2620 Commits

This Branch

This Branch

All Branches

Author	SHA1	Message	Date
enricobuehler	0170da2a5f	fix(client/apple): stop dropping rotation, and stop inventing it ... G13 — the three capture-fidelity findings from the gyro sweep, two fixed and one argued. **The 4 ms floor was a DROP, and it was shedding real rotation.** A sample arriving 3.9 ms after the last one was discarded outright. That is the wrong shape for this signal: buttons and sticks are absolute state, so a dropped frame costs nothing — the next one says everything it would have. Angular velocity is a RATE, and a consumer integrates it into an angle, so a dropped sample is rotation that happened and can never be recovered. GameController's delivery jitters around the pad's own ~250 Hz, so a floor set AT that rate does not shed a rare extra sample; it sheds a steady fraction of every turn. And the error is one-signed, so it accumulates — aim drifting short, which reads as bad sensitivity rather than as a bug. Nothing needed the ceiling. GC delivers at the sensor's rate rather than faster, the SDL client has always forwarded every sample, and the host's idle watchdog is a 100 ms timeout this cannot outpace. The throttle's two fields went with it: `lastMotionNs` was left set-but-never-read once the guard was gone, and `motionIntervalNs` had no other consumer. (Notes elsewhere say `flush` parks motion and reads it — that is PR #88's branch, not this one. Checked rather than assumed.) **An X-Box pad was streaming gyro it does not have.** Capture attached to any `GCMotion`, and an X-Box controller exposes one that reports gravity and NOTHING else. So the client sent a permanently-zero `rotationRate` to the host as authoritative gyro, under a declaration saying this pad has one. That is worse than having no motion plane at all: a game sees a controller being held perfectly still forever, and there is nothing to fall back to and nothing to notice. Now gated on `hasRotationRate`, which is GameController's own answer to the question we actually mean. The settings badge had the same bug from the same cause — `hasMotion` was `motion != nil`, so an X-Box pad got a gyroscope icon. It now reads `hasRotationRate` too. One wrong predicate was driving both the UI promise and the wire behaviour, which is why they were wrong together. That also simplifies G8's "your gyro can't reach this session" notice, which had to test `hasRotationRate` itself to avoid nagging about a gyro the pad never had. With the attach gated on it, the notice is just the else-branch. **Motion stays on the main queue, and this is the argument for why.** GameController's `handlerQueue` is a property of the CONTROLLER, not of an element, so moving motion off main moves buttons, sticks, the touchpad and the escape chord with it. This class is `@MainActor` throughout — eight `assumeIsolated` sites, the slot table, the gesture timers — so that is a rewrite of the isolation model rather than a queue assignment, and it would put the tvOS escape chord (the only controller way out of a stream there) on a background queue. That is a real risk for a speculative gain. The comment says so at the call site, and names the measurement to make first if it ever does bite: the host's per-pad motion inter-arrival histogram already reports exactly this and would say whether the delay is client-side or on the wire. Gate: macOS `swift build` + the full suite (215 tests, 5 skipped, 0 failures) and the iOS-triple typecheck green. No test pins the throttle removal or the capability gate: both are properties of live `GCMotion` delivery, which this module cannot fake — there is no injectable seam, and inventing one to assert "we called sendMotion twice" would test the mock. They are argued at the call sites instead, in the same spirit as the parts of `DsCapture` that are not unit-testable in their module either. On-glass verification is owed with the two already outstanding on that rig.	2026-08-07 19:12:02 +02:00
enricobuehler	d996449a82	fix(host/pads): a virtual pad at rest said it was in free fall ... G14, unblocked by the frame measurement in efb7f991 — the plan deliberately left this one alone until the up axis was known, on the grounds that a confidently wrong constant would be worse than an obviously wrong zero. It is known now. A virtual DualSense, DualShock 4 or Steam Deck that had received no motion reported acceleration `[0, 0, 0]`. That is not "no data": zero proper acceleration means free fall, which is a definite claim about the physical world and one that is never true of a controller sitting on a desk or held in someone's hands — both read 1 g up. Anything that interprets the accelerometer gets a confident wrong answer rather than a boring right one. It is worst exactly where it is least visible. A pad with no gyro at all — an X-Box