1153 Commits

Include renames

Author	SHA1	Message	Date
enricobuehler	0a710fdfe5	refactor(frame): narrow make_device's unsafe to the FFI calls themselves ... First slice of the `unsafe_op_in_unsafe_fn` work on the Windows-gated code, and a deliberate demonstration of the shape the rest should take. `cargo fix` is NOT the tool here. Its suggestion for this lint wraps the entire function body: pub unsafe fn make_device(..) -> Result<..> { unsafe { ..40 lines.. }} One block spanning everything, `{ unsafe {` / `}}` bracing, and — the actual problem — zero change in how much code runs unchecked. It makes the status quo explicit rather than smaller, which is the opposite of the goal. Narrowed by hand instead: four blocks, each around the FFI call that needs it (`D3D11CreateDevice`, two `SetGPUThreadPriority`, `SetMaximumFrameLatency`), each with its own SAFETY note. The `Option` handling, the `?`, the `if let Ok(..)` casts and the `tracing!` calls are back outside, where the compiler checks them. Roughly 40 unchecked lines become 6. Fixed two of my own mistakes in the same pass: `elevate_process_gpu_priority` and `auto_priority_gate` are ALREADY safe fns, and I had wrapped their calls in `unsafe` out of momentum. Unnecessary `unsafe` is not free — it is the same "this is scary" noise the lint exists to remove, and the compiler only warns about it. That mistake is a good argument against doing the remaining ~590 sites mechanically. pf-frame drops 23 -> 19 E0133. The remaining three functions in this file need a different answer, not just narrowing: `enable_inc_base_priority` takes no arguments and touches only the current process's own token, so it has NO caller contract and should become a SAFE fn with unsafe blocks inside — which removes an `unsafe fn` rather than decorating it. Same likely for `hags_enabled(LUID)`, whose only argument is a plain value. Verified on Windows .47: pf-frame checks clean at EXITCODE=0, no errors, no unused-unsafe warnings.	2026-07-28 21:31:42 +02:00
enricobuehler	a3843b6996	test(encode): the VAAPI dmabuf path verified on AMD silicon, and its owner-only fields say so ... Closes the last unverified leg of the AvBuffer work. `DmabufInner` was the heaviest ownership change in the crate — four owned objects (DRM device, derived VAAPI device, DRM-PRIME frames ctx, filter graph) whose eight failure branches each repeated the same four-line unwind, once inside a macro, plus a ninth copy in `Drop` — and it had no coverage at all. `vaapi_cpu_encode_smoke` does not reach it: that drives the swscale/CPU-upload path, which uses `VaapiHw` and never builds a graph. `dmabuf_inner_alloc_drop_cycles` loops construct/drop, which is the whole contract now that every handle releases itself. **It passes on a Radeon 780M**, as do the two pre-existing VAAPI tests, so `VaapiHw` and `DmabufInner` are both hardware-verified rather than compile-only. Also fixes three more owner-only fields the AMD box surfaced: `graph`, `vaapi_device` and `drm_device` are `never read` since the hand-written `Drop` that used to read them is gone. Same call as the decoders and the QSV pair — annotate rather than delete (removing `graph` would free it while `src`/`sink` still point into it) or underscore-rename (which hides what they hold). `drm_frames` is NOT annotated: `submit` genuinely reads it per frame. I had missed these twice: my Linux log greps searched "never constructed|never used", which does not match "never read". The check now includes all three spellings. Two environment notes worth keeping, both diagnosed by testing OUTSIDE our code first (ffmpeg's own CLI reproduced each): * Inside a distrobox on an immutable host, VAAPI needs `LIBVA_DRIVERS_PATH=/run/host/usr/lib64/dri`. The container's mesa (25.3.6) is older than the host's (26.0.4) and every encoder open fails with a bare ENOSYS — "Function not implemented", naming nothing. The test doc says so. * `cargo check --workspace` in that container fails on `glib-sys` (no GTK dev headers). Unrelated to this branch; .21 checks the full workspace clean. Verified: AMD .116 (Radeon 780M, Mesa 26.0.4) full pf-encode suite 33 passed / 0 failed plus all 3 ignored VAAPI hardware tests green. Linux .21 workspace check exit 0 / zero errors, pf-encode 33/0, pf-client-core 34/0, and zero dead-code warnings across all three spellings.	2026-07-28 21:31:42 +02:00
enricobuehler	4241bc0383	fix(encode/windows): enable D3D11 multithread protection before libav sees the device ... libav turns on `ID3D11Multithread::SetMultithreadProtected` in `d3d11va_device_create` — the path where it creates the device itself. We take the other path, `d3d11va_device_init`, because we hand it the capturer's existing `ID3D11Device`, and that path does not. Nothing enabled it on our side either, so it was simply off. Two consequences, one of them shipping since long before this branch: QSV zero-copy could not open at all. `av_hwdevice_ctx_create_derived(QSV <- D3D11VA)` ends in `MFXVideoCORE_SetHandle`, and MFX rejects a device without multithread protection as `MFX_ERR_UNDEFINED_BEHAVIOR (-16)`. libav reports that as "Error setting child device handle", which names neither the cause nor the cure. AMF — the DEFAULT Windows zero-copy path — has been running with a lock that does nothing. We deliberately leave `lock`/`unlock` null so libav installs its `d3d11va_default_lock`, and that lock is `ID3D11Multithread::Enter`/`Leave`, which are documented no-ops while protection is off. The lock libav installs to serialise our capture thread against its encode thread has therefore never actually serialised anything. That is the more serious half of this fix, and it is not QSV-specific. Measured on the Intel VM (UHD 750, FFmpeg 7.1.5 + libvpl), sweeping device flags x adapter x protection: BGRA -> QSV derive FAILED (-16) BGRA +MT -> QSV derive OK BGRA|VIDEO +MT -> QSV derive OK `D3D11_CREATE_DEVICE_VIDEO_SUPPORT` makes no difference in either direction, and the same result holds on both Intel adapters the box enumerates. Protection is the only variable that matters, and it read back `was=false` every time, confirming nothing else had enabled it. The hypothesis came from ffmpeg's own CLI succeeding at the identical derive (`-init_hw_device d3d11va=d3d -init_hw_device qsv=q@d3d`) — the only difference being who created the device. With this, `zerocopy_qsv_alloc_drop_cycles` passes on real Intel silicon, which also makes the QSV arm of the AvBuffer refactor hardware-verified rather than compile-only. Its doc comment now points back here, since a future regression will present as that same opaque "child device handle" line. NOT flipped: `zerocopy_active` still defaults QSV off. The derive works and the handles construct and release cleanly, but that is not the same as a validated streaming session, and the default should move on glass evidence, not on this. Verified on .47 (Intel UHD 750): full pf-encode --features amf-qsv suite 39 passed / 0 failed, d3d11hw_alloc_drop_cycles green, zerocopy_qsv_alloc_drop_cycles green, and the native VPL path (qsv_encode_live_smoke) still green. Linux .21: workspace check exit 0 / zero errors, pf-encode 33/0, pf-client-core 34/0.	2026-07-28 21:31:42 +02:00
enricobuehler	2d78d2a514	test(encode): D3d11Hw verified on real silicon; the QSV derive is broken upstream of us ... Ran the QSV probe on the Intel VM (UHD 750, FFmpeg 8). It fails — and not in the code it was written to cover: [AVHWDeviceContext] Error setting child device handle: -16 ZeroCopyInner::open(QSV) failed on iteration 0: derive QSV device from D3D11VA That is `MFX_ERR_UNDEFINED_BEHAVIOR` out of `MFXVideoCORE_SetHandle`, raised inside `av_hwdevice_ctx_create_derived` — libav's own code, called with the same arguments as before this branch. Confirmed pre-existing by A/B: the same probe, written against unmodified `origin/main` with the raw-pointer struct, fails byte-identically on the same box at the same first iteration. The ownership change is not implicated. It is also not the hardware. `qsv::tests::qsv_encode_live_smoke` — the native VPL backend, untouched by this branch — encodes real H.264 on that box via VPL 2.15 with its own D3D11 zero-copy. So QSV works; libav's QSV-from-D3D11VA derive is what does not. That matches `zerocopy_active` already defaulting QSV **off** pending Intel validation, and is now a measured fact rather than a suspicion. The test stays, `#[ignore]`d, with the finding in its doc comment so the next Intel driver / FFmpeg bump re-checks it instead of the question being quietly dropped. What that leaves is real coverage of the half that IS reachable anywhere: `d3d11hw_alloc_drop_cycles` loops construct/drop on `D3d11Hw`, the hwdevice + frames-pool pair both Windows zero-copy vendors share, so it covers the AMF path's ownership without needing AMD hardware. **It passes on the Intel box: 8 cycles, no abort.** Adapter selection is factored into `test_hw_device`, which prefers a vendor and skips the Microsoft Basic Render Driver — a punktfunk host enumerates our own virtual-display adapter too, so `EnumAdapters1(0)` is not a safe assumption. Verified: QSV VM .47 full `pf-encode --features amf-qsv` suite 39 passed / 0 failed (EXITCODE=0), with `d3d11hw_alloc_drop_cycles` green on real silicon. Linux .21 workspace check exit 0 / zero errors, pf-encode 33/0, pf-client-core 34/0.	2026-07-28 21:31:41 +02:00
enricobuehler	564797a12a	test(encode): a QSV construct/drop smoke test, and AvFilterGraph goes Linux-only ... Two loose ends ahead of running this on real Intel silicon. `zerocopy_qsv_alloc_drop_cycles` covers the one ownership question in the crate that was genuinely ambiguous. `ZeroCopyInner::open` builds a `D3d11Hw` and then DERIVES a QSV device + frames ctx from it, and the tuple it used to return handed those two derived pointers out twice — once as the encoder's args, once as the pair moved into `Self`. Free for raw pointers, two owners for `AvBuffer`. Looping construct/drop is what separates the outcomes: a double-unref aborts in the CRT, a missed one leaks an Intel device per session. Nothing else reaches this code — `zerocopy_enabled` defaults QSV OFF, and the native VPL backend supersedes this whole file unless `PUNKTFUNK_QSV_FFMPEG=1` — so the test calls `open` directly and sidesteps both gates, the same shape as `cuda_hw_alloc_drop_cycles`. `AvFilterGraph` is now `#[cfg(target_os = "linux")]`. The Windows gate had been reporting `struct AvFilterGraph is never constructed` since a960dff8 and I had been filtering it out of my own log greps (searching for "never read", which does not match "never constructed"). It is true: the VAAPI dmabuf path is the only filter-graph user, and the AMF/QSV backends build no graph. Cfg'd out rather than `allow`ed, so it cannot outlive its last caller unnoticed. Verified. Windows .133: `cargo test -p pf-encode --features amf-qsv --no-run` at EXITCODE=0 with the AvFilterGraph warnings gone (107 -> 105), and the test binary lists `ffmpeg_win::tests::zerocopy_qsv_alloc_drop_cycles`. Linux .21: workspace check exit 0 / zero errors, pf-encode 33 passed / 0 failed, pf-client-core 34 passed / 0 failed, and zero dead-code warnings there (AvFilterGraph is live on Linux). NOTE for running it: the test binary needs `C:\Users\Public\ffmpeg\bin` on PATH at RUNTIME. `FFMPEG_DIR` is build-time only, and without the DLLs the harness exits silently with no output at all — which reads exactly like a test that does not exist.	2026-07-28 21:31:41 +02:00
enricobuehler	e60fee7da6	refactor(encode): the QSV derived pair stops being two owners of the same pointer ... `ZeroCopyInner` held the QSV device + frames ctx as nullable raw pointers, where null meant "AMF, which feeds D3D11 frames directly" — a convention documented in a comment and enforced by three `is_null()` checks. `Option<AvBuffer>` says it in the type, so the checks and the hand-written `Drop` both go away. The reason this one was left out of the previous commit is the aliasing. `open` built a five-element tuple whose QSV arm handed the SAME two pointers out twice: once as the encoder's `dev_ref`/`frames_ref`, once as the pair moved into `Self`. For raw pointers that is free; for an owning type it is two owners and a double-unref. Ownership and borrowing are now separated — the pair is owned in `qsv_frames`/`qsv_device`, and the encoder's arguments are `as_ptr()` views taken from whichever owner applies. `open_win_encoder` takes its own refs of what it is handed, so lending transfers nothing. Net: all 13 `av_buffer_unref` calls in this file are gone (the last two hand-written Drops with them), including the encoder-open failure arm, which collapses to `?` now that every handle releases itself. Field order pinned and commented, as with the others: QSV frames, QSV device, then `enc`, then `hw` — reproducing the old `Drop`, which ran ahead of all fields and so released the derived pair before the encoder's AddRef'd copies and the D3D11 refs. `qsv_device` picked up an `#[allow(dead_code)]`: `qsv_frames` is still read by the send path, but the device is now purely an owner (the frames ctx and the encoder each hold their own ref). Same call as the decoders — deleting it would free the device early, an underscore name would hide it. Verified on BOTH: Windows .133 `cargo check -p pf-encode --features amf-qsv --all-targets` EXITCODE=0, no dead-code warnings, and confirmed non-vacuous (log shows `Removed 23 files, 36.2MiB` then a real `Checking pf-encode`). Linux .21 `cargo check --workspace --all-targets` exit 0 / zero errors, pf-encode 33 passed / 0 failed, pf-client-core 34 passed / 0 failed, and `cuda_hw_alloc_drop_cycles` still passing on real CUDA. The QSV path itself still wants Intel silicon to exercise.	2026-07-28 21:31:41 +02:00
