Add a runtime-free Wake-on-LAN sender in punktfunk-core (per-interface subnet-directed broadcast + 255.255.255.255 on ports 9/7, repeated, optional last-known-IP unicast) exposed both as a Rust fn and a punktfunk_wake_on_lan C-ABI (ABI v3), plus a parse_mac helper. The host enumerates its wake-capable NIC MAC(s) and advertises them in a new mDNS `mac` TXT record (routed NIC first), and best-effort detects & warns (never modifies) when the NIC isn't armed for WoL.
MAC delivery is via the unauthenticated mDNS TXT rather than the connection handshake by design: a spoofed MAC only makes a wake fail (the packet is inert; the cert fingerprint still gates the connection), and it avoids threading through the hot connect path.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The Steam-conflict gate scanned /sys/bus/hid/devices for non-virtual 28DE
devices, but the usbip/gadget virtual Decks present a REAL USB device (vhci
resolves through vhci_hcd, not /devices/virtual/) — so a just-ended session's
pad still detaching, or a concurrent session's live one, read as "physical
Steam controller attached" and degraded every back-to-back Deck session to
DualSense (observed live on Bazzite). Exclude our pads by their PFDK… serial
(HID_UNIQ), with the vhci_hcd path as belt and braces.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Video (Deck): the VAAPI zero-copy path renders corrupt/gray/washed-out on the
Deck — root-caused to Mesa >= 25.1 exporting radeonsi VCN decode surfaces TILED
(the Flatpak runtime's Mesa 26 drives both the decoder and GTK's GL, and GTK's
tiled-NV12 dmabuf import mishandles it; desktop Tier-1 validations ran distro
Mesa with linear export). `auto` now resolves to software on a Deck (clean,
correct-colour, easily handles 1280x800 HEVC); PUNKTFUNK_DECODER=vaapi still
forces the hw path, with the descriptor modifier dump + GSK_RENDERER as the
bisect levers. Also reserve extra_hw_frames=4 on the VAAPI decoder: the
presenter pins mapped surfaces past receive_frame, and the fixed pool recycling
a surface the renderer still samples is intermittent block corruption anywhere.
Input (Deck): with Steam Input ON for Punktfunk, SDL sees only Steam's virtual
X360 pad — the right trackpad arrives as a plain right stick and the left
trackpad/paddles/gyro not at all, silently. The client now checks once the
post-attach enumeration settles and raises a toast + warn naming the fix
(disable Steam Input for the shortcut). The host logs a one-shot warning when
InputPlumber is running (Bazzite default) since it can grab the virtual Deck
pad and re-emit it under a different identity.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Mac → Windows mic passthrough crackled heavily while the identical
stream was clean on the Linux host. Cause: clients push mic audio in
BURSTS on their own clock (the Mac input tap yields ~two 20 ms Opus
packets every ~42 ms) while the WASAPI render loop pulled a block every
~10 ms device period and greedily drained whatever was queued, padding
the rest with zeros — the queue sat near-empty and most periods
inserted mid-stream silence. The Linux backend has absorbed this since
day one with its priming jitter buffer; the WASAPI loop had none.
Port the same semantics: emit silence until ~48 ms is buffered (covers
the worst inter-burst gap), then play from the cushion (zero-filling
only a momentary shortfall), re-prime only after a genuine full drain
(client went quiet). Queue cap raised 80 → 120 ms for burst headroom;
steady-state added latency ≈ the 48 ms cushion.
Diagnosed live on .173: probe tone recording from CABLE Output proved
the endpoint wiring, then the burst-vs-period math explained the
crackle. Build-verified on Windows; on-glass listen pending.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The unified-stats equation's host+network stage was one opaque number
because the wire carried nothing but pts_ns. Now the host reports its own
share per frame: when the client's Hello sets VIDEO_CAP_HOST_TIMING (0x08),
the send thread emits a 13-byte 0xCF datagram — [tag][pts_ns u64][host_us
u32] — right after the AU's last packet leaves the socket, so host_us =
capture→fully-sent (capture read/convert, encode, FEC+seal, paced send)
against the same anchor the wire pts carries. Clients correlate by pts_ns
and derive network = (received + clock_offset − pts) − host_us; the two
terms tile per frame by construction.
Back-compat is free in all four combinations: old clients ignore unknown
datagram tags, old hosts ignore unknown cap bits (client keeps the combined
stage). The hardened data-plane format is untouched — this rides the
established QUIC side-plane pattern (0xC8…0xCE). NativeClient ORs the bit
in unconditionally and exposes next_host_timing(); the C ABI gains
PunktfunkHostTiming + punktfunk_connection_next_host_timing (additive).
The synthetic host emits 0xCF too, so pure-loopback protocol tests cover
the plane.
The probe reports the split (host_p50/p95_us · net_p50/p95_us) and is our
direct analogue of Sunshine's "host processing latency" — ours additionally
includes the paced send.
Validated on loopback (synthetic host + probe, debug build): 240/240 AUs
matched, host_p50 6.5 ms + net_p50 6.4 ms ≈ capture→received p50 13.0 ms.
Core suite + new 0xCF roundtrip/truncation test green; host+core+probe
clippy clean.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
One stat model everywhere (design/stats-unification.md): four measurement
points (capture/received/decoded/displayed), three stages that tile the
interval exactly, and a HUD that shows the addition explicitly —
end-to-end 14.2 ms p50 · 19.8 p95 · capture→on-glass
= host+network 9.8 + decode 2.1 + display 2.3
replacing each client's ad-hoc mix of overlapping absolutes (the Apple HUD's
three arrow lines that looked sequential but weren't), mean-vs-median decode
times (Windows/Linux), missing same-host-clock flags (Windows/Linux), and
three different names for the same capture→received measurement (probe's
"reassembled", Apple/Android's "client", Windows/Linux's post-decode "lat").
