punktfunk/CHANGELOG.md

Changelog

Protocol, ABI, driver and embedder detail, one section per stable release, newest first.

This is the technical half of a release. The other half — what changed for people who use Punktfunk — is docs/releases/vX.Y.Z.md, and it deliberately contains no internal names. The two were one document through v0.24.0; they split at v0.25.0 because the engineering section had grown long enough to bury the user-facing half it was appended to. See docs/releases/README.md.

If you embed punktfunk-core, package Punktfunk, or write a plugin, this file is for you. Start with the version table of the release you are moving to, then read Breaking changes.

---

v0.25.0

407 commits since v0.24.0.

Versions

	v0.24.0	v0.25.0	Notes
Wire protocol	2	2	unchanged — every addition below is optional or capability-gated
C ABI	14	17	three steps; see below
Workspace crate dirs	22	26	pf-bitstream (+ vendored cros-codecs), pf-vkdecode, pf-dxvadec, pf-vaadec added; pf-ffvk removed
Virtual-display driver protocol	6	6	unchanged (minimum accepted still 3)
Windows virtual-gamepad channel	3	3	unchanged
Plugin index schema	1	1	unchanged
api/openapi.json	0.23.0	0.24.0	tracks API edits, lags one release by convention

crates/pf-driver-proto is byte-for-byte identical to v0.24.0 — if you ship the virtual-display driver or the gamepad channel, nothing in this release touches you.

Why the wire did not move. It grew a lot and still did not break: an optional trailing max_shard_payload: u16 on Hello (absent/0 = legacy, doubling as the renegotiation capability flag and the jumbo receive ceiling); two control messages ShardPayloadChanged (0x08) and ShardPayloadAck (0x09); a redundant desktop-audio datagram tag 0xD2 beside the plain 0xC9; a controller-audio plane at 0xD1; a new 0xCD kind 0x06; arrival flag bits 8/9; and MAX_DATAGRAM_BYTES 2048 → 9216. Old peers never send or read any of it. Bump WIRE_VERSION only when the handshake or planes change incompatibly — riding a C-ABI bump onto the wire once locked every new client out of every deployed host (ABI mismatch: client 3 host 2, observed live).

C ABI 14 → 17

• v15 — the rumble policy engine's C surface. punktfunk_connection_next_rumble_cmd, punktfunk_connection_set_rumble_quirks, PUNKTFUNK_RUMBLE_QUIRK_*. These symbols are not new: they landed while the constant still read 7 and no bump was made, so every core since has exported them while advertising a version that never promised them. A shipped binary says what it says, so this cannot be corrected retroactively — v15 is the floor that guarantees them. At or above 15 the surface is present; below it, probe for the symbol. No code changed with this bump.
• v16 — the controller-audio client surface. punktfunk_connection_next_pad_audio (the 0xD1 per-gamepad DualSense haptics/speaker plane), punktfunk_connection_set_pad_audio_caps, and the PUNKTFUNK_CLIENT_CAP_PAD_AUDIO / PUNKTFUNK_HOST_CAP_PAD_AUDIO mirrors.
• v17 — session end reason. punktfunk_connection_end_reason + the PUNKTFUNK_END_REASON_* vocabulary: after a session ends, ask why — this client closed it, the host's launched game exited (its close carried APP_EXITED_CLOSE_CODE, which the host had been sending for a long time with nothing consuming it), the host ended it cleanly, the host reported a failure, or the connection was lost. Purely a read of state the core already had: no new call is required of an embedder, a client that never calls it is unchanged, and the host sends identical bytes either way.

⚠ Breaking changes

1. 149 unprefixed macros are now PUNKTFUNK_-prefixed (139 #defines renamed in the checked-in header). Names as generic as MAX_PADS, TAG_LEN, ABI_VERSION, WIRE_VERSION, INPUT_MAGIC and the whole BTN_* / AXIS_* family were landing in the namespace of every program that included the header.

What to do: add the prefix. Values are identical; the change is mechanical.

It cannot break silently. The old spellings cease to exist, so this is always an undeclared-identifier error, never a wrong value — which is precisely the failure being removed. A colliding #define does not fail to compile: the preprocessor silently takes the last definition, so an embedder whose own header defined MAX_PADS previously got a wrong value at runtime. Associated constants are untouched; the generator already qualifies those by type name.

2. Linux hosts: the virtual Steam Deck controller moved to its own punktfunk group. The capability rode on input, which every gamepad guide tells users to join — but it can emulate arbitrary USB hardware. Operators must usermod -aG punktfunk "$USER" and re-login or the pad stops attaching. Ordinary virtual gamepads are unaffected.

3. Plugins may no longer set launch.command or the pre-launch command. Both run through a shell and are now operator-token only; a plugin that sets them is refused. Third-party plugins that populated them need updating — use the launcher_ui / xbox launch kinds instead.

4. Plugin UIs moved to their own origin on a second listener (default PORT + 1, PUNKTFUNK_UI_PLUGIN_PORT). Reverse proxies and firewalls must forward that port; a self-signed console needs it trusted separately.

Capability bits

Four added, all in the handshake's client/host capability bytes:

Bit	Constant	Meaning
client 0x04	CLIENT_CAP_AUDIO_RED	can decode the redundant desktop-audio plane
host 0x20	HOST_CAP_AUDIO_RED	is sending it
client 0x08	CLIENT_CAP_PAD_AUDIO	can render controller audio
host 0x40	HOST_CAP_PAD_AUDIO	is sending it

⚠ Pressure worth watching: client_caps has four bits free; host_caps is down to its last one (0x80); video_caps has been full since 0.23.0 (VIDEO_CAP_MULTI_SLICE = 0x80). The next video capability needs a second byte and an ABI bump — plan for it rather than discovering it.