enricobuehler	5b142a7e85	refactor(encode): the Windows D3D11VA hwdevice owns its refs ... `D3d11Hw::new` is the third instance of the same shape as `CudaHw` and `VaapiHw` — alloc a hwdevice, deref to fill it, init, alloc a frames ctx, deref, init — with the same hand-written unwind: three `av_buffer_unref` calls spread over the failure branches plus a `Drop` repeating the pair. It gets the same treatment, so all three libav hwdevice wrappers in the crate now share one release path. Field order carries the semantics here as in the other two: frames declared before device, so declaration-order dropping reproduces what the hand-written `Drop` did. `ZeroCopyInner`'s `qsv_device`/`qsv_frames` are deliberately NOT converted in this commit. They are nullable (null for AMF, which feeds D3D11 frames directly), and the QSV branch hands the same two pointers out twice — once as the encoder's `dev_ref`/`frames_ref` and once as the owned pair moved into the struct. Modelling that needs `Option<AvBuffer>` plus a borrow/own split so the aliasing does not become two owners, which is a different change from the mechanical one this commit makes. Verified on the Windows runner .133: `cargo check -p pf-encode --features amf-qsv --all-targets` at EXITCODE=0. Confirmed non-vacuous — the log shows `cargo clean -p pf-encode` removing 23 files / 36.2 MiB followed by a real `Checking pf-encode`, because a fast green on that box can otherwise just be cargo reusing artifacts whose timestamps outrank the freshly-extracted sources.	2026-07-28 21:31:41 +02:00
enricobuehler	06a406adf2	refactor(client): the D3D11VA decoder owns its refs too, and the owner-only fields say so ... Completes the decoder trio. `D3d11vaDecoder::new` had the same cascading unwind as the VAAPI and Vulkan constructors — four branches each unref'ing the hwdevice by hand — and `d3d11va_decode_supported` opened a frames context it released on one exit path while relying on the null check for the other. Both are `AvBuffer` now. The one surviving `av_buffer_unref` in the file is deliberate: it runs before ownership is taken, on the `av_hwdevice_ctx_init` failure ahead of `from_raw`. The Windows gate then caught something worth keeping: `field hw_device is never read`, in all three decoders. It was true and it was new — the hand-written `Drop` used to be the field's only reader, so moving the unref into the type left a field that is *purely* an owner. The compiler cannot express that, so each one now carries a doc comment saying what it is and an explicit `#[allow(dead_code)]`. Deleting the field would free the device early; renaming it `_hw_device` would silence the lint by hiding what it holds. Neither is what we mean. Verified on BOTH platforms this time. Windows runner .133 (the toolchain env from the runner notes, plus `PF_FFVK_VULKAN_INCLUDE` for pf-ffvk): `cargo check -p pf-client-core --all-targets` at EXITCODE=0, clean of the dead-code warnings. Linux .21: `cargo check --workspace --all-targets` exit 0 / zero errors, pf-client-core 34 passed / 0 failed, pf-encode 33 passed / 0 failed.	2026-07-28 21:31:41 +02:00
enricobuehler	a42ab075a8	refactor(client): the hardware decoders own their hwdevice ref ... `VaapiDecoder::new` and `VulkanDecoder::new` create a hwdevice and then do several more fallible things with it — resolve `vkWaitSemaphores`, find the decoder, alloc the context, open it — and every one of those branches unref'd the device by hand, with a `Drop` doing it once more. Six hand-written unrefs between them, each one a line somebody has to remember when adding a step. `video_libav::AvBuffer` owns it instead, so an early `bail!` releases it and the branches carry nothing. It is a deliberate second copy of `pf-encode`'s type: the crates do not depend on each other (host encode and client decode share no code path), and this one needs something the host's does not — see below — so hoisting it to a shared crate would mean giving that crate an ffmpeg dependency and both sets of semantics to save about twenty lines. That extra piece is `into_raw`. `pick_vulkan` hands its frames context to the codec (`(*ctx).hw_frames_ctx = fr` — the codec unrefs it when the context closes), so the wrapper must give up ownership rather than drop. Making the transfer explicit is the point: dropping an `AvBuffer` there too would be exactly the double-unref this type exists to prevent. Two `av_buffer_unref` calls survive in `video_vulkan.rs` on purpose. One runs before ownership is taken (the `av_hwdevice_ctx_init` failure, ahead of `from_raw`); the other releases the codec's OWN pre-existing frames ctx before we replace it, which was never ours to model. Field order preserved: `hw_device` stays declared after `ctx`, so it still releases after each `Drop` frees packet/frame/context — the order the hand-written unref had. Verified on .21 (CachyOS, FFmpeg 62): `cargo check --workspace --all-targets` clean at exit 0 with zero errors — the workspace-wide check owed since the AvBuffer commit — plus pf-client-core 34 passed / 0 failed and pf-encode 33 passed / 0 failed. The decoders themselves still need real VAAPI/Vulkan playback to exercise.	2026-07-28 21:31:41 +02:00
enricobuehler	eb9c5be20d	refactor(encode): the VAAPI dmabuf path stops unwinding by hand ... `DmabufInner::open` builds four owned objects — a DRM device, a VAAPI device derived from it, a DRM-PRIME frames context, and a filter graph — and every one of its eight failure branches unwound them by hand. The same four-line block (`avfilter_graph_free` + three `av_buffer_unref`s) appeared eight times, once *inside a macro*, plus a ninth copy in `Drop`. Adding a step to that function meant remembering to extend the unwind at exactly the right depth; getting it wrong leaks a device per failed session (the persistent listener accumulates them) or frees one twice. All eight are gone. `AvBuffer` already owned the buffer refs; `AvFilterGraph` now does the same for the graph, so each handle is owned the moment it exists and an early `bail!` releases whatever was built so far. `open` lost ~40 lines of cleanup and gained none. Field order in `DmabufInner` is load-bearing and says so: graph, frames, VAAPI device, DRM device, then `enc` LAST. Fields drop in declaration order, and that sequence reproduces the old hand-written `Drop` exactly — including that it ran ahead of every field, so all four were released before ffmpeg-next dropped the encoder. Everything here holds its own reference, so refcounting makes any order sound; the ordering is pinned so a future reorder cannot quietly change what ships. One subtlety preserved deliberately: the buffersrc parameters take `drm_frames` BORROWED, not ref'd (`av_buffersrc_parameters_set` takes its own ref). That is now `drm_frames.as_ptr()` — same borrow, same single owned ref, no new leak. Verified on .21 (CachyOS, RTX 5070 Ti, FFmpeg 62): `cargo check -p pf-encode --all-targets` clean at exit 0, `cargo test -p pf-encode` 33 passed / 0 failed, and `cuda_hw_alloc_drop_cycles` still passes against real CUDA. vaapi.rs now contains zero `av_buffer_unref` and zero `avfilter_graph_free` calls, down from 40 and 8. The dmabuf path itself still needs AMD/Intel silicon to exercise end to end.	2026-07-28 21:31:41 +02:00
enricobuehler	7adc1db672	refactor(encode): AVBufferRef ownership moves into an RAII handle ... `CudaHw::new` and `VaapiHw::new` are the same shape — alloc a hwdevice, deref it to fill fields, init, alloc a frames ctx, deref, init — and both unwound by hand: `av_buffer_unref` on every failure branch (three in CUDA, two in VAAPI) plus a hand-written `Drop` repeating the pair. That shape has two failure modes and the compiler can see neither: add a branch and forget the cleanup (leak), or let two cleanup paths run (double-unref, an abort inside glibc). `libav::AvBuffer` owns the ref instead. It null-checks on the way in — the check each caller open-coded — and unrefs exactly once on drop, so an early `?` releases whatever was built so far and the failure branches carry no cleanup at all. Both hand-written `Drop` impls are gone, and `linux/mod.rs` goes from seven `av_buffer_unref` calls to zero. The one subtlety, called out at both structs: these two fields must stay declared frames-BEFORE-device. Fields drop in declaration order, and the code being replaced deliberately unref'd frames first (a frames ctx holds its own reference on its device). Refcounting makes either order sound, but a field reorder should not silently change what ships, so the ordering is load-bearing and commented as such. Also adds `cuda_hw_alloc_drop_cycles` — the RAII path had NO test coverage: the NVENC smoke tests take the CPU path and never construct a `CudaHw`, and the VAAPI twin's tests need AMD/Intel silicon. Looping construct/drop is what catches the double-unref (abort) and the leak (allocator growth) this refactor is about. Verified on Nobara (RTX 5070 Ti): `cargo check -p pf-encode --all-targets` clean at exit 0, `cargo test -p pf-encode` 33 passed / 0 failed, and `cuda_hw_alloc_drop_cycles` passes against a real CUDA device — eight construct/drop cycles, no abort. `VaapiHw` is compile-verified only; it is the identical shape but no AMD/Intel box was reachable to run its two ignored tests.	2026-07-28 21:31:41 +02:00
enricobuehler	5219107177	chore(unsafe): the workspace adopts the drivers' unsafe discipline ... `packaging/windows/drivers/*` has run `deny(unsafe_op_in_unsafe_fn)` + `deny(clippy::undocumented_unsafe_blocks)` for a while, with `forbid(unsafe_code)` on the modules that need no unsafe at all. The main workspace had no lint config whatsoever, so nothing stopped a clean crate from quietly growing an `unsafe`, and nothing distinguished the handful of genuinely-unsafe lines inside a 600-line `unsafe fn` from the safe ones surrounding them. Three things, all mechanical: * `#![forbid(unsafe_code)]` on the eight crates that already contain zero unsafe (`pf-driver-proto`, `pf-host-config`, `pf-paths`, the three clean clients, both tools). These were clean by accident, not by contract; now they are clean by contract. * `unsafe_op_in_unsafe_fn = "warn"` workspace-wide. `unsafe fn` states a contract the CALLER must uphold — it was never meant to switch off checking for the whole body. Measured fallout is 300 sites on Linux, and they are concentrated: six files carry all of them, while `punktfunk-core`, `pf-frame`, `pf-clipboard` and `pf-vdisplay` are already at zero. `warn` (not `deny`) so the build stays green while those six are worked down; it flips to `deny` once they are. This is also the Rust 2024 default, so it pays off the edition migration early. * `proc::current_uid()` replaces eight `unsafe { libc::getuid() }` blocks. Each site had copied out the same SAFETY note verbatim, which is the tell: `getuid()` is parameterless, always succeeds and touches no memory, so there is no contract for a caller to uphold and no reason for the unsafe to be visible eight times. One `unsafe` behind a safe wrapper, none at the call sites. Verified: `pf-vdisplay` builds clean on Linux (Nobara) at zero E0133; the macOS-buildable crates build clean locally. No behaviour change.	2026-07-28 21:31:41 +02:00
enricobuehlerandClaude Opus 5	eab4829630	feat(client): the brain layer — one connect plan, one wake machine, one spawn ... C0 of design/client-architecture-split.md. Wake-then-connect exists three times today — GTK's `WakeConnect`/`wake_fallback`, the WinUI shell's `wake_and_connect` (whose comment says it "mirrors the Apple HostWaker"), and Apple's `HostWaker` itself — and the deep-link and profile work was about to make it five. `pf-client-core` already held the ingredients (wol, discovery, trust, library, session); what was missing was the layer that composes them, so every front-end composed them itself. `ConnectPlan` is a resolved intent: which host, which pin, which launch id, which profile, the effective settings, whether to wake. One constructor per door — a card click, a CLI verb, a URL — one type out. `ConnectPlan::resolve` is pure (stores in, plan out), which is what lets the URL router be tested without a config directory; `for_host` is the convenience that loads them. `plan_from_link` is where the deep-link security rules actually live, rather than in each shell: a contradicted fingerprint refuses and says which host it was about, an ambiguous name refuses instead of picking the first, a profile the catalog can't honor refuses BEFORE anything is dialled, and an unknown — or known-but-never-pinned — host becomes a confirmation sheet carrying the claimed name and expected pin, so the first connect is verified rather than blind TOFU. Preempting a live session stays with the caller: only the front-end knows one is running. `WakeWait` is Apple's `HostWaker` cadence as a pure step function — packet at 0 s and every 6 s, presence polled every second, 90 s budget, and a PARK (not an error) at the end, because "it didn't wake in 90 s" is usually "give it 10 more". As a step function each front-end drives it from its own loop and they still agree, and the whole cadence is testable without waiting 90 seconds. Session spawn, its argv and its stdout contract move here too, and the GTK shell now spawns through them — so "which flags does a stream get" and "what does ready mean" have one answer for the shells, the console and the coming CLI. `--profile` deliberately does NOT ride a plain card click: the session resolves the host's binding through the same helper, and passing it would be a second source of truth for one decision. Not yet adopted: the two shells' wake paths. That swap changes the most-used path in the product and the handoff wants it verified on glass with a genuinely sleeping host, which this session couldn't do — so the state machine lands, the duplication stays until it can be verified away. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-07-28 20:45:07 +02:00