Per client: Apple threads receivedNs through the VT decode via the frame
refcon bit pattern so the decode stage exists at all (stage-1 fallback
honestly degrades to a capture→received headline); Windows carries
FrameTimes through the existing frame channel to the render thread and adds
e2e p50/p95 post-Present; Linux stamps received at AU pop and rides
decoded_ns on DecodedFrame to the paintable-set site; Android pairs receipt
stamps with MediaCodec output buffers via the codec's pts round-trip (JNI
stats array 14→16 doubles, indexes 0-13 unchanged). fps now uniformly counts
received AUs; lost/(received+lost) per window, hidden at zero.
docs-site gains "Understanding the Stats Overlay": what each line means, why
the equation only approximately sums (percentiles), and a line-by-line
Moonlight/Sunshine matrix — including that Moonlight has no end-to-end
number and its "network latency" is an ENet control RTT, so punktfunk's
headline must not be compared against any single Moonlight line.
Verified here: linux client + probe + core check/clippy/fmt green, android
native cargo-ndk arm64 check green. Pending: Windows CI + on-glass, swift
test on the mac, on-device Android.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Found by a live boot-order test (host started before the user session's
PipeWire): PwMicSource::open returned Ok before the daemon connection was
attempted, so a PipeWire that wasn't running surfaced as an instantly-dead
instance instead of an open failure — and the pump churned
open→die→reopen at heartbeat rate (1 Hz "virtual mic ready" log spam)
instead of backing off.
- PwMicSource::open now has a bring-up handshake (mirrors the Windows
backend): ready only after connect + stream connect succeed, so a
down daemon is an open ERROR and the pump's backoff engages.
- The pump triages deaths: an instance that lived >= 5 s (a one-off
daemon restart) reopens immediately with the backoff reset; one that
died right after opening counts as a failed open and backs off
(2 s → 60 s cap). New pump test rapid_death_backs_off.
Re-validated live: host started with PipeWire stopped → throttled
"unavailable" warns, zero churn; daemon started → mic node up on the
next retry; exactly one pump + one loop thread (no leak).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Mic passthrough silently died on real hosts. Root causes, all fixed:
- No liveness anywhere: a PipeWire restart (Linux) or any WASAPI device
error (Windows) killed the backend worker; push() fed the dead queue
for the rest of the host's life. VirtualMic now has a liveness
contract (push -> bool, alive(), discard()) and the new shared
audio::MicPump reopens with backoff, probing on an idle heartbeat so
the mic heals BETWEEN sessions too. Validated live: systemctl restart
pipewire -> node back in ~0.5 s, tone flows through the reopened
backend.
- Lazy creation: the mic device didn't exist until the first 0xCB
frame, but games bind their capture device at launch and never
re-follow. The pump opens eagerly at host start (node exists with
zero clients, elected default source).
- Windows headless dead-end: with VB-CABLE as the ONLY render endpoint
(exactly what the installer ships), the anti-echo guard rejected the
cable as the default render endpoint -> mic permanently dead. The new
wiring_plan (pure, unit-tested on every platform) assigns the mic its
endpoint FIRST (cable reserved for the mic), points the loopback at a
DIFFERENT endpoint, and the capture side now yields (explicit
endpoint or honest error) instead of the mic dying. Plan recomputed
per (re)open — endpoints churn at boot/logon/driver installs.
- Stale bursts: buffered audio from a previous session played into a
newly-attached recorder (observed live). Timestamped chunks + a
consumer-gap check in the process callback age everything past 1 s.
The Linux node mechanism stays the stream-based Audio/Source with
RT_PROCESS + priority.session: the canonical null-audio-sink adapter
recipe was tested on this box (PipeWire 1.6.2) and never gets a clock
(QUANT 0 -> pure silence), and WirePlumber reroutes a feeder targeting
it to the default sink (echo). Decision documented in the module docs.
Live-validated on this box (synthetic host + probe --mic-test,
pw-record): eager node, both attach orderings, PipeWire-restart
self-heal, post-session silence. Windows side compile/CI + on-glass
validation pending.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
DWM presents a display only when something dirties it. On an idle desktop a
perfectly healthy session sat at E_PENDING: the driver attached but no
first frame ever landed, so wait_for_attach's 4 s gate failed the open (and
a mid-session ring recreate hit the same stall against the 3 s
recover-or-drop). A real client escaped only because its own input soon
dirtied the desktop; a headless probe / input-less connect never did.
kick_dwm_compose() injects two net-zero 1 px relative mouse moves via
SendInput — pf-vdisplay has no hardware-cursor plane, so a cursor move is
composited into the frame, a guaranteed real present onto the IDD
swap-chain (the mechanism --input-test always relied on; the pointer ends
where it started). Wired into wait_for_attach (first kick at 600 ms, then
every 800 ms) and, rate-limited, into the GB1 recovery window.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
On-glass the lazy (first-session) claim let a second host started while the
freshly-restarted service sat idle win the mutex and ADD a monitor on the
real driver — priority backwards. The claim is now a process-global,
retryable slot (a failed claim is not memoized, so it heals once the other
instance exits), and `serve` claims it before any client can connect;
ensure_device keeps the lazy claim for standalone punktfunk1-host runs.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
On-glass the SCM service creates Global\punktfunk-vdisplay-manager as
SYSTEM, so a second elevated-admin host's CreateMutexW fails ACCESS_DENIED
(the implicit open is checked against the SYSTEM DACL) before the
ALREADY_EXISTS branch can fire — right refusal, wrong message. Map it to
the same loud "another instance is live" error.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Batch C of the audit's medium tier (M7+M8+M9):
- M7: GameStream sessions now run the same begin_idd_setup dance as
punktfunk/1 before creating the shared monitor. A GS connect could
previously ADD/reconfigure the monitor while a native session was
mid-build (and vice versa), and its sealed-channel delivery replaced the
native ring (newest-wins) — each plane could freeze the other. GS has no
cooperative stop plumbing, so it registers a flag nobody reads: a later
session signals it, waits the 3 s grace, then force-preempts — the
intended handover.