Wire planes

• Controller audio, 0xD1 — [0xD1][u8 pad][u8 kind][u32 seq LE][u64 pts_ns LE][opus payload], one Opus frame per datagram behind a 15-byte header. PAD_AUDIO_KIND_HAPTICS = 0 is the pad's BACK channel pair (the voice coils) at 5 ms frames; PAD_AUDIO_KIND_SPEAKER = 1 is the FRONT pair at 10 ms. Best-effort like every audio plane: loss is a sequence gap concealed by the gap tracker, silence is a frozen sequence under the mic-mute discipline, host gating at −60 dBFS with a 250 ms hangover. 0xD2 (redundant desktop audio) deliberately skipped 0xD1 to reserve it for this.
• HidOutput::AudioCtl — 0xCD kind 0x06, carrying the DualSense output report's volume/routing bytes, change-only and value-deduped. Older clients drop it as an unknown kind.
• Arrival flags — bits 8 (haptics) and 9 (speaker), sent only toward a HOST_CAP_PAD_AUDIO host.
• Adaptive-trigger effects are length-bounded on encode and decode against one shared constant; the header emits uint8_t effect[PUNKTFUNK_HID_EFFECT_MAX] in place of a literal 11 (same value, so the struct layout is byte-identical). A zero-length effect body is now rejected rather than decoding as an empty — that is, a release — effect.
• Out-of-range pad indices are dropped before either rumble consumer sees them. The reorder gate bounds-checked and the legacy queue did not, so an embedder draining it could be handed an index it would use to subscript its own array. The client also clamps the host's rumble lease receive-side at 5 s, where the ceiling had been sender-side only.

Host environment variables

Variable	Default	Notes
PUNKTFUNK_AUDIO_QUALITY	high	low/standard/high; high = stereo 256 kbps. standard reproduces the pre-0.25 encoder exactly for an A/B. A typo warns once rather than silently downgrading.
PUNKTFUNK_AUDIO_REDUNDANCY	unset = automatic	on when the client supports it and the budget allows
PUNKTFUNK_AUDIO_OUTPUT_MODE	client_only	client_only/host_and_client/follow_default. Windows host only.
PUNKTFUNK_PAD_AUDIO	on	0 disables controller audio host-wide
PUNKTFUNK_PAD_AUDIO_SLOTS	1	max 4; multi-pad needs an operator to raise it
PUNKTFUNK_PAD_AUDIO_STAMPS	unset	debug bisect hook
PUNKTFUNK_WIRE_MTU	unset	pins on-wire IP MTU for all sessions; above 1500 also enables jumbo
PUNKTFUNK_JUMBO	unset (off)	fixed 9000-MTU profile
PUNKTFUNK_UI_PLUGIN_PORT	PORT + 1	the plugin-UI origin
PUNKTFUNK_LIBRARY_ART_ROOTS	platform default	art-serving roots; POSIX now defaults to $HOME
PUNKTFUNK_DECODER	client	values changed: native-vulkan · native-vaapi (Linux) · native-d3d11va (Windows) · software. Legacy vulkan/vaapi/d3d11va still accepted and migrated. Now trimmed — a trailing space used to fall through to auto silently.
PUNKTFUNK_VAAPI_DEVICE	client	new — pin the VAAPI render node
PUNKTFUNK_DUMP_VIDEO / PUNKTFUNK_AU_DUMP	client	new — capture exact decoder input / the AU as it arrived from the host
PUNKTFUNK_AU_FAULT=drop|truncate|flip[:period]	client	new — deliberate decoder-input corruption for recovery testing; native rungs only
PUNKTFUNK_NVENC_SPLIT_ARBITRATE=1	host	new — opt-in live split-encode arbitration (Linux-wired)
PUNKTFUNK_NO_AUDIO_MINT	host (Win)	new — opt out of minted endpoints; restores the name ladder
PUNKTFUNK_GPU_PRIORITY	host (Win)	removed — superseded by PUNKTFUNK_GPU_PRIORITY_CLASS, a strict superset
PUNKTFUNK_FFMPEG_LOG	client	removed with the av_log machinery

Legacy PUNKTFUNK_HOST_AUDIO=1 and PUNKTFUNK_KEEP_DEFAULT=1 still work, mapping to host_and_client and follow_default; follow_default wins if both are set. New devtest command: punktfunk-host pad-endpoint ensure|remove|status.

Security

• Origin isolation. A second listener serves /plugin-ui/** and nothing else; the console origin refuses those paths and the plugin origin refuses everything else, /api/** above all. Different origin (scheme+host+port) so same-origin policy is the boundary; same site so the SameSite=Lax session cookie still flows. Bind failure disables plugin UIs rather than falling back. x-pf-listener is stripped inbound and set by the entry; active ports republish as *_PORT_ACTIVE; the plugin origin's CSP names the console as its only frame-ancestors; the proxy allowlist drops the plugin's Clear-Site-Data, Access-Control-Allow-Origin and Set-Cookie. ⚠ The kit's postMessage(..., "*") is load-bearing — narrowing it to location.origin would target the plugin's own origin and drop every message.
• Authorization is an allowlist with a build-time gate. plugin_may_access is a list of permitted (method, path) pairs with {} segment matching, enforced by a test that walks the live route table and fails the build on any unclassified route — the block-list it replaces let new endpoints through silently. Field authority is tracked separately from route reachability: requests carry the lane that authorized them, and prep / launch.kind = "command" are operator-token only.
• Art serving gained an extension whitelist plus magic-byte sniffing, canonicalize-or-refuse, UNC refusal, config-dir exclusion and root checking, with file:// percent-decoded before canonicalization so %2e%2e cannot hide. Validation also runs at write time, so an unservable path can no longer be persisted.

Native decode — FFmpeg is gone from the client

268 files, +129k / −25k. cargo tree -p punktfunk-client-session finds zero ffmpeg. The host keeps libavcodec unconditionally (pf-encode); no host workflow, packaging script or licence file was touched.