enricobuehlerandClaude Opus 5	3350d3cfbf	feat(client): the punktfunk:// grammar — one parser, one emitter, one vector file ... D0 of design/client-deep-links.md. Apple has shipped `punktfunk://connect/<uuid>` for a while; this is the same grammar, generalised and given a Rust implementation the GTK and WinUI shells, the session and the coming CLI all share: punktfunk://connect/<host-ref>[?fp=…][&host=addr[:port]][&launch=…][&profile=…][&name=…] The rule the grammar exists to keep is that a URL may only ever do what a click on an existing card could do, minus trust decisions. So it carries REFERENCES to things that already exist on this device — never values: no resolution, no bitrate, no codec, so a web page cannot shape a session beyond picking among the user's own configurations. And `pair` is not a route: it parses, and refuses, so a link can never start a trust ceremony. Everything hostile is rejected here, once, for every front-end — 2 KB total cap, per-parameter caps, strict percent-decoding (a half-escape or invalid UTF-8 is an error, not a shrug), control characters refused after decoding, launch ids held to the charset the host and Decky already agree on, and `fp=` that must actually be 64 hex. Resolution follows the same order everywhere: stable record id → unique host name → `addr[:port]`, with `host=`+`fp=` as the reinstall recovery path, which is why self-emitted links carry all three. An ambiguous name refuses rather than guessing; a stale id with no recovery address refuses rather than dialling "11111111-…" as if it were a hostname; an unknown-but-plausible address becomes the confirmation sheet, never an auto-connect. `pf://` parses as an input alias and is never emitted or registered — claiming a two-letter scheme on MSIX/Apple is unconditional squatting, and a link that resolves on one platform only is a trap. The 44-case `clients/shared/deeplink-vectors.json` is the cross-language contract: this module's tests run every case, and the Swift and Kotlin ports will run the same file, so three parsers cannot drift into three different security postures. Refusal codes are part of that contract, not just the happy path. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-07-28 20:32:32 +02:00
enricobuehlerandClaude Opus 5	a2be3f4de9	feat(client): settings profiles — the override catalog, the one resolver, and the session flag ... A client setting is global today: pick 4K@120 HDR and the retro box gets it too. The one escape hatch (per-host `clipboard_sync`) proves the shape works but covers a single field. This is P0 of design/client-settings-profiles.md — the headless core the shells, Apple, Android and the deep-link grammar all build against. No UI yet. A profile is a NAMED BUNDLE OF OVERRIDES, not a snapshot: `SettingsOverlay` is sparse `Option`s, so a field the user never touched keeps following the global value live, and fixing a global once fixes it everywhere. `Some(x)` equal to today's global is still meaningful — it pins x against a later global change — which is why the UI will write `Some` on touch and `None` on an explicit reset, never by diffing. The catalog is its own `client-profiles.json`, deliberately not part of the settings file: that file has five whole-file load-modify-save writers with no merge, so a profile written by one would be dropped by the next. Which host uses which profile is a field on the host record instead of a map keyed by host — that is what dissolves the client's long-standing "what identifies a host record" question for this feature. Resolution has exactly one implementation, `trust::effective_settings()`: effective = overlay(profile).apply(global) profile = one-off pick ?? host binding ?? none Both session construction sites go through it (`main.rs` AND `console.rs` — touching one and not the other is a Windows-only build break), so the console, and therefore Decky, honor host bindings with no work of their own. Nothing here can fail a connect: a deleted binding resolves as "no profile", and a one-off that can't be honored falls back to the DEFAULTS rather than to the host's own profile — "connect with Work" must never quietly stream "Game". `--profile <id|name>` is the one-off door for the shells' coming "Connect with ▸" and for `punktfunk://…&profile=`; `--profile ""` forces the defaults on a bound host. The active profile's name closes the stats overlay's first line, so "which profile am I on?" is answerable without leaving the stream. Also here, because this effort touches every one of these anyway: - `KnownHost` gains `profile_id`, `pinned_profiles` and a lazily minted stable `id`. The id is groundwork (design §4.5) — no lookup is re-keyed, `fp_hex`/`addr:port` stay. - `upsert` now preserves the state the USER set on a record — the profile binding, its pins, the clipboard decision, the id — against refreshes that carry none of it. `clipboard_sync` survived that only because the function happened not to mention it. - `KnownHost::default()` exists so construction sites read `..Default::default()`; five hand-written literals were exactly how a new field gets silently dropped on re-pair. - All three client stores now write temp+rename. A torn settings file loads as `Default` — i.e. silently resets every setting the user has. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-07-28 20:22:33 +02:00
enricobuehlerandClaude Opus 5	33a31427ae	fix(vdisplay): locks held across seconds-long work, an AB/BA inversion, and a panic that poisons the manager ... Phase 7's contained half. 7.1's writer split and 7.3's reservation are NOT here — see the plan; both need more than a local edit. **Lock-order inversion, gamescope.** `create_managed_session_steamos` held MANAGED_SESSION while taking STEAMOS_TOOK_OVER (and `persist_takeover`'s locks), while `do_restore_tv_session` takes STEAMOS_TOOK_OVER first and MANAGED_SESSION second — AB/BA between a connect and the restore worker, which genuinely run concurrently (`vd.create` runs off the registry lock; the worker is its own thread). The connect path now drops its guard before the pair and re-acquires after `poll_managed_node`, the shape `create_managed_session` already used. And `takeover_live` held all four statics at once — in a `||` chain every `.lock()` temporary lives to the end of the statement — for four reads; scoped bindings drop each guard before the next, taking it out of the ordering graph entirely. **`observe_session_instance` decided and acted under one lock.** It held LAST_INSTANCE across `invalidate_backend` (which drops keep-alive displays through the registry lock) and a `systemctl` shell-out on a 10 s budget, from three call sites including a per-second watcher. It decides under the lock and acts outside now; the baseline is advanced inside it, so a concurrent observer cannot run the action twice. **`admit` held the live-session table across the budget checks** — `manager::snapshot()` blocks on the manager `state` lock, which is itself held across DDC round trips and three 3 s activation ladders. Every connect, disconnect and mgmt read queued behind one slow display operation. The guard is scoped to `decide` alone. **GameStream never registered**, so its display was invisible to both Windows budgets (they gate on `!live.is_empty()`) and a native connect could be admitted past capacity the box had already spent. It registers now, anonymously, for the session's lifetime. **`ensure_exclusive_watch` panicked on a failed thread spawn** while holding BOTH `exclusive_watch` and (via its caller) `state` — poisoning the two locks the whole manager runs on, so every later acquire and mgmt read would panic on `.lock().unwrap()`. It logs and degrades instead: no re-assert watchdog is a degradation, a wedged manager is not. Also fixes two things only a Windows build shows, both mine: Phase 2.3 left `gamescope_route` unused there (every reader is Linux-gated), which `-D warnings` in Windows CI would have rejected — I had only run the host clippy on Linux. And Phase 5.4's helper-safening reaches a FOURTH crate, pf-inject, which xcheck does not lint. Verified on .173: `cargo clippy -p punktfunk-host -p pf-vdisplay --all-targets` clean across the whole tree; local fmt, both xcheck targets and 53 tests green. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-07-28 19:46:15 +02:00
enricobuehlerandClaude Opus 5	093c66cb1a	fix(vdisplay/windows): a failed resize left a phantom monitor, and a refresh-only one silently didn't happen ... Three defects in the mid-stream resize path. **A failed re-arrival put a DEAD monitor back in the slot.** `re_add` REMOVEs the old driver monitor before it ADDs the new one, so its only fallible step past that point means the old monitor is gone — yet the caller answered an `Err` by re-inserting the `Monitor` struct it had handed in, complete with a dead `key`, a departed `target_id` and a stale `gdi_name`. Every later `acquire` then joined a monitor that did not exist. It now rolls back: on ADD failure it re-ADDs at the OLD mode, so the session keeps streaming what it already had. The outcome is a three-way `ReAdd` enum because the distinction matters and a `Result` cannot carry it — `Arrived` stores the new monitor, `RolledBack` stores the RECOVERED one (a re-arrival mints a fresh `target_id`, so the struct the caller handed in is stale either way), and `Lost` means both ADDs failed and the slot must be left EMPTY so the next acquire creates a monitor rather than joining a phantom. **A refresh-only change reported success at the old refresh.** `wait_mode_advertised` compared resolution only, so a same-WxH new-Hz request answered "already advertised": the fast path skipped the `IOCTL_UPDATE_MODES` that would have taught the driver the new rate, `set_active_mode` fell back to the best advertised rate <= requested, and the resize returned Ok. It matches `dmDisplayFrequency` now, so an unadvertised refresh routes to UPDATE_MODES / the re-arrival instead. And `mon.mode = mode` recorded what was ASKED FOR. `set_active_mode` deliberately falls back to a lower advertised refresh rather than lose the client's resolution, so the field could claim a rate the display was not running — and it is what the next resize diffs against and what `/display/state` reports. It records what actually committed now, read back via the new `active_mode` (`active_resolution` delegates to it), and logs when the two differ. Deliberately NOT done by tightening `wait_mode_settled`: that gates the re-arrival fallback, so requiring an exact refresh there would force a full re-arrival every time the OS legitimately picked a lower rate. **A short IOCTL_ADD reply leaked an IddCx slot.** The IOCTL had SUCCEEDED — the driver already created the monitor and took a slot; only the reply was short — and the bail undid none of it. The slot pool is small enough that ~16 leaks wedge every later ADD at 0x80070490, which is the wedge the ghost-reap in this same function exists to recover from. It now issues the compensating REMOVE with the session id already in scope, and says so at error level if that fails too. Windows runner: clippy --all-targets clean, pf-vdisplay 55 passed, pf-capture 18 passed. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-07-28 19:46:15 +02:00
enricobuehlerandClaude Opus 5	50a153b72a	refactor(win-display): the CCD helpers were unsafe fn for serialization, not soundness ... `pf-win-display`'s 21 CCD/GDI helpers were `pub unsafe fn` without a memory-safety obligation between them. Every one takes `Copy` scalars or borrowed Rust data, returns owned values, and discharges its FFI preconditions internally — `retry_set_display_config` even binds the input desktop itself, which is the one thing a caller could plausibly get wrong. Their own `# Safety` sections say as much: one reads "safe to call from any thread", and the rest say "call under the manager `state` lock". That is a SERIALIZATION requirement — calling one unlocked races the topology mutator and returns a stale answer, a correctness bug, not undefined behaviour. Encoding it with `unsafe fn` cost 55 `unsafe` blocks across THREE crates whose proofs could only restate "this takes a u32 and returns a String". That is how `unsafe` stops meaning anything: the blocks that wrap a real FFI obligation read exactly like the ones that wrap nothing. The requirement is now prose on the functions that have it, and `unsafe` marks only the FFI calls inside the helpers. win_display.rs 97 -> 63 `unsafe`, manager.rs 56 -> 30, pf_vdisplay.rs 34 -> 32, plus 10 sites in pf-capture's IDD-push paths. 55 blocks and 55 orphaned `// SAFETY:` proofs gone; `#![allow(clippy::missing_safety_doc)]` with them, since nothing is `pub unsafe fn` any more. Two things the compiler handed back once the blocks went, both of which existed only to carry a proof: the `let-else` parens in `force_mode_reenumeration` and `apply_source_positions`, and the nested `if` in `wait_mode_settled` — which is `&&` again, still short-circuiting, so the second CCD query on a 25 ms poll is unchanged. Scope note: the sweep and the plan both scoped this as two crates. It is three — pf-capture calls the same helpers from 10 sites, and `unused_unsafe` under `-D warnings` makes that non-optional. Found by re-deriving the count, which the plan asked for because the underlying finding had been REFUTED in verification; the refutation was wrong about the magnitude (22 of manager.rs's 45 blocks, against its claimed 24). Verified on the Windows runner: clippy --all-targets clean over pf-frame, pf-win-display, pf-capture and pf-vdisplay; pf-vdisplay 55 passed, pf-capture 18 passed. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-07-28 19:46:15 +02:00