controller forwarded as a DualSense, which is what "Automatic" does for anything not Sony or Valve — never sends motion, so it sits on that neutral for the entire session, telling every game that reads it that the controller is falling. `switch_proto` has always done this correctly on its own up axis, which is what made the gap visible in the first place. Which axis, and why it took a measurement. The wire is a unit passthrough into the virtual pad's report, so the wire's up axis is the pad's own, and on 2026-08-07 a real DualSense read over raw HID put `+0.997 g` on report axis 1 at rest, in a frame pinned the same session as (Right, Up, Backward). So `MOTION_NEUTRAL_ACCEL` is `[0, 10000, 0]` — NOT the z-up the notes had assumed from `switch_proto`'s documentation, which is why guessing would have shipped a backend confidently disagreeing with the hardware. The constant lives in punktfunk-core beside the units it is expressed in, and every backend derives from it rather than restating it. The Deck's neutral in particular goes through `steam_remap::motion_wire_to_deck`, the same rescale a real sample takes, so the neutral and the live path can never end up with two opinions about what 1 g is — its `hid-steam` resolution stays in exactly one place. The DS4 needs no separate change: it reuses `DsState`. `switch_proto` is deliberately NOT touched, and the test says so. It is a different device on a different driver, its up axis is its own, and nobody has measured its frame — aligning it to the DualSense for consistency would be the same unmeasured guess this commit exists to avoid, just in the other direction. Non-vacuity proven both ways rather than assumed. Moving the up axis to slot 2 (the old z-up assumption) fails on the wire constant itself, which is what makes the measurement load-bearing rather than decorative; reverting both neutrals to `[0, 0, 0]` fails on the DualSense assertion with the message naming the defect. Each backend is checked in ITS OWN units, because hard-coding "1 g" three times is how the halves of a unit contract drift apart. Gate (Linux CI image): fmt, build, `clippy --locked --all-targets -D warnings` across punktfunk-core / pf-inject / pf-client-core, and both test suites — green, with `Running tests/motion_contract.rs` and the new case's own `... ok` line observed in the log rather than inferred from a green exit (`cargo test` stops after the first failing binary, so a green-looking run can mean the contract test never executed at all).	2026-08-07 19:07:30 +02:00
enricobuehler	efb7f99129	fix(client/apple): motion arrived in the wrong frame — measured against a real pad ... G16 step 1, and the second half of what 9e9bb9f4 started. That commit fixed the SIGN of acceleration (Apple reports the gravity vector, pointing down; a pad reports proper acceleration, pointing up). This fixes the FRAME, which is a separate defect and was never going to show up as an inverted axis — it shows up as roll where the game reads yaw. The wire is a unit passthrough. `dualsense_proto::write_report` puts gyro[0..3] and accel[0..3] straight into the virtual pad's report bytes 16.. and 22.., in order, with no permutation — the same slots a real DualSense fills. So the frame the wire is DEFINED in is the pad's own report frame, and forwarding GameController's x/y/z unconverted was speaking a different language with the same vocabulary. Both frames measured 2026-08-07 from ONE physical DualSense on one desk, read twice — over raw HID and through GameController — so this is two readings of the same controller in the same orientations rather than two documents: DualSense report frame: (Right, Up, Backward) axis 0 pitch, 1 yaw, 2 roll GameController frame: (Right, Forward, Up) Right is already slot 0; Up is GC's z and moves to slot 1; slot 2 wants Backward, which is GC's y negated. Hence (x, z, -y), applied to gyro AND acceleration because it is a change of basis and both live in that basis. Notable: the wire's documented naming was right all along — gyro[0]=pitch, [1]=yaw, [2]=roll is exactly what the hardware does. And Android needs no remap at all: it forwards the pad's own axis order un-remapped, which is correct. Its old reading was purely the scale bug f6de620f fixed. Only Apple was converting nothing. How the hardware frame was established, since a wrong frame here is invisible. Gravity at rest put +0.997 g on axis 1. Yaw clockwise-from-above drove axis 1 negative (98% of the rotation), pitch nose-down drove axis 0 negative (100%), roll right-side-down drove axis 2 negative (95%) — plain right-hand rule, and (a0 x a1 = a2) confirms the triad is right-handed. The accelerometer then corroborated the gyro's assignment independently: under pitch-down axis 2 rose 0.160 -> +0.339 (nose down raises the back, so world-up gains a Backward