Platform	v0.24.0	v0.25.0
Linux desktop	ffmpeg-next: Vulkan hwcontext (pf-ffvk) → VAAPI → libavcodec sw	pf-vkdecode (ash, presenter's own VkDevice, zero-copy) → pf-vaadec (dlopen'd libva, DRM-PRIME dmabuf) → openh264 + rav1d
Windows desktop	ffmpeg-next Vulkan → libavcodec D3D11VA half	pf-vkdecode → pf-dxvadec (plans into ID3D11VideoDecoder) → openh264 + rav1d
Android	MediaCodec (never had FFmpeg)	unchanged
Apple	VideoToolbox (never had FFmpeg)	unchanged

Workspace members: added pf-bitstream (+ vendored cros-codecs, compiler-enforced unsafe-free), pf-vkdecode, pf-dxvadec, pf-vaadec; removed pf-ffvk. Deleted: video_vulkan.rs, video_vaapi.rs, video_libav.rs, the libavcodec half of video_d3d11.rs, the av_log machinery, ffmpeg::codec::Id as decoder vocabulary, DecodedImage::VkFrame/::Dmabuf, the ffmpeg-fallback feature, and swscale — and with it the BT.601 default its correction code existed to undo.

Software rung: openh264 = "0.9" (BSD-2) and rav1d = { version = "1", default-features = false, features = ["bitdepth_8"] } (BSD-2). dav1d-sys was rejected because it is system-deps- only and would add a system library plus a .pc to every client package. default-features = false drops asm — rav1d's build.rs panics without nasm, unlike openh264-sys2, which degrades quietly. bitdepth_8 only ⇒ software AV1 refuses 10-bit by contract, read from the sequence header before any byte reaches the decoder.

⚠ HEVC has no CPU floor. An HEVC session that exhausts its hardware rungs tears down and re-dials advertising HEVC-less caps, and the host picks H.264 (last_rung_verdict / NoSoftwareRung). This is a first-class path, not a failure.

Rung × codec × hardware evidence (native_evidence) — the admission filter is driven by this, so an unproven rung yields only to one that is both verified for the codec and usable on the device:

Rung	Codecs	Evidence
native-vulkan	H.264, H.265 Main/Main10/4:4:4	yes — bit-exact vs libavcodec, 250/250 AUs on 3 drivers + 92-min soak
	AV1	yes — 250/250 bit-identical on one vendor, no soak
native-d3d11va	H.264, H.265	yes — frame-hash parity on RTX 4090 + AMD iGPU, 30-min soak
	AV1	not proven — decoded 4K60 once, no parity, no soak ⇒ excluded from the filter
native-vaapi	H.264, H.265, AV1	NO — has never decoded a frame anywhere; no VAAPI hardware was reachable
software	H.264 (openh264), AV1 (rav1d)	not proven; openh264 has never run on glass. No HEVC at all.

Vendor order (unchanged): Linux NVIDIA/AMD vk → vaapi → sw; Linux Intel/unknown vaapi → vk → sw; Windows NVIDIA/AMD vk → d3d11va → sw; Windows Intel/unknown d3d11va → vk → sw.

AV1 advertisement now answers from device facts (av1_hardware_decodable: Vulkan DECODE_AV1 queue op, or the Windows D3D11 import path) rather than ffmpeg::decoder::find(AV1), which was true on any build linking libdav1d. Settings migration: stored vulkan/vaapi/d3d11va migrate to native-* at decoder construction and at each dialog's lookup — the second is load-bearing, since an unmatched value renders as "Automatic" and a save would silently rewrite the preference.

The three decode data-loss bugs

AV1 sub-frame truncation — shipped in v0.24.0, host-side. NVENC sub-frame readback has two halves armed by different conditions: build_init_params arms the writer from subframe_on alone, while the chunked reader additionally requires slices >= 2 — and resolve_slices returns 1 for AV1 unconditionally, because AV1 partitions via tiles, not slices. So an AV1 session told the driver to publish tile-by-tile and then took only the first tile. Measured at 4K60: every AU carried a header declaring two tile rows plus a single Tile Group OBU with tg_start = tg_end = 0; libdav1d rejected 835/836 AUs. NVIDIA's hardware decoder accepts it (so Vulkan Video looked healthy at 60 fps); its DXVA path did not. 1080p is one tile and unaffected; 4K splits into two tile rows and loses half the picture. Fixed by disarming sub-frame for AV1 while leaving split_mode untouched — AV1 keeps every engine. Arming the reader instead is not a drop-in: the reader cuts at bitstreamSizeInBytes on the reasoning that slices are contiguous Annex-B, which AV1 OBUs are not. Post-fix 654/654 clean. The test that had pinned the old behaviour as correct is replaced by one pinning the disarm, plus one comparing the reader's gate against the writer's — the comparison nothing made.

HEVC DPB from the level ceiling — new in this release, client-side. dpb_limit computed max(A-2_level_ceiling, sps_max_dec_pic_buffering_minus1 + 1). HEVC equation A-2 is a ceiling on what an SPS may legally signal, not a statement of need, and it branches on picture size against the level's MaxLumaPs. The host is blameless: NVENC autoselects L5.1 because the bitrate exceeds L5.0's ceiling, and signals six pictures at every resolution. At 720p and 1080p the A-2 branch yields 16 frames / 17 slots — one more than NVIDIA's maxDpbSlots of 16 — so every AU fell outside device caps, flushed, waited for an IRAP, and the fresh IDR needed 17 again; rungs exhausted, and there is no software HEVC. It hid because the path was only ever exercised at 4K, the one size that falls through to the honest answer. Fixed to buffering.min(16): the max() bought no tolerance, since Dpb::needs_bumping already evicts at the signalled depth — it only over-allocated ten surfaces per 1080p session. H.264 escaped by luck (its ceiling lands at 13 for 1080p) and is left alone, because H.264's DPB size genuinely is level-derived absent a VUI bitstream_restriction.

rav1d aborts the process — new in this release, client-side. rav1d 1.1.0 abort()s on any decode error while holding one frame context: the c.fc.len() == 1 branch decodes inline, always finishes in rav1d_decode_frame_exit which unconditionally takes frame_hdr, then on Err re-enters an on_error whose first act is frame_hdr.as_ref().unwrap() on the None it just left. The panic unwinds into dav1d_send_data, which is extern "C" ⇒ panic_cannot_unwind ⇒ abort(). No catch_unwind, no rung demotion and no refusal can catch it, and every rav1d_* entry is pub(crate), so no in-process guard is possible. 4K was only where the first error happened — the CPU rung does 35–39 fps against a 60 fps stream, the backlog stopped draining, the pump flushed to live, and the next AU referenced undecoded frames. Fixed by opening with n_fc >= 2 and asking dav1d_get_frame_delay what the settings actually bought. Decode now drains past the first EAGAIN, which is why two frame contexts cost no latency (20–42 ms/unit at n_fc=2 vs 21–53 at n_fc=1). On glass: 4K60 AV1 was SIGABRT on the second frame every run; after, exit 0 with 1204 frames and 13 decode errors recovered across 17 backlog flushes. Reported upstream as rav1d#1497 with a reproducer. Does not make the CPU rung panic-proof.