- M8: the linger-expiry teardown now runs UNDER the state lock. Running it
outside let a concurrent acquire see Idle and ADD+isolate while the old
monitor's pinger-join / CCD-restore / REMOVE were still in flight — a
failed or de-isolated session exactly at the expiry boundary. A racing
acquire now waits the few teardown seconds instead. Lock order stays
state → device everywhere; the pinger takes only the device lock.
- M9: a named mutex (Global\punktfunk-vdisplay-manager) makes a SECOND host
process fail its vdisplay open loudly instead of firing a startup
CLEAR_ALL that razes the live host's monitors mid-stream (the admin
footgun the shared watchdog then masked).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Three levers to lower and steady decode latency on Snapdragon (Adreno) devices:
- ADPF (Adaptive Performance Framework): a new dlsym-resolved hint session
(native/src/adpf.rs; API-33+, resolved at runtime so there's no build-time
link dependency and libpunktfunk_android.so still loads on API 31/32) tells
the CPU governor the video pipeline runs a per-frame real-time workload, so it
keeps those threads on fast cores at high clocks. It now covers all three
latency-critical threads — the pf-decode feed/drain/present loop, the core
data-plane pump (UDP receive + FEC reassembly), and the audio thread — via a
new generic hot-thread registry on NativeClient (register_hot_thread /
hot_thread_ids; the pump self-registers). The session is built lazily on the
first presented frame, since ADPF createSession rejects a set containing any
not-yet-live tid.
- operating-rate -> Short.MAX ("as fast as possible"): pushes the Qualcomm
decoder to run each frame at max clocks instead of merely sustaining the
display rate at a power-saving clock that adds per-frame decode latency.
- appCategory="game": makes the app eligible for OEM Game Mode / Game Dashboard
performance profiles.
The core registry is cross-platform (gettid on Linux/Android, a no-op
elsewhere) — no Android-specific pollution of the shared core. Host workspace +
64 core tests green; Android arm64-v8a + x86_64 (platform 31) build + clippy
clean. On-device Snapdragon validation pending.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Fault-injection on-glass showed a killed/crashed WUDFHost leaves the devnode
"started" but HOSTLESS: PnP Status OK, no WUDFHost process, zero device-
interface instances — is_available() then fails every future session at the
vdisplay::open gate (and a reopen inside VdisplayDriver::open finds nothing),
until something cycles the device. Port reset-pf-vdisplay.ps1's adapter
disable→enable step in-process (restart_vdisplay_device): the open gate now
uses ensure_available() (cycle once + bounded re-probe; a genuinely
uninstalled driver — no adapter devnode — still fails fast), and
VdisplayDriver::open retries open_device over a short arrival window after a
cycle, covering the manager's reopen path too.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Batch A of the audit's medium tier (M1+M2+M3):
- M1 driver-death detection: a dead WUDFHost stops publishing, which at the
ring is indistinguishable from an idle desktop — SDR sessions streamed a
frozen frame forever (next_frame's 20 s bail is unreachable once anything
presented). The ChannelBroker's process handle now doubles as a liveness
probe (SYNCHRONIZE at OpenProcess); while no fresh frame arrives,
try_consume polls it (rate-limited) and fails the capturer, landing in the
session's bounded in-place rebuild.
- M2 reopenable control device: the manager's OnceLock-cached handle is now
a retire/reopen DeviceSlot — a gone-classified IOCTL failure (driver
upgrade / WUDFHost restart; pinger, create, or REMOVE) retires the handle
and the next use reopens + re-handshakes. Retired handles are deliberately
kept alive forever: bare-HANDLE holders (pinger, ChannelBroker) rely on
never-closed, and a retired handle only fails IOCTLs. CLEAR_ALL runs on
the FIRST open only (a reopen races live-ish sessions); acquire retries
the monitor create once after a reopen. The JOIN path now probes the
active monitor's WUDFHost pid and preempts a DEAD monitor instead of
handing the rebuilding session its stale target — without this the whole
recovery chain starved to the rebuild budget.
- M3 interface discovery: enumerate ALL interface instances with an
SPINT_ACTIVE filter (a Code-10 devnode at index 0 no longer shadows the
live interface), HDEVINFO behind RAII (error paths leaked one per probe),
the raw device handle wrapped before GET_INFO (leaked on handshake
failure), and the detail-sizing result guarded before the cbSize write.
- pf-driver-proto: SetFrameChannelRequest doc now states the real
adopt-on-success contract (the old wording invited a driver-side
close-on-error — a cross-process double-close against the host's reap).
- install: pf_vdisplay_present() passes /connected so a phantom devnode
can't suppress creating a live ROOT node.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Five verified findings from the IDD-push/pf-vdisplay deep audit:
- Keyed-mutex acquire (BOTH endpoints): AcquireSync returns WAIT_TIMEOUT
(0x102) / WAIT_ABANDONED (0x80) as SUCCESS-severity HRESULTs, which the
windows-rs Result wrapper erases — a busy slot read as "acquired", so
driver and host could race the same ring texture (torn frames) and the
designed busy-skip backpressure was dead code. Both sides now classify
the raw vtable HRESULT; WAIT_ABANDONED counts as acquired (ownership
transfers — refusing it would wedge the slot forever).
- Host SDR hot path leaked one ID3D11VideoProcessorInputView per converted
frame: the D3D11_VIDEO_PROCESSOR_STREAM ManuallyDrop field suppressed the
release after VideoProcessorBlt. Released by hand now, success or not.
- Driver leaked IddCx's per-acquire surface reference (from_raw_borrowed on
a TRANSFERRED reference — the MS sample Attach/Reset's it): the swap-chain
surface set survived swap-chain destruction, the likely true root cause of
the ~50 MB-per-reconnect VRAM loss that device pooling only mitigated.
Now adopted via from_raw (publisher or not) and dropped pre-Finished.
- PUNKTFUNK_IDD_PUSH removed: capture is unconditionally IDD-push, but the
vdisplay manager still gated the lingering-monitor preempt (and render
pin) on the knob, whose default was OFF — dev/CLI runs reused a lingering
monitor whose IddCx swap-chain is dead (black reconnect). The preempt and
the render-GPU pin are now unconditional; host.env comments no longer
promise the removed DDA/WGC fallback.