component) and under roll-right-down axis 0 went +0.021 -> -0.197, while yaw left acceleration untouched. Two different physical quantities agreeing on one triad. Apple's frame took four attempts, and the failures are worth recording because each was a different way to be confidently wrong: - peak |w| over a window containing BOTH the tip-down and the return stroke can record the return, with the opposite sign. Yaw (a continuous one-way spin) was unaffected; pitch and roll were exactly the two that disagreed with everything else. - reading `gravity + userAcceleration` when `hasGravityAndUserAcceleration` is FALSE yields a constant (0,0,1) in every orientation. It looks like data. The tell is that it never moves. The client's own else-branch on `m.acceleration` is the correct read and is what the instrument now mirrors. - `da/dt = -w x a` holds only for gravity, so testing it during vigorous waving — when `m.acceleration` carries inseparable linear acceleration — fits nothing. The frame that survived all of that: static poses, three of them, three repetitions each. Nose-down moved axis 1 by -0.635 (so axis 1 is Forward), right-side-down moved axis 0 by -0.686 (so axis 0 is Right), flat put +0.99 on axis 2 (Up). That conclusion holds whether or not the acceleration negation is right, because negating flips the measured vector and the physical direction it represents together. Confidence, stated honestly. The accelerometer half is solid: nine pose measurements, and mapping the flat pose through gives (+0.005, +0.992, +0.192) against the hardware's own (+0.021, +0.997, +0.160) — all three components, including the small tilt term that is what distinguishes this mapping from the five other permutations that also put gravity on slot 1. That the gyro shares the frame unmodified rests on a weaker measurement: a gravity-dominated consistency test that preferred (+x,+y,+z) by 1.22x, which is a margin, not a landslide. It is corroborated by the yaw reading (the one rotation measured without the return-stroke ambiguity) agreeing with right-hand rule in that frame, and by the peak-vs-return mechanism explaining the two that did not. A device-side confirmation is still owed and is listed below. The tests carry the measurements, not just the conclusion. Resting gravity is asserted against BOTH readings of that pose; each rotation is asserted to reach the slot the wire reads it from; and two properties guard the shape rather than the numbers — that the conversion is an isometry (a basis change may not stretch anything) and that it preserves handedness. That last one matters most: a permutation with the wrong number of sign flips is a REFLECTION, which looks plausible axis by axis and inverts every rotation. Mutation-checked: dropping only the negation fails 6 assertions across 4 of the 5 cases, the handedness test among them. Owed, and not claimed done: on-glass re-verification through a real iOS device, together with the two already owed on that rig (the 9e9bb9f4 sign fix and the Android calibration read) — one pass covers all three. G14's DualSense neutral acceleration is now unblocked by this measurement (1 g on slot 1, not the z-up the notes assumed) but is deliberately left to its own change; and that constant must NOT be propagated to switch_proto, which is a different device whose frame nobody has measured. Gate: macOS `swift build` + the full suite (215 tests, 5 skipped, 0 failures) with the five new cases observed in the run's own output, and the iOS-triple typecheck green.	2026-08-07 18:41:54 +02:00
enricobuehler	7cab7ae6bc	feat(client/android): say when a captured pad's gyro can't reach the session ... G8's Android half, and the last of the three clients. Same failure as the other two: a controller with a gyro, in a session whose virtual pad has no motion plane, does nothing when tilted — silently, with no way from the couch to tell that apart from a broken sensor. The fix is the Controller type setting, so the notice names it. Android read neither the requested nor the resolved backend, so this needed a plumb. What it did NOT need was a third copy of the rule. `nativePadMotionReaches` takes the kind a pad declared and answers off `pad_motion_reaches` in punktfunk-core, where the argument and the tests already live. The rule is subtler than it looks — the host builds each pad from its OWN declaration and folds what it cannot build, so neither the declaration nor the session echo answers it alone — and every way of getting it wrong is silent. A Kotlin transcription would have been a third thing to keep in step with the host, which is exactly how the SDL half got it wrong the first time. Asked once per pad, at claim, in `openExternal` — where the pad's kind is already being declared to the host — and the answer held for the pad's lifetime on the `ExternalPad`. Not per sample: this runs at a DualSense's full report rate. `hasGyro` gates only the NOTICE, and defaults to false. `DsCapture` passes true — every pad it captures is a Sony one whose IMU is a headline feature, forwarded on the rich plane. `Sc2Capture` keeps the default, because the Steam Controller 2's motion rides inside the opaque passthrough report that `hidReport` carries, which nothing here may second-guess: warning about motion for a pad that never calls `motion()` would be a notice about a feature the player never lost. The suppression itself is on `motion()` regardless, where it costs a dead pad nothing and stops a live one paying to send samples the host will decode and discard. The notice sits at the BOTTOM of the stream overlay, unlike the mic-chord confirmation at the top. The two can coincide — a pad is claimed at roughly the moment someone might be muting — and one landing on the other would cost the user both. It holds 6 s rather than the mic chord's 1.6: that one confirms something the user just did, this one explains something they did not, in a sentence they have to read. Nulled at teardown beside `onExitArmed`/`onMicChord`, for the same reason those are — a slot closing during release must not poke Compose state on the way out. Not covered by tests, and this is a limit of the module rather than a choice: `GamepadRouter` needs Android plus a live JNI handle, there is no Robolectric here, and the predicate it defers to is pure Rust that already has its table. So the parts that carry the reasoning are argued in comments, as `DsCapture`'s claim/teardown ordering already is. What IS mechanically verified is the piece that a compiler cannot catch and a device would fail on: the JNI symbol `Java_io_unom_punktfunk_kit_NativeBridge_nativePadMotionReaches` is present and global in the built arm64-v8a `.so`, so the `external fun` resolves rather than throwing `UnsatisfiedLinkError` at the first pad. Gate: `:kit:compileDebugKotlin`, `:kit:testDebugUnitTest` (62 cases, 0 failed, read out of the JUnit XML rather than inferred from a green build — unchanged from this branch's previous count), `:app:compileDebugKotlin` and `:app:testDebugUnitTest` (67 cases, 0 failed), with `:kit:cargoNdkRelease` rebuilding the JNI crate clean across all three ABIs, plus `cargo fmt --check` on it. On-glass verification is owed on the rig the earlier legs used, and is worth doing as one pass with the two already owed there.	2026-08-07 17:02:04 +02:00
enricobuehler	aaa58ad817	feat(client/apple): say when a pad's gyro can't reach the session, and stop powering it ... G8's Apple half — the UI hint 77797a9e left owed, plus the suppression, which on this client is worth more than it was on the SDL one. The failure being fixed is entirely silent. A controller with a gyro, in a session whose virtual pad has no motion plane, simply does nothing when tilted: nothing in the app says so, and from the couch a session that resolved an X-Box backend is indistinguishable from a broken sensor. The fix is the Controller type setting, so the hint has to name it — a badge that only said "motion unavailable" would leave the player exactly as stuck. Asked per pad, off what the slot declared, via the predicate punktfunk-core now carries. `GamepadCapture` is the one client where this is naturally per pad already: `openSlot` computes `manager.declaredKind(for:)` and puts it in `slot.pref`, so the question is answered where the pad is opened rather than on every sample. `GamepadType.motionReaches(declared:asked:resolved:)` is static and pure so it can be tested without a live session; the connection's instance method fills in the two halves it owns, and `requestedGamepad` is stored beside `resolvedGamepad` for the same reason it exists in the Rust client — the echo is only this pad's answer when the pad declared what we asked for. Where Apple differs from the SDL client, and better: it never powers the IMU. The existing code already declined to activate sensors when forwarding was off, reasoning that with nothing to forward there is no reason to make the pad stream gyro over Bluetooth and burn its battery — `closeSlot` is careful to power them back down for exactly that reason. A host that built this pad a backend without a motion plane is the same situation, so it takes the same branch. No per-sample check, no handler attached, and a DualSense in an X-Box-class session stops paying for a sensor nobody reads. The hint fires only for a pad that really has a gyro (`motion.hasRotationRate`). A gravity-only GCMotion — what an X-Box controller exposes — would otherwise produce a notice about a feature the player never had. That is a narrower condition than the capture path itself uses, deliberately: making the capture gate agree is G13's job and its own change. The badge sits