Settings loader BOM — shipped in v0.24.0, client-side. .and_then(|s| from_str(&s).ok()) turned every parse failure into Default. Set-Content -Encoding UTF8 writes EF BB BF, serde_json correctly rejects at byte 0, and every setting vanished silently. A shared load_json_or_default now strips the BOM and warns with path plus serde line/column, covering settings, known-hosts (where a BOM silently unpaired every host) and profiles on both desktop clients. The result is deliberately still Default, never an error.

Other decode/encode

• Intel Arc pNext ordering. vkGetPhysicalDeviceVideoCapabilitiesKHR was called with the codec caps struct chained before VkVideoDecodeCapabilitiesKHR (push_next prepends). Arc/Windows fills those two by position, not by sType, and returned them swapped — we read a level as a capability bitmask. Measured A/B: decode_flags_raw=12 max_level_idc=1 before, decode_flags_raw=1 max_level_idc=12 after. NVIDIA and RADV dispatch by sType, which is why the fleet stayed green. ⚠ This does not yet give Arc Vulkan Video — the refusal only moves down: the device advertises only COINCIDE, and its NV12 coincide entry does not advertise SAMPLED usage, which the zero-copy presenter needs. Unresolved whether that is ours or an Intel constraint.
• NVENC split encode. The 10-bit rule sat above the pixel-rate arm and took no codec, so it vetoed 10-bit 4K120 — the exact case the pixel-rate arm exists for — and applied an HEVC-Main10-on-Ada result to AV1 10-bit, which has no such measurement. Re-measured on Ada and Blackwell: 4K60 2.06×, 5120×1440@240 1.31×, 4K120 1.89× — split wins at every mode on both architectures, including the configuration the veto came from. New order: env override → pixel-rate arm (now taking max_forced_split_mode(engines), not a hard-coded 2) → HEVC-Main10-below-the-bar → AUTO. Operator over-asks are clamped with a warning because the driver honours an over-ask and silently encodes narrower. Also newly logged: HEVC + plain AUTO + sub-frame is silently single-engine — the fleet's default shape, and nothing said so. ⚠ Unvalidated consequence: 5120×1440@240 Main10 now clears the pixel-rate bar and will be forced to split — the exact configuration the old veto came from. PUNKTFUNK_SPLIT_ENCODE=0 is the escape.
• PyroWave on Windows stamped over the host's GPU scheduling policy. It raised the process WDDM class to HIGH at every session open, while auto_priority_gate already owns that process-wide — starting at HIGH, upgrading to REALTIME once safe, and leaving a monitor that drops back when VRAM tightens (REALTIME + NVIDIA + HAGS + near-full VRAM is a documented NVENC hang). Opening PyroWave stamped HIGH back and orphaned the monitor's decision. Removed rather than reconciled.
• A pf-vkdecode AV1 use-after-free fix had stabilised the wrong pointer — OwnedStdAv1SequenceHeader kept the Std struct inline, so pStdSequenceHeader was a dead stack address; it worked only because NVIDIA happened to retain pColorConfig instead. Std structs are now boxed inside each owning wrapper, and create-time arrays are fields of the stored parameters assembled at their final address. The same shape was fixed pre-emptively in H.264/H.265.

A/V sync — it did not previously exist

The host has always stamped pts_ns on every audio datagram. Every client decoded it into AudioPacket / AudioPCM and never read it. Video's pts_ns was used end to end; audio free-ran at whatever depth its jitter ring reached; nothing compared them. The A/V offset was an emergent property of buffer depths — it moved whenever the ring ratcheted under underrun pressure, and it got worse every time video got faster, because a quicker decoder lowers the video leg and leaves audio's where it was. That is why shaving milliseconds off the audio budget had never helped.

Two host defects were prerequisites:

• pts_ns was stamped at encode time, inside the loop draining an already-accumulated chunk, so every frame of a chunk carried near-identical timestamps describing when we got round to encoding. Now derived from the chunk's arrival instant minus queued-frame duration, re-anchored per chunk.
• The host did not pace. One capture callback hands over a whole quantum (5 ms honoured, 21.3 ms on a VM, where stock PipeWire raises min-quantum to 1024), drained into back-to-back send_datagram calls — a 4–5 frame burst then ~21 ms of nothing, which a ring could only absorb by standing a burst period deep. Frames now leave on the audio clock (FRAME_INTERVAL 5 ms, PACE_MAX_SLEEP 10 ms, PACE_REANCHOR 100 ms). Costs no average latency.

audio_e2e = (now + buffered_ahead + clock_offset) − pts_ns
av_offset = audio_e2e − video_e2e            (> 0 ⇒ audio behind the picture)

AvSync EWMAs it (AV_EWMA_TAU_MS = 2000), ignores anything inside AV_DEADBAND_MS = 10, waits AV_MIN_OBSERVATIONS = 100 before a first correction, and refuses rather than clamps beyond AV_SANE_LIMIT_MS = 1000 — a wall-clock step must not steer the ring.

⭐ Video is the master, and continuity outranks sync. JitterPolicy::set_sync_target takes only a request, clamped between the existing underrun-driven adaptive floor and the hard cap: a link whose jitter genuinely needs more buffer than the picture is away keeps its buffer, and the residual is reported rather than forced. None/nil reproduces prior behaviour bit-identically, which is how the four rings adopted it one at a time.