- Driver D3D device: dropped D3D11_CREATE_DEVICE_SINGLETHREADED (unsound
since DEVICE_POOL shares one device across processors) and the pooled
immediate context is now SetMultithreadProtected — two concurrent
monitors' workers otherwise race an unlocked context (UB in the UMD).
No wire-contract change (pf-driver-proto untouched); the driver fixes take
effect on the next pf-vdisplay redeploy.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
On-box A/B on the .173 Ryzen 7000 iGPU (720p60, real composition via input
injection — an idle virtual desktop composes ~1 fps and gives meaningless
encode timings): the encode-time-first `quality=speed` preset + explicit `bf=0`
cut host-side encode_us from ~36 ms to ~19.5 ms.
The blocking-poll idea from the prior commit was WRONG and is reverted to a
single non-blocking receive (default PUNKTFUNK_FFWIN_POLL_MS=0): libavcodec's
hevc_amf holds ~2 frames before releasing the oldest (needs frame N+2 to flush
N), so a spin between submits provably never yields the owed AU — verified with
a 150 ms cap pegging at exactly 150 ms across every usage preset and pipeline
depth. That ~2-frame buffer is inherent to the libavcodec wrapper, not host
scheduling; the real latency lever is a direct AMF SDK encoder (the AMF
analogue of the direct-NVENC path), tracked as the next AMD work item. The
env knob is retained for a future VCN/driver where a bounded spin can help.
Also measured and rejected: PUNKTFUNK_ZEROCOPY=1 on AMF is ~2x WORSE (68 ms vs
36 ms) — the D3D11 import path adds sync overhead beyond the readback it saves,
so the system-memory default stays. GPU-priority elevation is already
process-wide (dxgi.rs), so it covers the iGPU encode session with no change.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The blocking poll landed but wait_us pegs at exactly the 2-frame-period cap:
AMF holds the AU ~2 frame periods regardless of retrieval. Field knobs to
bisect on-box (usage preset × poll cap) without rebuild cycles.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The session loop's pipeline deferral was designed around direct NVENC, whose
poll() BLOCKS in lock_bitstream; libavcodec's AMF wrapper is truly async
(EAGAIN until the ASIC finishes), so a single non-blocking receive quantized AU
retrieval to the submit cadence: +1–2 frame periods flat (~43 ms p50 at 720p60
on the Ryzen iGPU vs ~3.5 ms of actual encode). FfmpegWinEncoder now tracks
in-flight frames and, while an AU is owed, spin-polls with short sleeps bounded
to ~2 frame periods (an overloaded encoder degrades to next-tick pickup instead
of stalling capture). Also: quality=speed (latency-first, iGPU-class VCN),
explicit bf=0 (h264_amf defaults >0 on RDNA3+), AMF low-latency submission
mode (FFmpeg ≥6.1, ignored on older).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Firewall (the service.rs core landed in 796d381): scope the web-console rule
(TCP 47992) to Domain+Private by default with a `--allow-public-network` opt-in
that deletes-then-re-adds the rule, and add the installer "Allow connections on
Public networks" task (unchecked) forwarding the flag to `service install` and
`web setup`. Default is now trusted-networks-only; Public is explicit.
Vulnerability disclosure: SECURITY.md (report to security@punktfunk.com, scope,
SLAs, safe harbor), a Gitea issue-template contact link, a README security line,
and a Reporting section on the docs Security page.
Docs: the Security page now documents the Private/Domain firewall default (and
how to fix a misclassified-Public network / opt in); removed internal design-doc
and CLAUDE.md links from the user-facing docs.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Two disk-write fixes:
- pf-xusb/pf-dualsense no longer write C:\Users\Public\pf*-driver.log
unconditionally — the file log is now opt-in (debug builds, or the
PFXUSB_DEBUG_LOG / PFDS_DEBUG_LOG system env var), mirroring the audit-§4.4
fix pf-vdisplay already got: a release driver never writes the world-writable
Public file (info-leak/DoS surface), and the per-report OUTPUT/SET_STATE hex
dumps stop being a sustained per-rumble disk-write path during gameplay.
OutputDebugStringA stays unconditional; the host's driver-silence WARN and
the gamepad-driver-health failure-mode table now say the log is opt-in.
- service.log/host.log get one-generation rotation: at each (re)open a file
over 10 MB is renamed to .old, so a crash-restart loop or a RUST_LOG=debug
left in host.env can't grow the append-forever logs without bound. Rotation
runs only before an open (never under a live appender — host.log's handle
lacks FILE_SHARE_DELETE, so a racing rename harmlessly fails).
Windows CI compile/clippy pending (drivers workspace + host are not
Linux-cross-checkable); rides along with the next pad-driver redeploy.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
punktfunk-host.exe embedded no icon or version resources, so Task Manager and
Explorer showed a bare lowercase exe name with a generic icon. build.rs now
embeds the branded .ico + FileDescription "Punktfunk Host" / ProductName
"Punktfunk" via winresource (same pattern as the Windows client and the tray;
Linux packaging builds skip the block). The tray gets a matching "Punktfunk
Tray" description, and the SCM display name moves off lowercase
"punktfunk streaming host" to "Punktfunk Host" (applied idempotently by
`service install` on upgrade).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The "Open web console" entry was gated on {exe dir}\web\web-run.cmd (Windows)
/ the punktfunk-web unit file (Linux) — which misses consoles run from a repo
checkout (the RTX box, caught on-glass) and shows a dead entry while an
installed console is stopped. The poller now probes https://127.0.0.1:<web
port>/ each cycle (any HTTP response = up, transport failure = down) and the
menu follows live on both platforms.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
New crates/punktfunk-tray — a small per-user companion showing the host service
state at a glance (running / stopped / starting / degraded / failed + the live
session in the tooltip) with one-click actions: open web console, approve a
pending pairing request, start/stop/restart, open logs. No more digging through
logs to learn whether the service came back after a reboot or an update.