enricobuehlerandClaude Opus 5	029979e824	fix(vdisplay/windows): own the device handle, prove the ladder once, and fix two proofs that were wrong ... Three of Phase 5's four items; 5.4 is held for a decision (see below). `open_device` is a SAFE fn returning an `OwnedHandle`. It never had a caller obligation — it takes no arguments and every precondition is internal — so the `unsafe fn` it used to be pushed a proof burden onto four call sites with nothing to prove, and handed back a raw `HANDLE` that two of them re-owned by hand with `CloseHandle`. That shape has already leaked once in this file (the wrap-IMMEDIATELY comment in `open` records it). Ownership is now a `Drop`, `probe()` is `open_device().map(|_| ())`, `is_available()` is `open_device().is_ok()`, and `CloseHandle` is no longer imported here at all. `resolve_target_gdi` ran the same 60 × 50 ms poll three times verbatim, and the 2nd and 3rd copies documented themselves as "SAFETY: as the resolve loop above" — a pointer to a proof rather than a proof, which rots silently the moment the block it points at moves. One `poll_gdi_name`, one proof, and the three-stage ladder now reads as the three stages its doc describes. Two proofs were simply wrong, both discharged everything except the one obligation that mattered: * The SetupAPI detail buffer was a `Vec<u8>` (align 1) written through as `SP_DEVICE_INTERFACE_DETAIL_DATA_W` (align 4). The SAFETY comment proved bounds and aliasing and was silent on alignment. Now a `Vec<u64>`. Its guard also compared the required size against `size_of::<u32>()` while stamping `size_of::<SP_DEVICE_INTERFACE_DETAIL_DATA_W>()` into `cbSize` — checking a different number than the one it promises the OS. * `GetMonitorInfoW` was handed `&mut info.monitorInfo`: `cbSize` promises 104 bytes but that pointer's provenance covers the 40-byte `MONITORINFO` prefix, so everything the OS writes past byte 40 — which is `szDevice`, the only field the caller reads — is out of bounds for it. A compiler may assume those bytes were untouched and fold the zeroed initializer into the reads, i.e. return an EMPTY device name. That name is what `panel_off_except`'s exclusion compares against, so an empty one darkens the panel it is meant to spare. The pointer now carries the full struct's provenance. pf_vdisplay.rs 42 -> 34 `unsafe`, manager.rs 58 -> 56. Windows runner: clippy --all-targets clean, 48 passed. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-07-28 19:46:15 +02:00
enricobuehlerandClaude Opus 5	f94191d927	fix(vdisplay/linux): stop destroying the user's portal config, and honour the /tmp promise the docs made ... Both wlr-family backends need ONE key set in a config file the user also owns — `chooser_cmd` in xdpw's `[screencast]`, `custom_picker_binary` in xdph's `screencopy { … }`. Both did it by `fs::write`-ing a COMPLETE file over whatever was there, so every other setting the user had in `~/.config/xdg-desktop-portal-wlr/config` or `~/.config/hypr/xdph.conf` was destroyed on first connect — silently, permanently, no backup. They now set that one key in place and keep every other line byte-for-byte, with a one-time `create_new` backup the first time we touch a file we did not write (once, so a later edit cannot overwrite the user's ORIGINAL with our own earlier output). The merge lives in its own module because a merge is only worth doing if it cannot corrupt what it merges into: seven tests cover the three cases per grammar — block absent, key present, key absent — plus idempotence (which is what keeps the caller from restarting the portal on every connect) and the empty-file shape. Four of them FAIL against the old whole-file write, so they are not vacuous. The module is declared unconditionally though only Linux calls it: the logic is pure string handling, and tests that guard a user's real config should run on every platform's CI, not only where the callers compile. Separately, seven call sites resolved the runtime dir as `env::var("XDG_RUNTIME_DIR").unwrap_or_else(|_| "/tmp".into())` while their own doc comments promised the opposite — "per-user, 0700 — NOT a world-writable /tmp path another local user could pre-create or rewrite between our write and xdph's read". One of those paths is an executable xdph then RUNS. `Ok("")` was the same defect's other half: it yielded a path relative to the process CWD rather than any runtime dir. All seven now go through one resolver that applies the rule `session.rs` already had (non-empty, else `/run/user/<uid>`, never /tmp). The picker shim also gets its mode at CREATION rather than a chmod afterwards, so it is never briefly world-readable while being a file something else executes. And identity.rs was the only writer in the workspace creating `pf_paths::config_dir()` — which holds the host private key, the pairing allow-list and the mgmt token — with a plain `create_dir_all` instead of the 0700 `create_private_dir` every other writer uses. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-07-28 19:46:15 +02:00
enricobuehlerandClaude Opus 5	2c015bba30	fix(vdisplay/windows): three teardown gates that each ended with the operator's panels dark ... All in the same ~120 lines, all the same shape: a recovery leg gated on the wrong flag, so the path that turned something off had no path that turned it back on. **The DDC wake was nested inside the CCD-restore arm.** Panels are commanded dark over DDC/CI *before* the isolate, but `panel_on_all()` sat inside `if let Some(saved) = ccd_saved.take()` — and `isolate_displays_ccd` returns `None` whenever its `query_active_config` fails. So a failed isolate meant the panels were never woken, and because the link stays live there is no returning signal to trigger a DPMS wake either: dark for the rest of the host's life. Hoisted out of that arm, next to the `pnp_disabled` restore which was already outside it for exactly this reason. **A later member's isolate could deactivate the physicals with nothing able to restore them.** If the FIRST member's isolate failed, `ccd_saved` stayed `None` and `ccd_exclusive` `false`; a second member's isolate then succeeded, blanked the physicals, and discarded its snapshot as "the group restores the first member's" — except there wasn't one. Teardown's restore is gated on `ccd_saved`, and the re-assert watchdog never started. The group now adopts the first SUCCESSFUL isolate's snapshot. **A non-last-member teardown ran the exclusive isolate on a Primary group.** The gate was `ccd_saved.is_some()`, but `Topology::Primary` stores a snapshot too (from `set_virtual_primary_ccd`) — so a shrink cleared `DISPLAYCONFIG_PATH_ACTIVE` on every non-kept path, blanking precisely the physical displays `Primary` exists to keep lit. It keys off `ccd_exclusive` now (what the re-assert watchdog already checks), and a Primary shrink re-promotes a survivor instead, since the departing member may have been the one holding primary. Also: `apply_group_layout` had only ever run from `acquire`, so a member LEAVING never re-arranged the survivors — they kept the origins computed while the departing monitor was still between them, leaving a dead gap in the desktop until the next acquire. Now called at the end of `teardown_removed`, which is where the slot map has already shrunk and after the REMOVE, so it covers all six teardown paths at once. The shrink gate is extracted as `shrink_action` so it can be pinned without a driver or a desktop — manager.rs had no tests at all, which is how a gate keyed on the wrong flag survived. Verified on the Windows runner: 48 passed (was 46), and both new tests FAIL against the old gate, so they are not vacuous. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-07-28 19:46:15 +02:00
enricobuehlerandClaude Opus 5	8b076f5cc9	fix(vdisplay): the gamescope sub-mode is a per-session value, not a process-global env write ... Completes what the env-lock fix started. `apply_input_env` published its decision into PUNKTFUNK_GAMESCOPE_NODE/_SESSION and `GamescopeDisplay::create` read it back out — with the lock released in between, and with the GameStream plane, the mid-stream switch watcher and the capture-loss rebuild all re-running the writer. So session B's routing decision could retarget session A's `create`: A asks for a bare spawn, B connects, A attaches to B's session instead. The decision now travels as a `GamescopeRoute` return value and is carried on the backend INSTANCE via `set_gamescope_route` — the discipline `set_launch_command` already documents two lines above it ("Carried on the backend instance — NOT a process-global env var — so concurrent sessions can't stomp each other's launch target"). Nothing writes the two keys any more; they are read once, as operator overrides, and `create`/`poolable_now`/`launch_is_nested` all answer from the session's own route. The route carries its payload rather than just the verdict, so the managed session flavour and the attach node id stop being separate env reads too. Threaded through both planes: handshake -> SessionContext -> `vd`, and the GameStream `open_source` -> `run`. Three re-resolution points needed the same treatment and each would have silently downgraded a session to a bare spawn without it — the switch watcher's rebuilt backend, the capture-loss rebuild (which can swap the whole instance), and the operator-pinned path, which deliberately skips the input-backend write but still needs a route. That last one is why the ladder is split out as `resolve_gamescope_route`: a `PUNKTFUNK_COMPOSITOR` pin must not silently mean "bare spawn" on a box pinned to the managed session. `#[must_use]` on both entry points found the capture-loss site, which I had missed. Verified on 192.168.1.21: whole workspace `cargo check -p punktfunk-host`, clippy --all-targets clean over the host and pf-vdisplay, 100 pf-vdisplay tests, and the live GNOME create/drop still green. Windows + Linux xcheck clean on the Mac. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-07-28 19:46:15 +02:00
enricobuehlerandClaude Opus 5	80e57c70be	fix(vdisplay): the env lock covered its writers and not its readers, and the gamescope ladder read back its own output ... Two halves of the same problem — process env used as shared mutable state. `apply_session_env` set XDG_RUNTIME_DIR / DBUS_SESSION_BUS_ADDRESS / WAYLAND_DISPLAY / HYPRLAND_INSTANCE_SIGNATURE / SWAYSOCK under ENV_LOCK, while detection read those same five keys with no lock at all, from five points spread across a `/proc` scan. That is exactly the hazard ENV_LOCK's own doc names: glibc `setenv` can realloc `environ` and free the old value string, so a concurrent `getenv` reads freed memory — UB documented there as "could crash the host". It is reachable today: the host's session watcher calls `detect_active_session` every second for a stream's lifetime while a second client's connect runs `apply_session_env` on a spawn_blocking thread. Detection now samples all five into an `EnvProbe` in ONE locked read and scans from that snapshot. Taking the lock around the whole of `detect_active_session` would have been simpler and wrong: it walks `/proc` every second, and holding a process-wide lock across a directory walk trades this problem for the one Phase 7 is about. The readers become pure functions of their inputs, which is also what let the tests stop mutating process env to steer them — `without_inherited_swaysock` is gone, replaced by handing in a probe. Empty values are now treated as absent, so `XDG_RUNTIME_DIR=""` falls back to /run/user/<uid> instead of yielding a relative path. The gamescope sub-mode ladder had a nastier version: `apply_input_env` WRITES PUNKTFUNK_GAMESCOPE_NODE/_SESSION to publish the sub-mode it chose, and read the same keys back as operator overrides. The Attach arm sets `_NODE=auto`, and `node_env` sits at rung 2 of the ladder — above `dedicated_launch` at rung 3 — so one Attach decision latched Attach for the rest of the host's life and silently overrode `game_session=dedicated`. Only rung 1 could escape it, because the Spawn arm that would have cleared the keys sits below the rung that by then always fired. The overrides are now sampled once, before this module has written anything, which keeps their actual meaning: "the operator set this before we ran". The live reads that remain (`launch_is_nested`, `poolable_now`) are deliberate — those consume the published decision. And gamescope's `create` read both keys unguarded, racing the writer and skipping the lock its two siblings take for the same keys; it now takes one guarded snapshot. Verified on Linux (.25): 100 passed, 1 ignored. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-07-28 19:45:36 +02:00