Per client: the Rust desktop reference is a new video_e2e_ns atomic beside clock_offset, written by the presenter and read by the audio thread. Android publishes OnFrameRendered — the one place that knows a frame latched — raw, not floor-shaved (the HUD shaves the OS present floor; sound must reach the ear when light reaches the eye), and stays inert below API 33 rather than substituting the release instant, which targets a future vsync 8–21 ms ahead of glass. Apple publishes its LatencyMeter sample as an expiring level, because that client has a backgrounded keep-alive that keeps audio playing while dropping video decode; its clamp raises the ceiling to the floor rather than min(max(…)), which on a device whose callback quantum alone exceeds the hard cap would otherwise hand back the cap, silently below the continuity floor.

Escape hatches: PUNKTFUNK_NO_AV_SYNC=1 everywhere, plus adb shell setprop debug.punktfunk.no_av_sync 1 on Android (a launcher-started app inherits no environment). Observability: buffer_ms/target_ms had only ever been a tracing::debug! line — and on a Deck the client runs under Steam's reaper with stdout on a pipe nobody can read, so the one number identifying a deep ring was unobtainable on the device reporting the latency. Now on the HUD and in the 1 Hz stats log on every client.

Decode-target aliasing — caught before it shipped

⚠ None of this ever shipped. git ls-tree v0.24.0 crates/ has no pf-vkdecode, pf-dxvadec, pf-vaadec or pf-bitstream; v0.24.0's decode rungs were libavcodec. This was a ship-blocker for the new stack, cleared — not a field bug.

Three of the four native rungs released a picture's surface inside the plan→submission conversion, then assigned the decode target a slot. SlotMap::assign returns the lowest free slot — the one just vacated. The submission then named one surface as both decode target and its own reference: CurrPicTextureIndex == RefFrameMapTextureIndex[k] on DXVA, or pSetupReferenceSlot sharing an array layer with pReferenceSlots on Vulkan. Decode into the surface you are predicting from.

• AV1 / D3D11VA — AV1 applies refresh_frame_flags after decode (7.20), so "read a slot then overwrite it" is the ordinary case: 268 of the vendored vector's 274 frames, first at frame 6.
• H.264 / both Vulkan and D3D11VA — H264Planner snapshots dpb_refs in begin_picture, before 8.2.5 marking and the C.4.5.3 bump, so a picture the sliding window unmarks and the bump evicts lands in both dpb_refs and dpb.removed. Both conditions coincide only in low-delay H.264 — and NVENC guarantees it (max_num_ref_frames = 3 alongside max_dec_frame_buffering = 3, plus max_num_reorder_frames = 0). Result: 297 of every 300 access units of every stream a punktfunk host emits, at every resolution, on both rungs.
• H.265 is exempt, now measured rather than argued — 0 of 120 aliases, with a counterfactual that moves the snapshot one call earlier and reproduces 115 of 120.
• VAAPI's exemption was incidental: the precondition is fully present (117 of 120 AUs) but plan_to_va never invents a surface. That held only because three call sites happened to write free_surface() and surface_table() adjacently; acquire_target now returns index, surface and table together so a later edit cannot split them.

Fix is uniform: the plans grow release_after_decode, conversions hand removals back, callers release once the decode op is issued. Costs no slot (SlotMap::new allocates max_dpb_frames + 1). Both rungs hold the Result rather than ?-ing it so the deferred release runs on failure paths — seven exits sat between conversion and release, each of which would have leaked a slot.

Why four gates missed it, all recorded: the conformance vector is structurally blind (level 1.3, no VUI bitstream_restriction ⇒ a 7-frame DPB against 2 reference frames, and it reorders) and passed 250/250 for two milestones; a test had encoded the bug as correct; another assertion was vacuous (it asserted the decode target was never also a reference while handing every picture its own never-reused surface id — distinct integers cannot collide); and it streamed clean — "the 2026-08-07 field sessions that looked clean were looking at wrong pixels."

gpu_parity is now 11 legs (not 9 — that note was written mid-PR): each decodes a vendored stream, reads back every output frame's NV12, crops to the display region and SHA-256s in display order against libavcodec goldens, frame count and flush tail included. The three new legs are our own encoder's output rather than conformance vectors — H.264 because the vector is blind to the shape, H.265 because an exemption with no stream behind it is how the H.264 defect survived two milestones, AV1 because the vector is one tile on all 274 frames while our encoder splits 4K into two tile rows, so every tile array the conversions fill had only ever been written at index 0. video_vaapi_native parity is new entirely: 7 legs, bit-identical on RDNA3.

⚠ Promoting D3D11VA AV1 to verified changes rung selection on Windows Intel/unknown vendors, not just a label. VAAPI stays verified = false deliberately — one vendor, never soaked; flipping it would move auto off Vulkan Video on every Linux AMD/Intel client including the Deck.

FFmpeg 9, and the Arch soname trap

pf-encode now builds against FFmpeg 9. The host still links libavcodec unconditionally; the client has none (see above).

⚠ pacman is the only one of our packaging formats that does not derive dependencies from ELF DT_NEEDED. rpm auto-generates libavcodec.so.62()(64bit), dpkg-shlibdeps emits libavcodec62, nix pins the closure — but a bare depends=('ffmpeg') let pacman -Syu walk the host across a soname bump with no warning and no conflict. FFmpeg 8 → 9 (2:9.0-5: libavutil .60→.61, libavcodec .62→.63, libavfilter .11→.12, libavdevice .62→.63, libswscale .9→.10) therefore bricked every Arch/CachyOS install: the dynamic loader cannot start the binary, so it is exit 127 before main() in a systemd restart loop, with nothing in the host's own log to explain it. ldd /usr/bin/punktfunk-host | grep "not found" is the one-line diagnosis.

⭐ The fix is SONAME deps, not a hand-written version bound: depends=(… 'libavcodec.so' 'libavutil.so' …). Arch's ffmpeg declares matching provides=(libavcodec.so=63-64 …), and makepkg rewrites each bare libfoo.so into libfoo.so=<soname>-<arch> by reading the built binary's DT_NEEDED — so the bound tracks whatever FFmpeg the builder linked against with nothing to maintain across the next bump. A literal ffmpeg<2:9 would go stale on every bump. pacman now refuses the upgrade instead of bricking the install. All seven libs are listed even though --as-needed currently drops two: an unlinked soname is left bare by makepkg and satisfied by any ffmpeg, so listing it costs nothing and a future link picks up the bound automatically.