Status is service-manager-FIRST (SCM / systemd user unit — a port squatter can
never fake Running), then the new loopback-only unauthenticated
GET /api/v1/local/summary (counts/booleans only; the mgmt token and cert.pem
are SYSTEM/Admins-DACL'd on Windows, so a non-elevated tray cannot bearer-auth).
Windows: windows_subsystem binary (a console exe in the Run key would flash a
terminal at sign-in), Shell_NotifyIcon + hidden window, per-session single
instance, TaskbarCreated re-add, --quit for the uninstaller; service actions
elevate per click via ShellExecuteW "runas" onto the new
`punktfunk-host service restart` (stop → wait Stopped → start).
Linux: ksni/StatusNotifierItem over zbus, systemctl --user actions (no polkit),
/etc/xdg/autostart entry whose --autostart self-gates to actual host users.
Icons: scripts/gen-tray-icons.py (pure stdlib) renders the brand lens + status
dot into committed .ico/hicolor assets; deb/rpm/arch ship binary+autostart+icons.
Live-validated: Linux on the headless KDE session (SNI registration, state
transitions, menu-driven start, dbusmenu layout); Windows on the RTX box
(session-1 launch with no NIM_ADD failure, single instance, --quit, restart
round-trip, summary loopback-200/LAN-401).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The nvenc build linked nvEncodeAPI64.dll's entry points at load time, so a
--features nvenc binary hard-crashed on any box without the NVIDIA driver
(AMD/Intel). Entry points now come from a runtime LoadLibrary table
(encode/windows/nvenc.rs load_api); a missing DLL just falls through the
encoder auto-detect to AMF/QSV/software. The generated import lib and all its
plumbing (gen-nvenc-importlib.ps1, nvenc.def, PUNKTFUNK_NVENC_LIB_DIR,
setup-build-env wiring) are gone.
Live-validated on the RTX 4090 box (NVENC session, 7000+ frames).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Frame ring (pf-vdisplay) and both gamepad SHM channels move off named Global\
objects (openable by any sibling LocalService) to UNNAMED sections/events whose
handles the host DuplicateHandles into the driver's verified WUDFHost with least
access — frame delivery over the SYSTEM+admins-only IOCTL_SET_FRAME_CHANNEL,
pads over a 32-byte named bootstrap mailbox (pid + handle value only, DoS-bounded;
HID minidrivers have no control device). Driver-validated pad_index kills
cross-pad redirects; v1↔v2 mixes fail closed with diagnosis logs on both sides.
Sibling-LocalService denial proven empirically (design/idd-push-security.md,
design/gamepad-channel-sealing.md).
Driver-side raw ops now live behind pf-umdf-util (checked shm accessors, the
forbid(unsafe_code) ChannelClient state machine, WDF request tokens) — the pad
drivers' logic is 100% safe Rust; whole drivers workspace clippy-gated in CI.
driver install --gamepad now sweeps SWD\punktfunk phantom devnodes: a re-created
SwDevice REVIVES the old devnode with its previously-bound driver (never
re-ranks), so an upgrade otherwise leaves the old driver serving — or, across
the v1→v2 fence, a dead pad (found live on the RTX box).
On-glass validated on the RTX 4090 box: frame path 7007 frames p50 2.06 ms
cross-machine; DualSense + XUSB "sealed pad channel mapped"/proto=2 attach via
both the test harness and a real streaming session; phantom-sweep repro.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The test burst 3×100 1200 B datagrams into an undrained loopback socket: at
~2.5 KB kernel truesize each, the default ~212 KB rmem holds only ~80, so on
a starved CI runner (parallel release builds) the kernel silently dropped the
overflow and the recv loop could never reach 300 — surfacing as WouldBlock
after the 3 s timeout. Size the burst (3×20) to fit the default buffer even
with zero concurrent draining, and give recv a starvation-tolerant 10 s.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
log-crate events arrive through the tracing-log bridge under the shim target
"log" — normalize them back to the real module path (NormalizeEvent) so the
console's target column and the noise gate see mdns_sd::… , and suppress the
bridge's log.* bookkeeping fields like the stderr fmt layer does.
Gate known-chatty third-party DEBUG targets (mdns-sd DEBUG-logs every
unparseable multicast packet — one AirPlay device floods thousands of entries
per hour) to INFO-and-up in the ring, so ambient LAN noise can't evict the
tail the ring exists to preserve. stderr under RUST_LOG is unaffected.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Port 3000 collides with half the dev-server ecosystem; 47992 sits next
to the mgmt API (47990) in the punktfunk port family. Updates the run
scripts, systemd/scheduled-task units, Dockerfile, Windows firewall
rule + installer, packaging, and every doc that referenced :3000.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The snapshot comparison now normalizes info.version on both sides and
compares structurally — a version bump alone can never invalidate the
checked-in spec again (the 0.5.0 release tripped on exactly this; the
API surface is what drift-control protects). Snapshot regenerated so
the docs-site copy shows the current version.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The rebase onto main picked up the pre-fix loop{match} variant of the
async retrieve drain — the exact shape the Windows clippy gate rejects
(run 6722 failed on it; the while-let form passed run 6724 on the CI
branch). Restore the gated form.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The gamepad drivers have no IOCTL plane (hidclass gates the stack), so
until now the host had ZERO visibility into whether a driver ever
bound: a pad could be "created" with no driver installed and nothing
was logged. Two health fields are carved from reserved shm space
(layout-compatible; pf-driver-proto pins the offsets): driver_proto —
stamped by pf-xusb at device add + per serviced XInput IOCTL (movement
= the game-visible path) and by pf-dualsense/DS4 from its ~125Hz timer
— and driver_heartbeat. Host-side, every pad owns a DriverAttach
watcher fed from the existing service() poll: INFO on attach (WARN on
proto mismatch), and after 3s of silence ONE diagnosis WARN combining
a cached pnputil /enum-drivers store check, the devnode's CM problem
code (CM_Locate_DevNodeW/CM_Get_DevNode_Status on the instance id now
captured from the create callback, with plain-language hints: 28 = not
installed, 52 = signature/Memory Integrity, …) and the driver's debug
log path. Also fixes a real bug both SwDeviceCreate wrappers shared:
the 10s WaitForSingleObject result was ignored and the callback
HRESULT zero-initialised, so a PnP timeout read as SUCCESS (now E_FAIL
init + explicit timeout error). Failure-mode table:
design/gamepad-driver-health.md.