enricobuehlerandClaude Opus 5	787756e68b	test(vdisplay/mutter): the on-glass lever for the GNOME backend, and what it took to see the leak ... `live_mutter_create_drop`, `#[ignore]`d in the same convention as the Windows backend's live tests: the real four-step handshake, hold, then drop the keepalive and let the RAII teardown revert it. Verified against gnome-shell on 192.168.1.21 — node_id came back in ~5 ms, preferred_mode as asked, and the topology reverted. Two things that would otherwise be re-derived by whoever validates this next, both of which cost a wrong answer here first. The virtual monitor does NOT appear in DisplayConfig's GetCurrentState just because `create` succeeded — its size follows the PipeWire negotiation, so without a consumer attached there is no monitor to census. A before/after monitor diff therefore proves nothing about this backend, which is the trap the first attempt fell into. And a short-lived process cannot exhibit the abandoned-thread leak at all: exiting drops the D-Bus connection, which makes Mutter tear the session down no matter which code is running. Reproducing it needs the process held alive past the failed create, and the observable is the live RemoteDesktop session object, not the monitor: `gdbus introspect -d org.gnome.Mutter.RemoteDesktop -o /org/gnome/Mutter/RemoteDesktop` lists a `node Session` child per live session. With a 1 ms create timeout forced and the process held 20 s, that probe reads 0 with b89a36a6 and 1 without it — the orphan thread parked on a live session. Both read 0 after exit, which is precisely why this only ever bit a long-running host. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-07-28 19:45:11 +02:00
enricobuehlerandClaude Opus 5	69bc92afed	fix(vdisplay/mutter): a create timeout no longer abandons the session thread with the desktop dark ... `create` gave up after 45 s and dropped its `stop` flag WITHOUT setting it — the `StopGuard` that owns the only `store(true)` was built on the success path only. The thread meanwhile discarded its own send failure with `let _ =`, so it sailed past the dead channel, made the virtual output PRIMARY, and parked in `while !stop` for the rest of the process's life holding the D-Bus connection that IS the monitor's lifetime. Its teardown was unreachable. Under the default topology that orphan applies a SOLE-monitor APPLY_TEMPORARY config. Every physical head goes dark, and Mutter reverts such a config only once the virtual monitor disappears — which the parked thread prevents. There is no in-process recovery: each retry adds another orphan and another virtual monitor. The other three portal/D-Bus backends never had this — kwin.rs, hyprland.rs and wlroots.rs all `?` on that exact send. Mutter is now the same shape, via a shared `report_node` that stops the session when the opener has gone. Two more layers behind it, because the send can also LAND in the moment `recv_timeout` gives up: `create` and `stream_existing_output` build their guards BEFORE the wait, so every error arm signals the thread on its way out, and the thread re-checks the flag before it touches topology — so a doomed session never applies a sole-monitor config at all, rather than applying one and reverting it a tick later. The fix lands once because the duplication that caused it is gone first: `connect` and `connect_monitor` shared ~72 verbatim lines, and the mirror path repeating the session path's `let _ = send` is precisely what that copy bought. Both now share `open_rd_sc` (steps 1-2) and `start_and_await_node` (step 4). Also stops a physical hotplug stealing the virtual monitor's identity. `wait_virtual_connector` took "any connector absent from the pre-snapshot" across a window spanning a ~10 s stream wait plus 6 s of polling; TOPOLOGY_LOCK keeps sibling sessions out of it but a hotplug is not ours to serialise, and the session's topology apply and scale persistence would then be aimed at the operator's new panel. It now prefers the connector advertising the client's exact WxH — the discriminator `make_virtual_primary` already matches on — falls back to the old behaviour when that singles nothing out, and warns when more than one candidate appears. Tests: `pick_virtual` split out of the polling loop so the choice is testable without a session bus. Verified on Linux (192.168.1.25): 99 passed, and the hotplug case FAILS against the old first-new-connector logic, so it is not vacuous. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-07-28 19:45:11 +02:00
enricobuehlerandClaude Opus 5	e7dc7e9d18	fix(win-display): every CCD/GDI operation now carries a proof, not just the blocks that needed one ... Companion to c16f2850, which closed the same hole in pf-vdisplay. The file header claims "Every `unsafe` block in this file carries a `// SAFETY:` proof", and that was true — it just did not cover the 84 unsafe OPERATIONS sitting directly in `unsafe fn` bodies, which under edition 2021 need no block and so carry no proof. That is the whole CCD surface: every QueryDisplayConfig, every SetDisplayConfig, every DISPLAYCONFIG union read. Rather than 84 restatements of the same argument, the shared contract is stated ONCE at the top — how the buffer-size/query pair keeps pointer and count in agreement, and why the two obligations the compiler cannot see (arrays only ever come from the OS or from a SavedConfig this module built; SetDisplayConfig serialised under the manager `state` lock and bound to the input desktop) hold at every site. Per-site comments then name the site-specific fact instead of repeating the invariant. The union reads justify differently and the header says so: `modeInfoIdx` and `ADVANCED_COLOR_INFO.value` overlay same-sized POD, so no discriminant can be got wrong, while `MODE_INFO.Anonymous.sourceMode` IS discriminated by `infoType` — and every read of it is guarded by that check. Writing that down found the one place the guard does NOT gate the read: the `then_some` in set_virtual_primary_ccd evaluates eagerly, so POD-ness is what carries it there, not the check. The comment says so rather than implying a guard that is not doing the work. Two shape changes, both deliberate. `wait_mode_settled`'s two calls are nested rather than `&&`-joined so the second CCD query stays short-circuited — it polls every 25 ms and must not run while the target has no active path. And the eager union read in `set_virtual_primary_ccd` is hoisted to a `let` so its proof has somewhere to attach; `then_some` already evaluated it eagerly, so nothing changes. Union field ASSIGNMENTS are safe in Rust — only reads are — so two of them keep no block. Verified on the Windows CI runner: clippy -D warnings clean across pf-frame, pf-win-display, pf-capture and pf-vdisplay, and both crates' tests pass. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-07-28 19:45:11 +02:00
enricobuehlerandClaude Opus 5	6782aa0054	fix(vdisplay): stop the suite reporting ok for tests that never ran, and let clippy see the KWin backends ... Three coverage holes, all of the same shape — a gate that looks green because nothing is behind it. The Windows half (~3,400 lines) had no executed tests. Its only two `#[test]`s opened with `if env::var("PUNKTFUNK_PF_VDISPLAY_LIVE").is_err() { return; }`, so without the hardware they passed without running, and no CI job compiled the crate's test targets at all — `-p punktfunk-host` builds pf-vdisplay as a dependency, which never reaches a `#[cfg(test)]` module. Both live tests are `#[ignore]`d now (an unrun hardware test should say `ignored`), and windows-host.yml gains the `--all-targets` lint plus a real test step. The link question that keeps the host and pf-encode on clippy-only does not apply: pf-encode is `default = []` and nothing here turns on nvenc/amf-qsv/qsv. Settled on the runner to the same standard as pf-capture's step — 46 passed, 2 ignored. `#![allow(clippy::all, ...)]` at the top of kwin.rs and kwin_output_mgmt.rs was meant for wayland-scanner output, but the generated modules already carry their own copy of that attribute — so the file-level one was only ever silencing 2,366 lines of hand-written backend, with `clippy::correctness` (deny-by-default) among the casualties. Dropping it found a dead match arm in the `kde_output_configuration_v2` dispatch; that one is now exhaustive on purpose, so a re-vendored XML that adds an event fails the build instead of dropping it silently. (The sibling `_ => {}` in the DeviceMode dispatch is NOT dead — `ModeEvent::Removed` really does reach it and get swallowed, which is a separate defect.) `dead_code` is now enforced on Linux, where ~10k of the ~17k lines live, instead of allowed crate-wide behind a rationale that had stopped being true. It turned up no genuinely dead code: four items, every one live from tests or another platform, now annotated with which. One deserved the comment most — `Entry::keepalive` is "never read" because it is an RAII guard whose `Drop` releases the compositor output, so deleting it as unused would release every pooled display the moment it was created. Two tests were not testing what they claimed. `another_uids_socket_is_ignored` asserted the sway-socket ownership guard but never reached it — `find_sway_socket` filters on the `sway-ipc.<uid>.` filename prefix first, so the foreign-uid file was rejected by name and the assertion held even with `md.uid() != uid` deleted. It is renamed to what it does cover, and the real leg is a new `#[ignore]`d root test that chowns a correctly-named socket (querying with a non-matching pid, or the exact-path shortcut would return before the guard and make it vacuous all over again). And proc.rs's tests were ungated while the module is compiled everywhere, so they would have gone red on Windows the first time anyone ran them: `#[cfg(all(test, unix))]` now, with a `cmd /c` twin that does run there. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-07-28 19:45:11 +02:00
enricobuehlerandClaude Opus 5	995cac5d03	fix(vdisplay): the three hardest FFI calls were exempt from the crate's own unsafe-proof rule ... `#![deny(clippy::undocumented_unsafe_blocks)]` cannot see an unsafe operation that sits directly in an `unsafe fn` body — in edition 2021 `unsafe_op_in_unsafe_fn` is allow-by-default, so a bare call inside such a body needs no block and therefore carries no proof. The file headers claim "every unsafe block carries a SAFETY proof", and that was true; it just did not cover the calls that needed it most. Turning the lint on named exactly three: `DeviceIoControl` and the `ioctl` wrapper in pf_vdisplay.rs — the whole host<->driver control channel — and `restore_displays_ccd` in manager.rs, the call the teardown path depends on to give the operator their physical panels back. Each now carries a real proof, and `ioctl` and `set_render_adapter` grew the `# Safety` heading their callers were owed. Also teaches scripts/wincheck.sh to cover pf-vdisplay, which is what let the above be compile-verified rather than reasoned: the crate's Windows half is ~3,400 lines that Linux CI never compiles. That needed a stub pf-encode (the real one drags ffmpeg-sys, and the admission gate calls exactly one predicate from it) and CompositorPref on the stub core. Verified non-vacuous with a planted type error. pf-win-display's half of the same lint is a separate change — it flags 62 further operations, and they want proofs written one at a time, not in bulk. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-07-28 19:45:11 +02:00
enricobuehlerandClaude Opus 5	9b38b6a22d	feat(stats): a capture records which encoder and GPU produced it ... The stage split is the thing an fps-shortfall report is diagnosed from, and it cannot be read without knowing the backend behind it: a p50 `submit` of 10 ms means "the GPU's CSC+encode throughput is the ceiling" on one backend and something else entirely on another. The capture's meta carried the mode, codec and client but neither the encoder nor the GPU, so every report so far cost a round-trip asking which one it was. Both come from `pf_gpu::active()` — the record the encoder open itself writes — so they name the branch that really opened rather than a re-derived guess that goes stale the moment a backend gains an internal fallback. Read once per capture, not per frame. The console shows them next to the mode in the recording's header. Both fields are `serde(default)`, so a recording made before this still loads and simply omits them. Refs #9 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-07-28 19:14:10 +02:00
enricobuehlerandClaude Opus 5	eb4ab91627	feat(host): a frame limiter for the game, not for the stream ... PUNKTFUNK_MAX_FPS caps how fast the compositor lets the game render. It deliberately does NOT touch the session: the client still negotiates and receives its full rate, because the encode loop re-encodes the held frame whenever the compositor produced no new one, and a repeat of an unchanged picture is an almost-empty P-frame. So a 60-capped game on a 120 Hz session still puts 120 frames a second on the wire — and the GPU time the game gives up goes to capture and encode instead, which is the point (on a laptop or a handheld, it goes to heat and battery too). Capping the stream would be a different and mostly unwanted feature: it hands the client fewer frames than it asked for and saves the game's GPU nothing. gamescope is the compositor that has the lever, so it is the one that gets it: the rate becomes --nested-refresh, which is what gamescope clamps the game to. All three sessions we own take it — the bare spawn's -r, the managed gamescope-session-plus wrapper's PF_HZ, and the SteamOS PATH shim's. In the managed path the limit lands on PF_HZ alone and NOT on CUSTOM_REFRESH_RATES, so the mode the session advertises stays the client's — that is what makes games see the real refresh rather than the box's EDID. Two scoping notes. Under gamescope the cap is the nested output's rate, so everything it composites moves at it, not the game alone; there is only the one output. And the attach path (PUNKTFUNK_GAMESCOPE_NODE) mirrors a gamescope we do not own, so it has no lever to pull and is untouched. Closes #13 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-07-28 19:14:10 +02:00