Linux playback filled the buffer ceiling

The PipeWire playback callback sized its writes from the mapped buffer's capacity — PipeWire's quantum limit, 8192 frames ≈ 170 ms — instead of the graph's per-cycle ask (pw_buffer.requested). Every cycle queued up to 170 ms of PCM downstream of the ring and taught JitterPolicy that the device drains 170 ms per callback, so the underrun floor (want + one frame) rose above any depth the A/V sync loop could request: sync measured audio ~280 ms late and was then forbidden — by its own continuity rule — from draining it. The first on-glass run of the latency overhaul showed exactly that: audio buffer 272 ms, a/v +284 ms, stable. Now honours requested (capacity remains both the ceiling and the fallback when requested == 0) and logs requested-vs-capacity once per stream. Needs libpipewire ≥ 0.3.49; every ship target clears it.

Windows audio substrate

The host now mints its own devnodes from Valve's INFs (SteamStreamingSpeakers.inf / SteamStreamingMicrophone.inf under {CommonProgramFiles(x86)}\Steam\drivers\Windows10\…) instead of bundling VB-CABLE.

• Two persistent endpoints, Punktfunk Speakers (client-only loopback sink — the wiring plan parks the default playback on it during a stream, its WASAPI loopback feeds the encoder, the host stays silent) and Punktfunk Microphone (host writes decoded client voice into the render side; the capture side surfaces as the mic). Both survive host restarts and re-resolve by marker.
• Identity is the recorded endpoint id, never the name — a minted instance is name-identical to Steam's primaries. Durable marker PunktfunkAudioRole (1 = Speakers, 2 = Mic) under Device Parameters. Name stamping is device-desc + device-name only: a wider stamp set makes AudioEndpointBuilder re-mint under a new GUID. Best-effort via the SYSTEM ACL route; on failure the endpoint still wires and simply keeps the driver's default name.
• Format stamps are per-direction. Render gets the PCM16-device / float-mix stereo split; capture gets the device-format key only — mix and host-format keys are render-engine properties, and stamping them onto a capture endpoint breaks its shared-mode graph (IsFormatSupported reports 2ch/48k fine, Initialize then fails 0x88890008).
• MintedIds is tier-0 in the wiring plan. The mic takes its minted device outright (paired by provider id — a name search cannot distinguish it from the primary); the loopback prefers the minted sink at the head of the silent tier. Below that the old ladder is unchanged: Steam primaries → cable → real hardware. PUNKTFUNK_MIC_DEVICE still beats everything.
• Mic-vs-loopback arbitration: the mic may hold the Streaming Microphone only while the loopback still gets a non-last-resort pick; otherwise the loopback takes it and mic_withheld is set. This fixes a field case where a headless Steam-only host streamed silence.
• New AudioReadiness — Full / AudioOnly / MicOnly / Nothing, logged on every plan change and surfaced at GET /api/v1/status → RuntimeStatus.audio (AudioWiring, Windows-only, absent before the first wiring pass; a status poll triggers no COM work or IPolicyConfig writes). The console Dashboard renders it as an "Audio wiring" card.
• Requires Steam installed (never running) — without the INFs the host streams video only, and picks the drivers up automatically if Steam is installed later. Opt out entirely with PUNKTFUNK_NO_AUDIO_MINT, which restores the previous name-based ladder exactly.
• ⚠ VB-CABLE is no longer bundled but is deliberately NOT uninstalled — it is a third-party shared component other apps may use, and it stays in the ladder as a live fallback. Demoting it was considered and rejected: on a box where minting transiently fails, that would let the Steam Streaming Microphone outrank an installed cable, steal the silent sink and make stream audio audible on the host.
• ⚠ The minted endpoints survive Punktfunk's uninstall by design (they are plain instances of Steam's drivers and are inert without the host). There is no user-facing removal path; cleanup is the devtest punktfunk-host audio-probe cleanup.
• New devtest: punktfunk-host audio-probe ssm|sink|sss-primary|mint|plan|micpitch|micpins|cleanup. plan is the field-triage command; micpins maps exclusive+shared IsFormatSupported across {1,2}ch × {16,32}bit × {44.1,48,96}kHz on both mic pins.

Apple audio

• The microphone was never in the render graph. On the combined (voice-processing) engine — made default a week earlier and never run on a device — the input node carried a tap and no connection, so nothing pulled it: the IO unit came up, the recording indicator lit for a beat, and not one buffer ever reached the tap, with no error and no failed start. The 10 s silence tripwire counts captured frames, so it never fired. Input now runs through a silent sink into the main mixer at outputVolume = 0 (Apple's own voice-processing sample topology). Two more: the tap read the input format before prepare(), and enabling voice processing swaps in the VPIO unit and renegotiates, so the pre-swap read could be 0 Hz / 0 ch; and a mic-chain failure on the voice-processed engine took the whole uplink down for the session — it now falls back to the split path, because the mic outranks the AEC.
• No packet-loss concealment on the one client that decodes Opus in core. Linux, Windows and Android all feed an AudioGapTracker and synthesize libopus PLC; the in-core path had the tracker sitting unused in the same crate and decoded only packets that arrived. At ~200 packets/s of 5 ms frames every lost datagram was a hard time-domain gap — one click per loss. The redundant plane (0xD2) hides single losses, so the survivors were exactly the burstier gaps that most needed concealing. Concealed frames now land in front of the arriving frame in one contiguous buffer, a DTX marker advances accounting without being decoded, and the output buffer is pre-sized for a full concealment run so the borrow-until-next-call pointer cannot dangle (50 ms cap).
• The Apple jitter ring never grew. The shared Rust JitterPolicy has an adaptive target floor; the hand-written Apple mirror mirrored the shed half but not the growth half, pinning its target at the 20 ms base forever. On Wi-Fi that bunches arrivals, 20 ms is regularly shorter than one delivery stall, so the ring re-primed through every stall for the whole session. Now the full note_read mirror: 3 underruns in a 5 s window grow the target 10 ms (capped at CoreAudio's 70), 30 s of quiet steps back, and the write-side hard trim follows the grown target.