Linux workspace green; Windows host + drivers CI-compile only, on-box
recipe at the bottom of the design doc.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Remote debugging without shell access: a tracing layer tees every
event at DEBUG-and-up — independent of the RUST_LOG filter gating
stderr/host.log, so console-side debugging never needs a restart —
into a bounded in-memory ring (log_capture.rs, 4096 newest entries,
OnceLock singleton like config()), installed at both init sites
(stderr path in main, the Windows service file path). The mgmt API
serves it cursor-paged at GET /api/v1/logs?after=&limit= — bearer-only
and deliberately NOT on the mTLS cert allowlist (log lines can name
client identities and host paths). The web console grows a Logs page
(follow/pause · min-level filter · text search · eviction-gap badge);
polling self-paces: a non-empty page advances the after-cursor (new
query key → immediate refetch, drains backlogs), an empty page idles
at the 2s interval. OpenAPI regenerated; ring pagination/eviction,
layer wiring, and the authed route are unit-tested; Storybook story
included.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The gpu-contention plan's §5.B lever: today submit and the blocking
lock_bitstream share one thread, so under a GPU-saturating game the
pipeline serializes on the WDDM scheduling wait (1000/17ms ≈ 59fps —
the depth-1 collapse; the old 'deeper pipeline just stacks latency'
result was a same-thread implementation, not a disproof). Async mode
opens the session enableEncodeAsync=1, registers an auto-reset
completion event per pool bitstream, and moves the wait+lock+copy+
unlock onto an internal retrieve thread feeding poll() through a
channel — the exact split the NVENC guide mandates. Register/map/unmap
stay on the encode thread; teardown drops the job channel, joins the
thread, THEN destroys the session. In-flight depth is bounded by
PUNKTFUNK_NVENC_ASYNC_DEPTH (default 4, hard cap POOL-1) — both for
output-buffer reuse and because NVENC encodes the capture ring's
textures in place. Idle latency cost ≈ 0 (same-tick pickup); under
contention completed frames queue instead of stalling capture.
CI-compile validated only — on-glass A/B under game load on the RTX
box still pending (box offline).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
PUNKTFUNK_VIDEO_DROP now also covers the native data plane (N% of
sealed wire packets discarded before send in paced_submit — the same
FEC-test knob the GameStream path has; no netem/root needed), and the
probe grows the real clients' recovery trigger: the data loop publishes
the session's unrecoverable-frame count and the control task sends
RequestKeyframe when it rises (100ms poll = natural coalescing).
Together these make the IDR-vs-intra-refresh recovery A/B runnable
against any host.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
nvclocks.rs -> gpuclocks.rs. PUNKTFUNK_PIN_CLOCKS=1 now also pins every
amdgpu card's power_dpm_force_performance_level to high for the host
lifetime (prior level restored on exit) — the measured AMD encode-
latency lever: VCN per-frame time doubles when a 60fps paced trickle
lets clocks sag (8 -> 4.4ms/frame at 1440p on the 780M with clocks
hot). Root-gated by sysfs ownership; non-root degrades to a logged
recipe (validated live on the AMD box). Opt-in stays deliberate:
box-wide power-management override, wrong on battery/Deck.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Chasing the 8ms submit at 1440p on the 780M: the sampled PUNKTFUNK_PERF
split (push/pull/send) shows desc+buffersrc at ~5us, hwmap-import+VPP
CSC at ~0.2-0.5ms, and avcodec_send_frame owning the rest — so neither
a VA-surface import cache nor CSC overlap would help. Two facts landed:
(1) async_depth>=2 in libavcodec's vaapi_encode is a structural
+1-frame latency (frame N's packet only materializes when N+1 queues;
measured 18ms vs 8.3ms p50 at depth 1) — depth 1 stays the default,
PUNKTFUNK_VAAPI_ASYNC_DEPTH exists for pixel rates beyond the ASIC's
serial budget, and poll() now does a bounded in-flight wait so a deeper
depth still ships the AU as soon as the ASIC finishes. (2) The residual
send_frame block tracks GPU CLOCKS, not the ASIC: ~8ms/frame at a 60fps
duty cycle vs ~4.4ms at 120fps pacing vs 3.5ms back-to-back (270fps CLI
benchmark, even at -async_depth 1) — the clock-sag fix lands in
gpuclocks.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A/B'd on the Bazzite box (RTX 5070 Ti, KWin 6.6, driver 595, 1080p60
over the LAN): pixel-correct decode (full desktop, no tint/banding),
latency-neutral idle (p50 1.47ms RGB vs 1.52ms NV12, both 2400/2400
frames), CPU-neutral — and it deletes NVENC's internal RGB->YUV CSC
from the SM/3D engine a game saturates (video 40%+SM 15% -> video
26%+SM 2% measured on Windows). Matches the Windows host default.