in the bottom-centre stack with the muted-mic badge and the start-of-stream banner, at every stats tier and with the overlay off, because this is not a statistic. Unlike the mic badge it is not a control: the setting is not reachable mid-stream on every platform and applies from the next session anyway. So it states the fact, names the setting, and leaves after the banner's same 6 s. Every platform including tvOS — a DualSense on an Apple TV is an ordinary way to play, and is exactly the pad this happens to. The model owns the expiry rather than the view, so a second pad's hint replaces the first cleanly instead of stacking, and ending the session cancels a pending clear rather than carrying a stale hint into the next stream. Non-vacuity proven by mutation, not assumed: collapsing the predicate to `resolved.hasMotion` fails 4 assertions, including the mixed-pad row that is the whole reason it is not a session-level check. The table mirrors the Rust one row for row — a client that disagrees with the host here either kills a working gyro or streams ~250 Hz into a void, and both are silent. Gate: macOS `swift build` + the FULL suite (210 tests, 5 skipped, 0 failures) with the two new cases observed in the run's own output, and the iOS-triple typecheck green (`arm64-apple-ios17.0`, iOS slices + hand-assembled xcframework per the memory recipe) — the badge and the overlay it joins are on every platform, so the macOS build alone would not have covered them. tvOS remains unverifiable from this Mac; the badge deliberately reuses the neighbouring banner's shape rather than introducing anything tvOS-specific.	2026-08-07 16:55:03 +02:00
enricobuehler	5c4969fd6b	fix(client/pads): the gyro cut-off asked about the session, not the pad ... Supersedes the check 77797a9e shipped an hour ago. The suppression, the log-once, and the "unknown must not suppress" rule all stand; the field it reads does not. 77797a9e read `Welcome.gamepad` — the backend the host resolved for the SESSION — and stopped sending motion when it had no motion plane. But the host does not build pads from that. It builds each virtual device from that pad's own `GamepadArrival` (`Pads::set_kind`) and falls back to the session default only for a pad that never declares one, which is precisely why `declared_kind` exists and why its doc comment says an explicit setting has to be re-declared per pad. So the check had a false negative, and it is an ordinary living-room setup. Under "Automatic" the Hello carries the ACTIVE pad's kind (`auto_pref`), so a couch with an X-Box pad on slot 0 and a DualSense on slot 1 echoes Xbox360 — while the host, reading pad 1's arrival, builds it a DualSense with a working motion plane. The old check read the echo, saw no motion plane, and killed pad 1's gyro. That is the exact failure 77797a9e's own commit message names as the worse of the two ("a false negative kills working motion"), introduced by the fix for the other one. The question is per pad, so the slot now carries what it declared, beside the physical `pref` it already held. The two are deliberately separate fields answering different questions: `pref` is the controller in the user's hands, which is what the local feedback paths must keep reading, and `declared` is the one the host is pretending to have. Three facts decide the predicate, and they are written out in `pad_motion_reaches` rather than at the call site because all three clients need the same reasoning: - the echo is not this pad's answer when the pad declared something else; - the host FOLDS what it cannot build — a Switch Pro on Windows, any UHID backend on a host whose /dev/uhid is unusable — and nothing client-side can predict it; - but the echo IS one observed sample of that fold, for the kind the Hello asked about, so it is authoritative for a pad that declared exactly that. Hence: trust the echo when declared == asked, else fall back to the declaration. That keeps both motivating cases — a generic pad under Automatic (declares X-Box 360, suppressed, the sweep's H5c) and an explicit Switch Pro folded to X-Box 360 by a Windows host (declared == asked, so the echo catches it, H5d) — where either field alone gets one of them wrong. `requested_gamepad` is kept on the client next to `resolved_gamepad` for this: the pair is what makes the echo usable per pad, and a lone field would only tempt the next reader back into the session-level question. The residual gap is a pad whose declared kind differs from the session's AND gets folded: we keep sending and the host keeps dropping. That is the direction to be wrong in, and it is what the session-level check was worth in the first place — wasted datagrams, not a dead gyro. Non-vacuity proven both directions rather than assumed. Reverting to `resolved.has_motion()` fails on the mixed-pad row; reverting to `declared.has_motion()` (no echo at all) fails on the Switch-Pro-on-Windows row. Each case in the table is a session someone can actually sit down to, and the comment on each says which of the three inputs decides it. Gate (Linux CI image, pf-lxcheck2): fmt, `build -p punktfunk-core`, `build -p pf-client-core`, `clippy --locked --all-targets -D warnings`, and both test suites — green, with the new case observed in the run's own `... ok` line rather than inferred from a green gate, and pf-client-core's 163 unchanged.	2026-08-07 16:46:06 +02:00