Clients

• Nothing in the desktop console had ever been clickable. SkiaOverlay::handle_event matched only KeyDown and TextInput, so every mouse button, wheel and touch contact fell past the console into the run loop, which routes pointer input exclusively at stream.capture — None while browsing. New Overlay::handle_pointer carries mouse/touch in swapchain pixels; the run loop converts (it owns the window and hence display scale); the console hit-tests the rects it drew last frame. Only direct touch devices are offered — an indirect trackpad already drives the mouse. Widgets act on press, not release, because both carousels scroll the focused item toward centre and what you pressed would slide out from under your finger. Host menu on Up from a saved tile; UpdateHost edits in place (remove-and-re-add would silently drop the fingerprint, learned MAC, pinned cards and profile binding), and ForgetHost arms on first press and fires on second.
• Discovery went permanently deaf three ways, each needing an app relaunch: a failed resolve was never retried (browseResultsChangedHandler fires only when the result set changes, and a host whose resolve failed is still in the set); a stuck resolve never ended (NWConnection has no timeout, so the throwaway UDP flow could sit in .preparing forever, and a service with a connection in flight was skipped); and an NWBrowser parking in .waiting was ignored — which is exactly where iOS's local-network privacy prompt lands on first launch, and granting it does not revive the browser that was already waiting. A 1 Hz sweep now times out stuck resolves, retries failed ones on a 1→30 s backoff, and re-arms a dead browser; the advert's TXT is re-read on every browse report. discovery::Rescan forces a fresh mdns-sd query — the browse otherwise re-queries on a doubling backoff capped at one hour, so a long-lived browse is effectively passive. ⚠ clients/windows/src/discovery.rs is a second copy of the browse that the earlier IPv4 pinning missed; it took an arbitrary first address, so a host whose OS responder answered AAAA rendered a card that failed on every click.
• Phone gyro mirror, off by default, player 1 / wire pad 0 only, and only while that pad has no motion source of its own. iOS/iPadOS only on Apple (DeviceGyro wraps CMDeviceMotion at ~100 Hz on a dedicated serial queue — the controller path's main-queue delivery is a known jitter source); Android phones with a gyroscope at ~200 Hz with maxReportLatencyUs = 0, since batching is poison for gyro aim. Both rotate from the device's natural frame into the controller frame by interface orientation, and both send one zero-gyro sample on stand-down — the host holds motion as state and re-emits it, so a leftover nonzero angular velocity reads as endless rotation.
• Safe-area resolution is purely a sizing change — no layout change, no input change; pointer mapping follows for free since both clients derive the picture rect from the live host mode. Full native height, width less left+right safe insets. Portrait settings screens report the housing on top with zero horizontal insets, so the portrait top inset stands in (gated so an iPad's status bar never fabricates one). Android adds the rounded-corner radius, which it does not count as cutout. Both even-floor and clamp, because validate_dimensions rejects odd dimensions and an inset subtraction lands odd about half the time.
• Gamepad UI: six sections (Stream · Video · Audio · Controller · Interface · Profiles, plus Input on the desktop console) walked with L1/R1 with per-section cursor memory; 12 palettes under one shared ui_palette key, Violet keeping its explicit sixteen colours so existing installs are an identity transform. Presentation only → device preference, never part of a profile. Palette maths ported three times (Rust/Swift/Kotlin) with the same assertions pinned in each language; every_palette_is_multi_tone fails under 45° hue spread and caught Ember at 35° and Graphite at 3°. Three render-only findings: additive blending blows out over a pale ground, a white scrim at the dark field's strength bleaches the gradient, and white glass over a bright field needs more body.

Session and game lifetime

• PunktfunkEndReason replaces a single "closed" bit (ABI 17, additive, wire untouched). Five values — local, game exited, host ended, host error, lost — classified by the connection watcher from close codes already on the wire (APP_EXITED_CLOSE_CODE had been sent for a long time with nothing consuming it). Latched before the shutdown flag, because the two are read by different threads and the reason must never arrive second. Exposed as punktfunk_connection_end_reason + is_normal(). Shells fall back to the old wording when there is no verdict (older core, or a close that raced the read).
• The Steam Running registry hint was an unbounded veto. Honouring it reset the absence window every pass, so a flag Steam left set — Steam crashed, was closed first, the game re-parented — pinned a lease in running for the life of the host process. The absence timer now runs regardless; past VETO_LIMIT (30 s) with nothing of the game on the box, the session ends anyway and logs at WARN. Extracted as a pure exit_confirmed(gone_for, hint_running) with tests — the watch loop polls a live process table and cannot be unit-tested, which is exactly how the unbounded veto shipped.
• New launchreg.rs: one record per (client fingerprint, library id), written at launch and independent of the termination policy. The old fingerprint-keyed reclaim only ran under GameOnSessionEnd::Always, so under the default Keep nothing was recorded — and a client retry re-sent Hello::launch verbatim, which the host obeyed unconditionally. Steam/Epic URIs hid it (the launcher just focuses the running copy) but a gog:/custom: target genuinely started a second instance over the same save files. The same retry also minted a fresh launch_stamp, so procscan refused to adopt a game older than 2 s and a reconnected session lost game-exit detection for the rest of its life. Identity now flows backwards from the watcher, which publishes the concrete ProcRefs it adopted; liveness is Scanner::alive over that recorded set, re-verified by (pid, start). Tradeoffs: a custom: command with no detection hints stays Unknown forever (trading exit detection for not double-spawning), and IN_FLIGHT_WINDOW is a fixed 90 s, deliberately not disconnect_grace_seconds.
• A launcher entry is LeaseKind::Untracked unconditionally, checked ahead of nested/child/spec. Its lifetime previously depended on invisible state: launcher not running → live child → Child lease → quitting the launcher ended the session; launcher already running → command forwards and exits inside SHIM_WINDOW → Untracked → session persists. Steam Big Picture is a mode, not a process (and on a Deck it is always running); Heroic is single-instance Electron. The real trap was the GameStream path, whose GsApp intermediate silently dropped the field.