PUNKTFUNK_NV12=0 restores the RGB feed; LINEAR/gamescope captures are
unaffected.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Two halves of the easy-scene p99 lever (host-latency plan Tier 1B):
CudaNoStablePerfLimit application profile (no root; NVIDIA's supported
opt-out of the CUDA/NVENC P2 memory-clock clamp, raw key 0x166c5e=0 per
open-gpu-kernel-modules#333, shipped for obs/Discord in R595) installed
into ~/.nv/nvidia-application-profiles-rc.d/ keyed on procname, opt-out
PUNKTFUNK_NV_PROFILE=0; and PUNKTFUNK_PIN_CLOCKS=1 arming an NVML
SetGpuLockedClocks(TDP, UNLIMITED) core-clock floor (base floor, boost
headroom — never a max pin) held for the host lifetime, reset-on-start
self-healing a crashed run's stale pin, NO_PERMISSION degrading to a
logged sudoers/oneshot recipe. libnvidia-ml is dlopen'd like libcuda —
no link-time dependency, clean no-op off NVIDIA.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Intra-refresh (opt-in PUNKTFUNK_INTRA_REFRESH=1 until on-glass
validated): NVENC runs a moving intra band + recovery-point SEI
(gop_size becomes the wave period, ffmpeg forces the real GOP infinite;
default fps/2, PUNKTFUNK_IR_PERIOD_FRAMES overrides; ENOSYS latches a
fallback to IDR-only). Clients request a keyframe on every
FEC-unrecoverable frame, so under intra-refresh the session glue serves
the first request instantly and suppresses the rest for a 2s window —
the wave heals loss without the 20-40x IDR spike cascade. VAAPI/software
keep IDR recovery.
Instrumentation: the wire pts now anchors at the PipeWire delivery stamp
(client-measured latency covers delivery + queue age, not just
submit->glass; repeats/synthetic stamps fall back to now), encode_us
keeps its submit->AU meaning via a separate inflight stamp, and a new
'queue' stage (delivery->submit age of fresh frames) rides
PUNKTFUNK_PERF and the web-console stats samples.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The GameStream video thread ran unboosted on Linux and the send thread
only got the Windows MMCSS call; both now use boost_thread_priority
(Linux nice -10/-5, Windows HIGHEST/ABOVE_NORMAL + session tuning).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
PUNKTFUNK_ZEROCOPY unset now defaults ON when the encode backend is
VAAPI — a stock AMD/Intel install gets the LINEAR-dmabuf -> GPU-CSC path
instead of three full-frame CPU touches (measured on the 780M at 1440p:
0.8s vs 7.9s CPU per 600 frames, pixel-identical). NVENC stays opt-in.
A dmabuf offer the compositor never accepts latches a one-shot downgrade
so the pipeline rebuild renegotiates on the CPU offer; explicit =1 keeps
erroring loudly. The EGL->CUDA importer is no longer built on VAAPI
backends (an NVIDIA box forced to PUNKTFUNK_ENCODER=vaapi now correctly
takes the passthrough instead of producing CUDA frames the encoder
rejects), and a VAAPI session landing on the CPU path warns with the
reason.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Gen12+/Arc iHD exposes ONLY EncSliceLP, so the default open fails with
'no usable encoding entrypoint'. Try full-feature first (AMD unchanged,
validated on the 780M), retry low_power=1, cache the mode per codec;
PUNKTFUNK_VAAPI_LOW_POWER pins it. Probes inherit the ladder. Docs note
the Intel HuC firmware requirement.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
wlroots injector: the virtual keyboard keymap now defers to the standard
XKB_DEFAULT_RULES/MODEL/LAYOUT/VARIANT/OPTIONS env vars (libxkbcommon
built-ins as fallback) instead of hardcoding evdev/pc105/us, matching the
libei path where the session compositor's own keymap applies. Android:
Keymap gains the same positional-key coverage for non-US layouts (+ tests).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
First-party punktfunk clients send US-positional VKs (the physical key's
US-layout VK), GameStream/Moonlight clients send layout-semantic VKs
(Sunshine's model). The SendInput injector previously resolved everything
through the SYSTEM service's layout - on a German host that is the y/z swap
and u-umlaut-on-o-umlaut scramble. GameStream ingest now tags its key events
KEY_FLAG_SEMANTIC_VK (stripped from punktfunk/1 wire events so a network
client can't flip the convention); the injector maps semantic VKs under the
foreground app's layout and positional VKs through a fixed scancode table.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- new crate::gpu (compiled on all platforms so the OpenAPI doc stays
platform-independent): DXGI / sysfs GPU inventory with reboot-stable ids
(PCI vendor:device + occurrence — LUIDs are per-boot), persisted auto/manual
preference (<config>/gpu-settings.json, atomic temp+rename with in-memory
rollback), one selection with precedence console preference >
PUNKTFUNK_RENDER_ADAPTER > max VRAM and graceful fallback when the preferred
GPU is absent, plus a live "in use" record (RAII session guard wrapped around
every encoder open_video returns)
- fix: windows_gpu_vendor derived the encoder backend from DXGI adapter 0
instead of the selected render adapter — on a hybrid box (e.g. Intel iGPU at
index 0 + NVIDIA dGPU) the backend could disagree with the GPU the capture
ring / IddCx render pin sit on. The NVENC 4:4:4 probe now also runs on the
selected adapter (was: OS default), the codec/4:4:4 probe caches are keyed
per selected GPU (were process-lifetime OnceLocks), and an explicit
PUNKTFUNK_ENCODER conflicting with the selected GPU's vendor warns up front
- mgmt API: GET /api/v1/gpus (inventory + mode + preferred + next-session
selection with reason + in-use GPU/backend/session-count) and
PUT /api/v1/gpus/preference (validates mode/gpu_id before writing);
openapi.json regenerated; the vdisplay render pin now also engages for a
console preference (not just the env pin)
- web console: GPU card on the Host page — list with vendor + VRAM,
Automatic / Prefer controls, Preferred / Next session / "In use · backend"
badges, missing-preferred-GPU warning and env-pin note; en + de messages
- Linux: a matched manual preference picks the VAAPI render node and the
NVENC-vs-VAAPI auto choice; auto mode is exactly the previous behavior
Validated live on the hybrid laptop (RTX 3500 Ada + Intel Arc Pro, which
enumerates twice — the occurrence ids disambiguate): enumerate, prefer,
bad-id 400, restart persistence, auto-restore keeping the stored pick.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- Modern branded wizard: WizardStyle=modern dynamic windows11 (Inno >= 6.6,
plain-modern fallback for older compilers; CI provisioning upgrades a
pre-6.6 Inno). Brand-mark wizard side panels + header tiles (100-200% DPI)
and a multi-size punktfunk.ico (SetupIconFile + Apps & Features), generated
AND committed by branding/gen-branding.ps1 from the canonical brand geometry.
