
+    in_args.TargetModeCount = targets.len() as u32;
+    in_args.pTargetModes = targets.as_mut_ptr();
+    // SAFETY: `object` is a live IddCx monitor handle (arrived — checked above; a concurrent REMOVE
+    // is serialized by the host, which only ever resizes a monitor its own session holds a lease
+    // on). `in_args` points at valid local storage (`targets` outlives the synchronous DDI call).
+    let st = unsafe { wdk_iddcx::IddCxMonitorUpdateModes2(object, &in_args) };
+    dbglog!(
+        "[pf-vd] IddCxMonitorUpdateModes2(session={session_id}, {width}x{height}@{refresh}) -> {st:#x}"
+    );
+    if !wdk_iddcx::nt_success(st) {
+        // Keep the stored list coherent with what the OS actually holds (the old one).
+        let mut lock = lock_monitors();
+        if let Some(m) = lock.iter_mut().find(|m| m.session_id == session_id) {
+            m.modes = old_modes;
+        }
+        return st;
+    }
+    crate::STATUS_SUCCESS
+}
+
 /// `IOCTL_REMOVE`: depart + drop the monitor for `session_id`. Returns true if one was removed.
 pub fn remove_monitor(session_id: u64) -> bool {
     // Pull out the IddCx handle AND the swap-chain processor under the lock, but drop the processor
@@ -611,6 +724,9 @@ pub fn remove_monitor(session_id: u64) -> bool {
             return false;
         };
         let mut entry = lock.remove(pos);
+        // Keep the departing monitor's advertised list for its id — the next same-id create
+        // unions it back in (P2 mode history; see MODE_HISTORY).
+        remember_modes(entry.id, &entry.modes);
         (entry.object, entry.swap_chain_processor.take())
     };
     // Drop the worker FIRST (it joins + deletes the swap-chain), THEN depart the monitor.

packaging/windows/drivers/wdk-iddcx/src/lib.rs

@@ -140,6 +140,16 @@ iddcx_ddi!(
         in_args: *const iddcx::IDARG_IN_ADAPTERSETRENDERADAPTER,
     ) @ IddCxAdapterSetRenderAdapterTableIndex as PFN_IDDCXADAPTERSETRENDERADAPTER -> ()
 );
+iddcx_ddi!(
+    /// Refresh a LIVE monitor's target-mode list (the HDR `*2` variant, IddCx 1.10 — the same API
+    /// family as the `*2` mode/buffer DDIs this driver already requires): the OS re-evaluates which
+    /// modes the target supports WITHOUT a monitor departure, so the host can then mode-set to a
+    /// freshly-advertised mode in place (the mid-stream resize, latency plan P2).
+    IddCxMonitorUpdateModes2(
+        monitor: iddcx::IDDCX_MONITOR,
+        in_args: *const iddcx::IDARG_IN_UPDATEMODES2,
+    ) @ IddCxMonitorUpdateModes2TableIndex as PFN_IDDCXMONITORUPDATEMODES2
+);
 iddcx_ddi!(
     /// Bind a D3D device to an assigned swap-chain. HRESULT-shaped (0x887A0026 → retry on monitor flap).
     IddCxSwapChainSetDevice(