Library and plugins

• Store claims keep identity across the scanner-to-plugin handover. library.json gains a v2 {entries, claims} shape that reads the old bare array unchanged and rewrites on first mutation. PUT /library/provider/{p}?store=<s> claims a store; entries then surface as <store>:<external_id> rather than custom:<id>, so entry ids, GameStream app ids, client art caches and Moonlight pins all survive. One provider per store (409 otherwise); while a claim is held the matching built-in scanner is skipped, so the two never double-list.
• GET/PUT /library/scanners is now a sources endpoint over the same disabled-set file.
• New entry fields: role: game|launcher; launch kinds steam_ui (bigpicture|desktop), launcher_ui (platform-gated, 400 on invalid) and xbox.
• Plugin kit 0.3.0 adds a ./library subpath: defineLibraryPlugin plus ported total parsers — text VDF/ACF, the binary shortcuts.vdf walker with CRC-32 appid derivation, read-only immutable SQLite, a registry wrapper that refuses HKCU, path-confinement joins. GET/PUT /__config returns {schema, value} and persists raw, so a plugin with settings need not ship an SPA.

Platform and packaging

• The client's config writer falls back to an in-place write when the atomic replace is unavailable, verifies it by reading the bytes back, and records the last persistence failure centrally so the UI can surface it. Scratch files are now per-process, closing a real collision between the five processes that write these stores (shell, session, console UI, CLI, Decky) — one could previously rename its half-written temp over another's target.
• Host send pacing gained a pure, unit-tested budget function: oversized frames are budgeted at the pacing rate with a 100 ms absolute ceiling rather than compressed into one frame interval. Steady-state schedules are byte-identical, the legacy behaviour stays reachable via an environment escape hatch, and the GameStream-compatible path is untouched.
• Mid-session shard renegotiation is gated off for PyroWave sessions, which parse the video stream in windows fixed at session start — re-sizing mid-stream would corrupt the parse. Those sessions get the next-session clamp only and are excluded from jumbo. The ABR decode-cap latch likewise does not apply to PyroWave, where adaptive bitrate is open-loop by design.
• The Deck's Vulkan compatibility layer is built from source, pinned to the same upstream revision as the host's own packaged build — bump both together. ~4 MB of app content replaces a 94 MB external extension, and Flathub is no longer needed at install time. ⚠ subprojects/vkroots is a gamescope submodule and flatpak-builder clones submodules by default; declaring it again as an explicit source breaks the build during extraction. glm and stb are .wrap files, not submodules, and do need explicit sources.
• Build-container images push to an authenticated registry endpoint, and :latest is reconciled against the content key on every push to main — an out-of-band tag move is detected and repaired rather than silently inherited.
• Windows pad drivers publish their sequence counters with release ordering (the host was already loading with acquire and pairing with nothing) and serialize the output-ring publish. The /dev/uhid event ABI, previously transcribed into all five Linux gamepad backends, is consolidated into one module.

Verification status

Honest about what has and has not been on hardware, because several things in this release have not:

• Controller audio has never run on a real DualSense. Its entire verification is unit tests and compile checks, and its rumble arbitration rests on an explicitly retracted assumption about whether the voice coils and the rumble motors are the same actuators. The evidence-based 500 ms idle window is correct either way, but the underlying exclusivity is unsettled. Android's arbiter is the evidence-based one; the desktop twin and the coil restore on Android's stop path are owed. Some Android OEM kernels refuse the isochronous claim outright, which degrades to ordinary rumble.
• The plugin-UI origin split is validated against a fake console and a fake plugin, not yet in a real browser.
• The packaging default-on changes have had no installer run or package build.
• No launcher tile has been clicked on a real host — the first source that would publish one does not exist yet.
• Desktop-audio, packet-sizing and iPad-pointer work is build-verified only.
• ⚠ The FFmpeg-deletion milestone itself has never executed on a GPU. It was gated on cross-clippy, 160 tests, a workspace check and an ffmpeg count of 0 in the client / 2 in the host. The software on-glass check, the D3D11 and VAAPI AV1 hardware legs and the field bake were all owed at merge; later commits closed some of that but not all. The "no FFmpeg" claim is verified by cargo tree and a notices-generator mention count, not by inspecting a shipped binary.
• ⚠ pf-vaadec has never decoded a frame anywhere — no VAAPI hardware was reachable. It is the first rung on Linux/Intel and unknown vendors; the evidence filter bars it there in favour of pf-vkdecode, but an explicit pin reaches it.
• openh264 has never run on glass; the H.264 software rung is unit-tested only.
• native-d3d11va AV1 is deliberately verified = false — one 25 s 4K60 session, no parity.
• Split arbitration is opt-in and Linux-wired only; the Windows arm is built and unit-tested but not on hardware. The 5120×1440@240 Main10 behaviour flip is explicitly unvalidated and is named as the first thing to re-measure.
• Software throughput is unmeasured in general — the CPU rung does 35–39 fps at 4K AV1 against a 60 fps stream, which is why the backlog flush that triggered the rav1d abort happens at all.
• The Apple mic fix is a proven root cause, not a verified session. Its own commits call it "a strong inference plus one proven logic defect rather than a confirmed fix" and close "awaiting the reporter's on-device confirmation" — which nothing later in the range records. It also leaves a known gap: nothing reports whether the uplink actually opened, so the HUD still offers a Mute Microphone button over a session that may be sending nothing.
• The Windows audio substrate is, by contrast, well-evidenced on hardware — repeated "measured on the target box", a live bisect on a fresh endpoint, and a micpitch proof reading 440 Hz in → 440 Hz out at exact peak. The one thing not evidenced is a real client speaking through the minted microphone end to end; the pitch proof is probe-driven.
• The phone-gyro mirror is not recorded as hardware-verified — remap matrices are pinned by unit tests in both languages, but there is no "played a game with a clip-on pad" evidence in the tree.
• The iOS gamepad-UI pale-palette sweep on glass is still owed, per its own commit.
• ⚠ The CI runner scripts are hand-installed (/usr/local/bin/ci-docker-prune.sh, /usr/local/sbin/ci-docker-reclaim.sh). Merging does not deploy them — both runner hosts need the files copied out of scripts/ci/, and the missing 192.168.1.58:5011 insecure-registry entry on one host is routed around, not fixed.