Gotcha encoded in the script: ISCC rejects all-PNG icons, so entries <= 64px
are classic DIBs (PNG only at 128/256), and the ICO is load-verified.
- VB-CABLE actually ships now: windows-host.yml never set VBCABLE_DIR, so every
published installer silently omitted the virtual mic (broken mic passthrough
in the field). CI provisions the pinned, SHA-256-verified official Pack45
(provision-windows-punktfunk-extras.ps1) and the pack now FAILS on a
supplied-but-invalid dir instead of shipping mic-less again. Attribution per
VB-Audio's bundling grant surfaced in the visible wizard task text (vendor,
vb-cable.com, donationware) on top of the licenses notice.
- GameStream (Moonlight) compat is a wizard task (checked by default) ->
service install --gamestream=on|off writes PUNKTFUNK_HOST_CMD=
serve[ --gamestream] into host.env. Only the two canonical values are ever
rewritten - a hand-customized command line survives upgrades. Silent
installs: /MERGETASKS="!gamestream".
- Driver uninstall (field report: our virtual-device drivers survived
uninstall): new `driver uninstall [--gamepad]` removes the pf-vdisplay
device node(s) + the pf-vdisplay/pf-dualsense/pf-xusb driver-store packages,
wired into [UninstallRun] after service uninstall. Locale-safe by
construction: devices matched on unlocalized VALUES (never pnputil's
localized labels), packages found by INF content scan - validated against a
German-locale box ("Instanz-ID:" parse; 7/7 punktfunk INFs matched, no
foreign hits). VB-CABLE is deliberately left installed (shared third-party
component with its own uninstaller).
Installer compile, cargo check/clippy/fmt, and the ASCII locale gate are green;
the wizard look + uninstall flow still need one on-glass pass on a disposable
box (this box runs the live host).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Sources reorganized (client: Home/Session/Settings/Stores/Support/Trust; kit:
Audio/Connection/Gamepad/Input/Support/Video/Views) with the big files split
along the same seams.
The gamepad mode is couch-complete, and now on macOS too (the living-room
Mac case), not just iOS/iPadOS:
- GamepadSettingsView: a console-style, fully controller-navigable settings
screen (X from the launcher) — up/down moves focus, left/right steps values
(clamped, boundary thud), A cycles/toggles, B closes; the focused row shows a
one-line description. Backed by GamepadMenuList, the vertical sibling of
GamepadCarousel, and SettingsOptions — the option lists hoisted out of
SettingsView statics and shared by the touch, tvOS and gamepad settings.
- GamepadAddHostView + GamepadKeyboard: register a host end to end with a pad
— field rows open an on-screen controller keyboard (dpad grid, A types,
X backspaces, B done); the launcher carousel ends in an Add Host tile, so
the dead-end "add one with touch first" empty state is gone.
- Launcher polish: contextual hint bar with the pad's real button glyphs,
controller name + battery chip, one shared console chrome.
- GamepadScreenBackground: an animated aurora (TimelineView-driven drifting
blobs in the brand's violet family, breathing radii, slow hue shift,
legibility scrim; freezes under Reduce Motion). Pure SwiftUI on purpose — a
.metal library only bundles reliably in one of the two build systems (SPM vs
the xcodeproj's synced folders) these sources compile under.
- macOS port: settings/add-host/library present as sized sheets (a macOS sheet
takes its content's IDEAL size, and the GeometryReader-driven screens
collapsed to nothing), NSScreen-based mode lists, scroll indicators .never
(the "always show scroll bars" setting overrides .hidden), tray scrims so
scrolled rows dim under the pinned title/hints, extra title clearance, and a
PUNKTFUNK_FORCE_GAMEPAD_UI=1 dev hook — launcher/settings/add-host/keyboard/
library render-verified live on a real Mac + LAN hosts.
- GamepadMenuInput: X button support, and (re)start now snapshots held buttons
so a controller handoff press never fires twice (the B that closed the
keyboard no longer also cancels the screen underneath).
- Cleanups: one "Connection failed" alert in ContentView instead of one per
home screen; HostDiscovery.advertises/unsaved shared by both home screens.
- host: can_encode_444 stub for the non-Linux/Windows host build (the macOS
synthetic-source loopback used by the Swift tests).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adds a client-selectable **preferred codec** and wires the core + ABI + probe + Linux client to
negotiate and decode it. (Windows/Apple/Android follow in 2b.)
**Core:**
- `Hello.preferred_codec` (a single CODEC_* bit, 0 = auto) — a soft hint appended after
`video_codecs`. `resolve_codec(client, host, preferred)` now honors the preference when the host
can also emit it, else falls back to precedence (HEVC > AV1 > H.264). Roundtrip + preference tests.
- `NativeClient::connect` takes `video_codecs` + `preferred_codec`; `NativeClient.codec` exposes the
resolved `Welcome.codec`.
- ABI: `punktfunk_connect_ex7` (adds the two codec params; `ex6` delegates to it advertising
HEVC-only) + `punktfunk_connection_codec` getter + `PUNKTFUNK_CODEC_{H264,HEVC,AV1}` constants
(drift-guarded against the wire values). Header regenerated.
**Host:** passes `hello.preferred_codec` into `resolve_codec`.
**probe:** `--codec h264|hevc|av1|auto` sets the preference (still advertises it can decode all
three); the dump extension already follows the resolved codec.
**Linux client:** advertises the codecs FFmpeg can actually decode (`decodable_codecs()`), threads
the user's `codec` setting as the preference, and builds the decoder — both the software and VAAPI
paths, plus the mid-session VAAPI→software demotion — from the negotiated `Welcome.codec` instead of
hardcoding HEVC. New "Video codec" dropdown in Preferences (Automatic/HEVC/H.264/AV1).
Live-validated on the dev box: probe `--codec hevc` against a software (H.264-only) host resolves to
H.264 (graceful soft-preference fallback), no failure. clippy + core (57) + host (133) tests green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
enricobuehler