enricobuehlerandClaude Opus 5	86b605f01f	feat(host): run the virtual display faster than the stream needs ... PUNKTFUNK_VDISPLAY_HZ_MULT runs the virtual display at a multiple of the session's rate without putting one extra frame on the wire. A compositor paints on its own vblank, so a frame finished just after the capture sampled waits nearly a full interval to be picked up — the jittery part of the latency budget, not the steady part; at 2 that worst case halves. It costs the compositor and GPU the extra composites, so it stays opt-in at 1. The pacing rate was previously just "whatever the backend achieved". It is now the session's rate floored by the achieved one — the same value as before whenever the knob is unset, and the only correct answer when it is set: never above what the client negotiated, never above what the display actually produces. Both build sites get it, the initial build and the in-place resize, so a mid-stream resize can't silently drop back to 1x. A backend that won't give the multiplied rate now says so at info rather than warn, since that is the expected outcome of an opt-in knob; falling short of the SESSION's own rate still warns, because that one costs the client frames. Closes #14 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-07-28 19:13:19 +02:00
enricobuehlerandClaude Opus 5	f98547c41f	fix(web-console): the paired-clients card counts both pairing planes ... GameStream and native (punktfunk/1) clients pair into separate stores, and `/status` only ever reported the GameStream certificate count. Native is the DEFAULT plane, so a host whose clients had all paired normally — and which Moonlight had never touched — showed "0 paired" on the dashboard. Report the native count alongside it, the way the tray's own summary route already does, and sum the two in the card. Closes #10 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-07-28 19:13:19 +02:00
enricobuehlerandClaude Opus 5	0c4e582b29	fix(gamescope): a managed session now proves our flags reached it ... The bare spawn builds gamescope's argv itself. The two managed modes hand the flags over indirectly — `gamescope-session-plus` through `GAMESCOPE_BIN` + `PF_HDR_ARGS`, SteamOS through a PATH shim — and a session free to ignore either would exec the distro's gamescope carrying none of them. Half of that failure was already loud: HDR dies on the bit depth the Welcome fixed before the display existed. The other half was silent. The host had been told the compositor would paint the pointer, so it stopped painting one, and nobody did — a stream that is correct in every respect except that it has no cursor, which nothing in the logs would have said. So both managed paths now read the running compositor's `/proc/<pid>/cmdline` once its node appears, and refuse the session when a flag we passed is missing. The plan is fixed by then (`cursor_blend` feeds the encoder open, which precedes the display), so this session cannot be corrected in place; instead the capability is latched off for the process and the spawn fails, and the retry resolves a correct SDR host-composited session. One rejected attempt per boot, then it converges. Fails OPEN at every ambiguity: no flags expected, or no readable gamescope in `/proc`, says nothing. And it accepts ANY running gamescope carrying the flags, because a box commonly has a second one — the Nobara test box runs its own game-mode `/usr/bin/gamescope --steam` beside ours. That leaves only false PASSES possible, and the flag that matters cannot produce one: `--pipewire-composite-cursor` exists nowhere but our patch set. Two things fell out of writing it. `gamescope_argvs` matches argv[0]'s basename with `ends_with`, not `==` — the old equality test would have missed our own `punktfunk-gamescope`, which is what `current_gamescope_output_size` was already scanning for. And `hdr-probe` gained a line for who paints the cursor, the one capability with no symptom of its own until you compare two streams. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-07-28 18:03:30 +02:00
enricobuehlerandClaude Opus 5	ee716d0137	fix(encode): pf-encode did not build on Linux without vulkan-encode ... `vulkan_encode_available_at` and `vulkan_encode_caps` were gated on bare `target_os = "linux"`, but the second returns `vulkan_video::VulkanEncodeCaps` and both reference that module — which only exists under the feature. Every CALL SITE was already correctly gated, so nothing pointed at them; the two definitions alone were enough to fail the build. That is CI's default-feature line (`cargo clippy --workspace --all-targets`), so this was going to be caught — it was caught on a Fedora 44 box first. `cursor_blend_capable`'s `ten_bit` goes unused in the featureless arm for the same reason, which `-D warnings` also rejects. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-07-28 18:03:30 +02:00
enricobuehler	3d3ecf1e82	test(encode/nvenc): the NVIDIA HDR leg, verified on an RTX 5070 Ti ... The NVIDIA half of "zero-copy, no host CSC, 10-bit HDR" had never met a GPU. Two `#[ignore]`d smokes now drive it, run on `.41` (Bazzite f43, RTX 5070 Ti, driver 595.58.03): * `nvenc_cuda_hdr10_packed_rgb` — a packed 2:10:10:10 CUDA payload straight into NVENC as `ARGB10`, HEVC **and** AV1. Asserts what would catch a mislabelled stream: the encoder DERIVED 10-bit and HDR from the input format rather than being told, and picked `ARGB10` for `X2Rgb10`. That derivation is what selects Main10 / AV1-at-10 and the BT.2020 PQ signalling. * `nvenc_cuda_hdr10_cursor_blend` — `cursor_blend.comp` MODE 3/4, the 10-bit channel-unpacking twin. Asserts the blend targets `SlotFormat::X2Rgb10` and not the 8-bit layout, which would tint the pointer and shift its channels. Both green, and the dumped bitstreams decode 0-error reporting `Main 10` / AV1 `Main`, `yuv420p10le`, `bt2020nc` / `smpte2084` / `bt2020`. There is no host colour conversion anywhere on this path: the frame arrives as a LINEAR dmabuf, crosses to CUDA through the Vulkan bridge, and NVENC's ASIC does the BT.2020 conversion following the VUI the session configured. The "no CSC" property AMD gets from the EFC, NVIDIA gets from the encoder itself. The whole pre-existing `nvenc_` suite was re-run alongside them — 16/16, no regression from the depth-derivation and buffer-format changes underneath it. Capture half checked too: the patched gamescope binary built on the AMD box runs unmodified on the NVIDIA one (same Fedora 43 base) and its node offers the same four formats with the colorimetry props — so headless gamescope + 10-bit PQ is not AMD-specific.	2026-07-28 18:03:30 +02:00
enricobuehler	576bf7e294	test(encode/vulkan): 10-bit smokes — and they pass on a real AMD GPU ... The Vulkan Video 10-bit path was the least-exercised code on this branch: nothing had ever created a Main10 video session, so the profile query, the `G10X6…3PACK16` picture allocation, the scratch→plane copy, the hand-packed AV1 sequence header and the colour signalling were all first-run-on-glass. Three `#[ignore]`d smokes now drive them, and they were run. `.116` (Bazzite f43, AMD 780M, RADV / Mesa 26.0.4), all three green: * `vulkan_smoke_10bit` — HEVC Main10 through the compute CSC; * `vulkan_smoke_10bit_av1` — AV1 at 10 bits; * `vulkan_smoke_rgb_10bit` — HDR with NO host CSC, the EFC converting BT.2020 off the packed 10-bit source. It did **not** soft-skip, which answers the one capability in this whole feature I had only ever read in a registry: RADV's VCN EFC really does advertise `MODEL_YCBCR_2020` and accept a 10-bit packed-RGB encode source. Every stream decodes 0-error and reports `yuv420p10le` + `bt2020nc` / `smpte2084` / `bt2020`. The AV1 one parsing at all is the load-bearing result there: `high_bitdepth` precedes the CICP bytes in `color_config()`, so a wrong bit would have thrown every later field out of phase rather than merely mislabelling the depth. Round-trip on solid frames, fed (160,160,800) as 10-bit codes: compute CSC -> (159, 158, 796) EFC -> (159, 158, 796) AV1 -> (158, 159, 794) Under 0.5%, all of it limited-range quantisation and lossy encode. The compute and EFC results being IDENTICAL is the strongest check available: two independent BT.2020 NCL implementations — my shader and AMD's fixed-function block — agreeing to within rounding. The smokes deliberately assert structure (submits encode, AU count, the depth the encoder settled on), not colour: a shader writing the 10 bits into the wrong end of the word still produces a decodable stream. Colour is the dump + ffmpeg round-trip above, which is what actually caught nothing this time. Also corrects a comment: I claimed a wrong store factor was a "~1.6% luminance error". It is not. The `<< 6` PLACEMENT is the load-bearing part (dropping it is 64x too dark); `1/1023` vs `64/65535` is ~0.1% and harmless.	2026-07-28 18:03:30 +02:00
enricobuehler	fd648aa776	feat(gamescope): put the cursor in the node, so a session can be zero-copy ... gamescope keeps the pointer out of its PipeWire node — it lives on a hardware plane for scanout, and `paint_pipewire()` composites a separate, reduced frame that never includes it. So punktfunk has always reconstructed it from XFixes and blended it in host-side. That blend is what has been blocking the zero-CSC encode path, and the cost is larger than it sounds: `VulkanVideoEncoder::open` refuses the RGB-direct (EFC) source for any session with `cursor_blend`, because that front end is fixed-function and has no blend stage. `cursor_blend_for` sets it unconditionally for gamescope. Net effect: a gamescope session paid a full-frame colour-conversion pass per frame, forever, for a pointer. So put the cursor where it belongs. A second carried gamescope patch adds `--pipewire-composite-cursor` (off by default — the node has never carried it, and a consumer that draws its own would get two), painting it with the same `MouseCursor::paint` call the scanout composite uses. It scales for free: paint_pipewire has already set `currentOutputWidth/Height` to the capture size, which is what that function scales against. The repaint test grows the cursor's state beside the commit ids — a pointer-only move produces no commit, so without it the composited cursor would freeze on a static screen while the real one moved, and a cursor that became hidden would never be erased. Host side, the `+pfhdr` marker becomes a monotonic PATCH LEVEL, so one probe answers every capability the session must know before it is planned (level 1 = HDR formats, level 2 = the cursor flag). `cursor_blend_for` and the `gamescope_cursor` resolver both consult it through one helper, because they have to agree: the reader without the blend is a wasted X11 connection, the blend without the reader is a stream with no pointer, and both together with a gamescope that paints its own would draw two. ⚠ The two indirect spawn modes carry the flag through `PF_HDR_ARGS`, so this shares a dependency with the HDR flags: a session that ignores `GAMESCOPE_BIN`/`PATH` and execs the distro's gamescope gets neither. HDR fails loudly there (negotiation timeout + SDR latch); a missing cursor would be silent. Noted in packaging/gamescope/README.md as worth a post-spawn `/proc/<pid>/cmdline` check if it ever bites.	2026-07-28 18:03:30 +02:00
enricobuehler	2a13bc5252	fix(gamestream): advertise AV1 Main10, and honour HDR per negotiated codec ... `ServerCodecModeSupport` layered `SCM_HEVC_MAIN10` and never `SCM_AV1_MAIN10`, on the stated theory that "the GameStream AV1 path is left off until live-confirmed". But the SDR baseline has always offered AV1 **Main8** to every client, so that path is either live or it is not — the DEPTH was never the uncertain part, and the omission only cost AV1-preferring clients their HDR. Now that the encoders probe 10-bit per codec, each bit is gated on that codec's own `can_encode_10bit` AND the SDR baseline already advertising it. A box that does HEVC Main10 but not 10-bit AV1 — or the reverse — advertises the truth instead of one bit standing in for both. Two gates follow from that: * `host_hdr_capable` becomes codec-agnostic (ANY 10-bit-capable codec makes the host HDR-capable). It was asking about HEVC alone, which would have hidden HDR entirely on a hypothetical AV1-only-10-bit box; * the RTSP honor gains the per-session half: a client that negotiated the codec this host CANNOT do 10-bit with degrades to 8-bit SDR there, rather than being handed a PQ label over an 8-bit stream. H.264 always lands there — there is no 10-bit H.264 encode anywhere. The unit test now pins each bit independently, including the two one-without- the-other cases a single shared flag got wrong in both directions.	2026-07-28 18:03:30 +02:00
enricobuehler	cb4690f216	feat(encode/vulkan): probe the device instead of guessing — AV1 10-bit + zero-CSC HDR ... Three fixes to the same mistake: deciding what the Vulkan Video backend can do from a table in our heads rather than from the driver, and routing everything that didn't fit to libav VAAPI — where a session loses real RFI recovery and the cursor blend for no reason the hardware asked for. **Capability probe, per codec AND depth.** `probe_encode_support`'s "is there an encode queue" boolean becomes `VulkanEncodeCaps { supported, eight_bit, ten_bit }`, answered by `vkGetPhysicalDeviceVideoCapabilitiesKHR` against the very profile chain the session open builds. So the dispatcher's prediction cannot disagree with reality: a capable device keeps the Vulkan path, an incapable one routes to VAAPI BEFORE burning a failed open, and the cursor-blend mirror stays honest for free. This is the shape the direct-SDK NVENC path already uses for its codec GUIDs. **AV1 10-bit.** It was excluded on a guess about driver coverage; now the device answers. `color_config()` carries `high_bitdepth` + the BT.2020/PQ CICP triplet in both the `StdVideoAV1ColorConfig` and the sequence-header OBU we bit-pack ourselves — they must stay identical or the driver's frame OBUs parse against a header we didn't write. `high_bitdepth` sits BEFORE the CICP bytes, so getting it wrong doesn't just mislabel the depth, it puts every following field one bit out of phase; the new test reads the packed bits back. **Zero-CSC RGB-direct in HDR.** The EFC probe assumed BT.709 and BGRA. It now asks for the model this session's colourimetry needs (`MODEL_YCBCR_2020` for 10-bit — the extension has always had it) and for the CAPTURED format as an encode-source format, and the session create-info selects the matching model. An HDR session with no pointer to composite therefore hands the captured buffer straight to the fixed-function front end and runs no host CSC at all. Sessions that DO composite a pointer keep the compute CSC, unchanged: the EFC cannot blend, and that rule outranks everything. Also: `can_encode_10bit` on AMD/Intel now reports the union of VAAPI's and Vulkan Video's answers instead of VAAPI's alone. `open_amd_intel` tries Vulkan first and falls back, so either one being able to encode Main10 makes the session 10-bit-capable — answering `false` because only one of them said yes stranded encodable HDR sessions at 8 bits.	2026-07-28 18:03:30 +02:00
enricobuehler	479f0965ee	feat(encode/vulkan): Vulkan Video encodes 10-bit, so AMD/Intel HDR keeps the good path ... The Vulkan Video backend was 8-bit for no structural reason — the API has `VK_VIDEO_COMPONENT_BIT_DEPTH_10_BIT` and `PROFILE_IDC_MAIN_10` in the very fields this pinned to 8 and MAIN, and AMD VCN and Intel both encode Main10. It was six hardcoded sites, and the cost of leaving them was paid twice over: an HDR session had to take libav VAAPI, losing real RFI loss recovery AND the compute CSC's cursor blend — which on gamescope is the only way the pointer reaches the stream at all, since gamescope has no embedded-cursor mode. An HDR session now opens a Main10 profile with 10-bit component depths, a `G10X6_B10X6R10X6_2PLANE_420_UNORM_3PACK16` picture + DPB, and an SPS carrying `bit_depth_*_minus8 = 2` with the BT.2020/PQ CICP triplet instead of BT.709. `rgb2yuv10.comp` is the CSC's twin, and the two interesting parts of it are: * it is a PURE 3x3 matrix. The samples arrive already PQ-encoded (gamescope composites into the PQ container), so BT.2020 NCL applies to the code values as they are — there is no transfer function to apply here and applying one would be wrong; * the scratch planes are `R16`/`RG16`, not the picture's plane formats. The 10-bit ycbcr plane formats are not storage-image formats, so the shader writes the value into the HIGH bits by hand (`code10 << 6`, hence the `64/65535` factor and not `1/1023`) into planes that are merely SIZE-compatible with the picture's — which is all `vkCmdCopyImage` requires. Scope and safety: * HEVC only. AV1 10-bit encode has far thinner driver coverage, and a session open is not the place to gamble on it — those stay on VAAPI, as does a device that fails the Main10 profile query inside the open (the pre-existing "failed Vulkan open falls back to VAAPI" net, no new probe needed). * HDR pins the compute-CSC arm over the EFC RGB-direct one, which the EFC could not serve anyway: its fixed-function conversion is 8-bit BT.709 narrow with no knob for BT.2020. * `open_inner` binds `hdr` to the parameter-set HEADER bytes, so the depth flag is `ten_bit` there — the one name collision this change had to route around.	2026-07-28 18:03:30 +02:00
enricobuehler	17f824c3e9	feat(encode/nvenc): an HDR capture stays zero-copy on NVIDIA ... AMD/Intel needed no new encoder code for HDR — the VAAPI path already ingests an XR30 dmabuf into `format=p010:out_color_matrix=bt2020`. NVIDIA did: the 10-bit formats were excluded from the GPU import outright, so an HDR session fell back to a CPU readback plus swscale, which is the one thing the capture path is not allowed to ship. It turns out no CSC kernel is needed. NVENC ingests packed 10-bit RGB natively as `ARGB10`/`ABGR10` and does the conversion itself following the configured VUI matrix — which `apply_low_latency_config` already sets to BT.2020 NCL for an HDR session. So the frame travels LINEAR dmabuf → Vulkan bridge → CUDA → NVENC unconverted: no host CSC pass, no depth loss, no extra work on a contended SM. * invariant 1 is restated rather than dropped: HDR must never take the TILED EGL de-tile blit (it renders into an 8-bit `GL_RGBA8` texture). The HDR pods are LINEAR-only by construction, so the plan may build the importer; the per-frame gate — which sees the negotiated modifier the plan cannot — is what enforces the tiled half, and falls back to the CPU path if a producer ever ignores our offer. * …but only where the encoder can actually take the payload (`linux_hdr_cuda_ok`). libav's HDR route builds a P010 hardware frames context and swscales into it, so on a host without the direct-SDK backend a packed-2:10:10:10 CUDA buffer would land in a P010 surface as garbage. Those keep the CPU path. * `nvenc_cuda` stops pinning 8-bit/SDR. Depth and HDR now follow the INPUT format, like the Windows backend: a 10-bit session whose capture came back 8-bit encodes AND labels 8-bit rather than mislabelling. * the cursor-blend compute shader gains two 10-bit modes, so the pointer gamescope leaves out of its node survives the HDR path. Same display-referred blend the CPU path's `composite_cursor_rgb10` already does — the samples are PQ, and a real sRGB→PQ cursor LUT is polish, not correctness for a pointer.	2026-07-28 18:03:30 +02:00
enricobuehler	689297c86e	feat(hdr): a gamescope session streams true HDR10, decided before the Welcome ... The Linux native plane has been 8-bit for a reason that stopped being true: `capturer_supports_hdr()` returned a flat `false` because Mutter's virtual monitors are SDR-only upstream. That is still right for Mutter, KWin and wlroots — and wrong for gamescope, whose node can now offer 10-bit BT.2020 PQ (packaging/gamescope). Open the three gates that held it off, together: * the capture-side gate becomes SOURCE-AWARE (`capturer_supports_hdr_for`): Windows keeps its platform answer, gamescope asks whether the resolved binary offers the formats, everything else stays false. It must be truthful rather than optimistic — the Welcome is irrevocable, and PQ frames on an 8-bit encoder are a deliberate hard error — so every term is a static fact resolved before anything is spawned, including "do we spawn gamescope at all" (an attach inherits someone else's flags and can promise nothing); * `want_hdr` reaches the capturer: `open_virtual_output` grew the parameter it never had, and the host arm stopped dropping `want.hdr` on the floor; * the spawn gains `--hdr-enabled --hdr-debug-force-support` on all three sub-modes (bare args, the `GAMESCOPE_BIN` wrapper, the SteamOS PATH shim), from one shared `hdr_args` — the force flag is what makes it work headless, and forgetting it looks exactly like a negotiation failure. Two things that would have gone wrong quietly: * the keep-alive reuse key gains `hdr`. gamescope cannot turn HDR on live, so a kept SDR display handed to an HDR session would give the game no HDR surfaces while the stream negotiated PQ over an SDR composite — wrong, and not obviously broken. Same for the managed session-plus/SteamOS reuse check. * the HDR-negotiation-failure latch becomes PER SOURCE. It was one process-wide flag, so a monitor that left HDR mode would have disabled gamescope HDR until the host restarted, and vice versa — two facts with nothing in common but the word HDR. GameStream follows the same shape: `host_hdr_capable` gains the gamescope arm, and the RTSP gate's live BT.2100 monitor probe is scoped to the portal source — a headless box has no monitor to be in HDR mode, so running it there would hard-refuse every gamescope HDR session. Off by default for one release (`PUNKTFUNK_GAMESCOPE_HDR=1`), and `punktfunk-host hdr-probe` now answers for both Linux HDR sources.	2026-07-28 18:02:13 +02:00
enricobuehler	28c50d1c5b	fix(encode): every backend signals its colour, so no decoder has to guess ... Three encode paths shipped a bitstream with no colour description at all, leaving primaries/transfer/matrix/range "unspecified": - Vulkan Video HEVC (`vk_build.rs`) built an SPS with no VUI whatsoever. This is the DEFAULT backend for AMD/Intel Linux hosts on HEVC/AV1. - Vulkan Video AV1 packed `color_description_present_flag = 0`. - The openh264 software path wrote nothing (it converts BT.709 limited and relied on decoders defaulting to that). - The libav-NVENC Linux path excluded packed-RGB 4:2:0, on the belief that "NVENC's internal CSC writes its own VUI". It doesn't: libavcodec derives `colourDescriptionPresentFlag` from the AVCodecContext colour fields, so leaving them unspecified emits none. Reachable on a CPU/dmabuf capture, a build without `--features nvenc`, or PUNKTFUNK_NVENC_DIRECT=0. Unsignalled looks fine on every punktfunk client — `csc_rows` falls back to BT.709 on "unspecified" — which is why this survived. Vendor TV decoders do not: they guess colorimetry from RESOLUTION, and an LG webOS panel reads a 4K SDR stream as BT.2020 and renders it visibly washed out. All four now signal BT.709 limited, which is what every host CSC actually produces (`rgb2yuv.comp`, `convert_bt709`, the swscale paths) and what the Welcome's `ColorInfo::SDR_BT709` already advertises out-of-band. NVENC, VAAPI, QSV, AMF and the Windows libav path were already correct. Two tests, both parsing the REAL emitted bitstream rather than re-asserting the constants: an independent bit-walk of the AV1 sequence header (the packed OBU must stay identical to the `StdVideoAV1ColorConfig` handed to the driver), and an H.264 SPS/VUI parse proving openh264 honours the request instead of dropping it. Not yet verified on hardware: the HEVC VUI depends on the driver's SPS writer emitting `vui_parameters()`. PUNKTFUNK_VULKAN_ENCODE=0 falls back to VAAPI if a driver mishandles it. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> (cherry picked from commit 3c56ff5717b2c9a0871953127da3dadd6a84220d)	2026-07-28 17:01:59 +02:00
enricobuehler	0868b6a364	fix(vdisplay): compositor availability follows the live session, not the env ... GNOME/Mutter reported "Unavailable" on a host sitting in a live Mutter session — and, on the same request, "Default", because the two columns had different sources. available() asked each backend, and those probes read the process env (XDG_CURRENT_DESKTOP for Mutter, WAYLAND_DISPLAY for KWin's registry handshake, SWAYSOCK for sway) — env a host started outside the session (systemd --user, a TTY, ssh) never inherited. It is only retargeted at the live session on the connect path, so the answer also flipped depending on whether anyone had connected yet. Both columns now come from the same /proc scan detect() already used: the live session's compositor is usable by definition, as is an explicit operator pin, and the per-backend probe stays as the fallback for backends that are not the live session (gamescope, which spawns its own). A live KWin without the zkde_screencast grant now surfaces as available and fails at create with that probe's precise message, which beats "no usable compositor" on a box visibly running KDE. Mutter's env sniff stays deliberately narrow — one var, not three. XDG_CURRENT_DESKTOP is the one apply_session_env owns end to end (written per connect, scrubbed when nothing is live); sniffing DESKTOP_SESSION alongside would resurrect the bug that scrub exists to prevent, where a stale value after a gnome-shell crash routes the next client into a dead session. Non-Linux hosts now report no compositors at all rather than five Linux backends flagged unavailable with no default — on Windows the pf-vdisplay driver is the only backend and vdisplay::open ignores the argument, so the old list read as broken detection instead of "not applicable here". The console says so explicitly. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> (cherry picked from commit eb7ba3d6177552f5c3ed6a15439404084869c636)	2026-07-28 17:01:59 +02:00
enricobuehler	5ea087ca47	feat(session): the stats overlay scales with the display's DPI ... The stream chrome — stats OSD, capture hint, start banner, resize label — was hardcoded at 14 px with 12/10/8 px insets. The overlay composites into the swapchain 1:1 in PHYSICAL pixels, so on a 4K panel at 200 % all of it rendered at half its intended physical size: fine on a 1080p monitor, a squint on a HiDPI laptop. FrameCtx now carries a scale — SDL's window display scale (DPI × the display's content scale) times a PUNKTFUNK_OSD_SCALE preference — and every metric moved into a `base` module that is multiplied by it. Re-read per frame and quantized into the damage key, so dragging the window to a differently-scaled monitor re-renders at the new size rather than keeping the stale one. Sanitized because SDL returns 0.0 when it cannot resolve the window's display, which would collapse the panel to nothing. Two details worth keeping: the face is re-derived at the scaled size rather than the canvas transformed, because Skia rasterizes glyphs at the requested size where a magnified 14 px bitmap would be mush; and the long capture hint is fit-clamped to the window — at 2× it is wider than a 1080p screen, so scaling it naively would have traded a small OSD for a truncated one. Linux and Windows share this path. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> (cherry picked from commit 744467d13a28b91ba88a23d6038da70263e9b502)	2026-07-28 17:01:59 +02:00
enricobuehler	b444308592	feat(host): PUNKTFUNK_HOST_NAME names the host in Moonlight and the clients ... A box called `bazzite-htpc` had no way to present itself as "Living Room" short of renaming the machine. The new knob overrides the name everywhere a human sees it: the GameStream serverinfo <hostname> element and the mDNS service instance name both adverts carry. Unset (the default) is the machine's own hostname, exactly as before. Free text is the point, so the DNS-level name is now a separate concern from the display name. The instance label may contain spaces and accents; an A-record target may not, and mdns-sd rejects the whole ServiceInfo if the target is not a legal name — which would take discovery down rather than merely look wrong. dns_label() sanitizes the target and passes an already-legal name through byte-for-byte, so hosts without the override advertise precisely what they always did. The display name loses `.` (it would split the label, and clients derive the name as the first label of the fullname, so "Ben's PC v1.2" would arrive as "Ben's PC v1") and is capped at the 63-byte DNS-SD ceiling on a char boundary. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> (cherry picked from commit bbf72261a12e0e67601f549d40db65ae61979268)	2026-07-28 17:01:59 +02:00
enricobuehler	188f55d3b1	fix(encode/nvenc): the host advertises what the driver lists, not a superset ... Every NVIDIA host advertised a static H.264|HEVC|AV1 superset, so a 1st-gen Maxwell (GTX 960M, no HEVC/AV1 encode) offered HEVC — a client that believed it got ~15 s of blank video and a disconnect instead of a stream. Both OSes now ask the driver itself (nvEncGetEncodeGUIDs) on one throwaway direct-SDK session: Linux on the shared CUDA context, Windows on the selected render adapter, wired into host_wire_caps AND the GameStream serverinfo mask (which had been left on the superset for NVIDIA on both OSes). Fails open — an unanswerable probe keeps the historical superset, so it can only ever narrow the advertisement to codecs the GPU really encodes. The HEVC 4:4:4 answer rides the same session on Linux instead of opening a libav hevc_nvenc FREXT probe: that open is the prime suspect for the field bug where one probe wedges NVENC process-wide (NV_ENC_ERR_INVALID_VERSION on every later session until a host restart), and the direct backend re-checks the same caps bit at session open anyway. The ffmpeg probe remains only for hosts that really stream over libav (PUNKTFUNK_NVENC_DIRECT=0 or a build without the nvenc feature), where ffmpeg's NVENC client runs regardless. The 10-bit probe deliberately stays libav — Linux HDR rides the libav P010 path. On-hardware: .136 (RTX 5070 Ti) 14/14 nvenc tests in one process incl. the probe followed by real sessions and dirty teardown; .173 (Windows RTX) probe + 47 release lib tests. The Windows probe test documents the pre-existing MSVC debug-link failure (LNK2019 via the sdk crate's unused lazy loader) — run it with --release, the same reason windows-host.yml gates with clippy. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> (cherry picked from commit 0346ec8090568eb499e8cb7d735305b28471185e)	2026-07-28 17:01:59 +02:00
enricobuehlerandClaude Opus 5	560e663aef	fix(drivers): the pad channel asks the devnode who to trust, not the mailbox ... A LocalService principal could take over a virtual pad's shared input section and forge HID input into the interactive desktop. The host duplicates each pad's unnamed DATA section into the driver's WUDFHost, and through gamepad proto v2 it learned that process from `driver_pid` in the named bootstrap mailbox. That mailbox has to be LocalService-writable — that is what the driver's own WUDFHost runs as — and the delivery gate, verify_is_wudfhost, only checks that the target's IMAGE is %SystemRoot%\System32\WUDFHost.exe. That image is world-executable. So anything running as LocalService — notably the deliberately de-privileged plugin runner — could spawn its own WUDFHost (CREATE_SUSPENDED parks it indefinitely with the right image path), publish that pid, and be handed SECTION_MAP_READ|WRITE on a live section. For pf-mouse that section drives a real absolute pointer, so it was desktop control; for the pads it was forged gamepad input plus a read of the remote user's controller state. The module docs claimed mailbox tampering "yields at worst a gamepad DoS, never a read or an injection". That was wrong, and the reasoning behind it — that a LocalService token is DACL-denied OpenProcess on a UMDF WUDFHost — only covers the REAL host, not one the attacker spawned itself. The pid now comes from the device stack (ChannelProof, proto 2 -> 3). The host asks the devnode it SwDeviceCreate'd who is serving it, looked up by the instance id PnP handed back, so a planted look-alike devnode is not a candidate and the kernel — not anything the attacker supplies — does the routing. Only the driver PnP actually bound to that device can answer. `driver_pid` survives as a liveness hint; a tamperer can still deny a pad, which squatting the name always allowed, but can no longer choose the recipient. Two rules keep the state machine honest around it: a delivery stands until its target process EXITS (judged on a retained SYNCHRONIZE handle, so a recycled pid cannot fake it, and UMDF's restart-after-driver-crash still re-attaches), and a pad with no SwDeviceCreate devnode refuses to deliver rather than fall back — unless an operator sets PUNKTFUNK_PAD_CHANNEL_TRUST_MAILBOX, which says so loudly. Three transports, because Windows carries different things to different driver shapes, and the obvious two did not survive contact with hidclass. Measured on .173 (Win11 26200): HidD_GetIndexedString is NOT forwarded to a UMDF HID minidriver at all — it failed for every index including ones the driver demonstrably serves through the named wrappers; and a private device interface registers and enumerates but cannot be OPENED (ERROR_GEN_FAILURE), because hidclass owns IRP_MJ_CREATE on a devnode it is the FDO for. That is exactly why pf-xusb was never affected: it is not a HID minidriver, so nothing sits above it. What works: * pf-xusb — a private IOCTL on its own GUID_DEVINTERFACE_XUSB. * pf-mouse — the HID serial string. Verified: PFCP:3:0:7296, and 7296 was a genuine service-spawned WUDFHost.exe. Safe here alone: nothing reads the virtual mouse's serial, whereas a pad's is SDL/Steam dedup material. * pf-gamepad — a HID feature report, and it cost NO report-descriptor change. The captured descriptors already declare far more Feature ids than the driver ever served: 0x85 is declared on DualSense, DualShock 4 and Edge alike and used to fail with STATUS_INVALID_PARAMETER, so hidclass lets it through and nothing can have depended on the old failure. The Deck's one feature report is unnumbered and Steam drives it command->response, so its proof rides that existing contract via a private two-byte command. Verified: feature 0x85 returned magic "PFCP", proto 3, pad_index 0, wudf_pid 18456 — and 18456 was a WUDFHost — with the product string still 'DualSense Wireless Controller'. Also renamed pf-dualsense -> pf-gamepad. One driver has always served four identities, so the old name read as if the other three lived elsewhere. ONLY the package identity moved (crate, INF/CAT/DLL, UMDF service, build script, CI lines, log file, env var). The four HARDWARE IDS are deliberately unchanged — they bind every devnode the host creates and every installed system — as are the Global\pfds-boot-<i> mailbox and PAD_MAGIC, which are wire contract. `driver install --gamepad` now retires the pre-rename store package first, matched on pf_dualsense.dll because that string appears only in the OLD inf; matching on the hardware ids would delete what we are about to install. On .173 that separated 14 stale packages from the 1 new one with 0 ambiguous, and the renamed package binds the old hwid (devgen root\pf_dualsense -> oem143.inf = pf_gamepad.inf). The repo's own pre-commit/pre-push rustfmt hooks named the old crate, so they caught the rename before the commit did — they now check pf-gamepad, and pf-mouse alongside it, which they had been missing relative to the CI line. Host and drivers MUST ship together: v2<->v3 fails closed in both directions by design, with the existing "update host + drivers together" diagnostic. The rename moved files that also carry the security change, so splitting this into two commits would mean reconstructing an intermediate state that was never gated. It is one commit on purpose. Gated on the windows-amd64 runner with cargo clean first (the box's clock lags, so stale artifacts would read as a vacuous green): clippy -D warnings clean for pf-inject, pf-capture and pf-driver-proto, drivers workspace build + the CI clippy line clean, cargo check --release -p punktfunk-host clean, 19 + 58 tests green. Also fixes pf-mouse still writing its debug log to world-writable C:\Users\Public, which the 2026-07-17 review moved for the other three drivers and missed here. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-07-28 16:54:40 +02:00
enricobuehlerandClaude Fable 5	9b3ec9204c	fix(capture): a refused EGL→CUDA dmabuf offer stops being asked ... The raw-dmabuf passthrough has had a negotiation-timeout latch since the hybrid-Intel case: one conclusive timeout and later captures negotiate the CPU path instead of re-paying it. The EGL→CUDA dmabuf-only offer had no equivalent — a compositor that accepts none of the importer's modifiers timed out the same 10 s negotiation on every session, forever, under the generic 'format negotiation never completed' diagnosis. Now the symmetric latch (note_gpu_dmabuf_negotiation_failed) gates build_importer in the one negotiation resolver, scoped to this offer alone (raw passthrough, worker-death latch, encoder untouched), with the same operator escape as the raw arm: an explicit PUNKTFUNK_ZEROCOPY=1 keeps erroring loudly instead of downgrading. One correctness detail beyond the handoff's sketch: whether the GPU offer was ACTUALLY advertised is a runtime fact of the PipeWire thread (the importer may fail to construct — no CUDA — in which case no dmabuf was offered and a timeout must not latch it off). plan.build_importer alone cannot answer that, so the thread records the made-offer on CaptureSignals (gpu_dmabuf_offer) and the timeout diagnosis branches on the offer that really happened, not the plan's intent. Also renames vaapi_dmabuf_forced → zerocopy_forced (it now guards both timeout arms; single caller). New plan invariant pinned in tests: the latch gates only the importer, never the raw passthrough. (V3 from design/pf-zerocopy-sweep-handoff.md — the deferred design call, now decided in favour of the symmetric latch.) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>	2026-07-28 13:15:36 +02:00
enricobuehlerandClaude Fable 5	e92a0aaa00	fix(zerocopy/vkslot): a timed-out blend can't corrupt its slot, and NV12 cursor chroma sits on the grid ... Two [GPU]-class defects from design/pf-zerocopy-sweep-handoff.md; the code is compile- and unit-verified, the visual halves still owe on-glass time. - blend_ref reused a fence and command buffer after a wait timeout: on TIMEOUT it reset a fence that still had a pending signal and re-recorded a command buffer the GPU might still be executing — reached exactly in the contended case the CPU-synced path exists for (no timeline export + visible cursor + a >1 s GPU stall). A failed wait now drains the device before the reset (the VkBridge::import_linear precedent), and free_slots quiesces unconditionally instead of only when a timeline exists, so teardown after a timed-out sync blend no longer frees objects an outstanding submission references. (C1) - cursor_blend.comp anchored NV12 chroma blocks to the cursor's oy, not the surface chroma grid: for odd oy every UV sample averaged luma rows one below the rows it covers (a one-row colour fringe on the cursor's edges, ~half of all cursor positions) and the cursor's last row's chroma was never written. Block rows now anchor to the chroma grid the same way spans anchor to the word grid — y0 = floor(oy/2)*2, per-row cy guards for the straddle block — and blend_geometry counts blocks from the same anchor (one extra block when oy is odd; covered by new tests, including negative oy). cursor_blend.spv rebuilt (glslangValidator, spirv-val clean, drift gate passes). (C6) Owed on-glass: a stalled-GPU cursor session for C1; an odd-oy cursor colour check for C6 (subtle fringe on the cursor's top/bottom edge). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>	2026-07-28 13:01:46 +02:00