enricobuehler	8e8d30202c	fix(client/android): a Sony pad's buttons no longer wait on its calibration ... Supersedes the parse gate in 26b0819f. The off-thread read, the claim token, the teardown ordering and its bounded wait all stand — only what happens in the gap changes. 26b0819f held every report back until the calibration read came home, so a pad that stalled on EP0 could feel dead for up to the link's 250 ms timeout: no buttons, no sticks, nothing. Reports are now forwarded immediately and their motion scaled by the nominal calibration until the real one lands. That gap is exactly the behaviour that shipped before f6de620f — acceleration ~18% short, gyro unscaled — for about a millisecond. Nobody can feel that. A controller that ignores a button press for a quarter of a second is not in the same category, and it is the only one of the two a user would ever report. It is also the safer of the two conservatisms available here. The rejected third option, forwarding motion as zeroes until the real numbers arrive, would have the host read a still pad as being in free fall — a lie about the physical world rather than an imprecision about it. The nominal constants are merely a slightly wrong scale. The token is more load-bearing under this, not less. With a gate, an unpublished calibration meant "parse nothing"; now it means "scale nominally", so begin() clearing the previous pad's value is the whole reason a re-claim falls back to the nominal constants instead of silently inheriting factory numbers belonging to a different unit — which are, in general, further off than nominal. The fallback therefore lives in the hand-off itself (MotionCalHandoff.effective) rather than as an elvis at the call site: restoring the gate now means changing the type's API, not deleting three characters in onReport. The tests moved with the contract. They assert the nominal calibration is what is in effect during the gap, rather than merely that the slot is empty — an empty slot is now compatible with either behaviour, so asserting on it would have let a regression pass. Added the case the change exists for: the same raw report, parsed either side of publication, forwards identical buttons and sticks while its gyro and acceleration convert differently. Mutation-checked three ways — dropping the nominal fallback fails all five cases, dropping begin's clear fails the inheritance case, dropping the token check fails three. Gate: `:kit:compileDebugKotlin`, `:kit:testDebugUnitTest` and `:app:compileDebugKotlin` green on a forced clean rerun, 62 cases across the module, 0 failed, with the five hand-off cases read back out of the JUnit XML. The on-glass re-verification f6de620f owes is still owed and unchanged.	2026-08-07 16:23:48 +02:00
enricobuehler	26b0819f5c	fix(client/android): plugging in a Sony pad could hitch the interface ... Supersedes the synchronous calibration read f6de620f shipped an hour ago. The ordering it protected is kept; the blocking it cost is not. f6de620f read the pad's calibration inline in DsCapture.startUsb, which runs on the main thread — the stream's setup path, and the USB-permission broadcast. The read is a blocking EP0 control transfer: a pad that is there answers in about a millisecond, but a pad that is stalling takes the link's whole 250 ms write timeout, and either way the interface was waiting on a controller. That is the wrong thread for it. It now runs on its own daemon thread, one per claim, named pf-ds-cal — the same shape HidUsbLink already uses for its reader rather than a second style. A pathological stall now delays the pad's motion by a moment instead of freezing the UI. What kept the ordering honest before was "assign the calibration before `model`", since `model` is what lets the link thread into the parse. That reasoning stands, so the gate simply moved: MotionCalHandoff holds the claim's calibration, starts null, and onReport parses nothing until it lands. No report is ever scaled by the last pad's numbers — those are per unit — nor by the nominal fallback the real read is about to replace. Dropping the first millisecond of a capture costs nothing: the reports carry absolute state, so the next one says everything the dropped one would have. The calibration is what got deferred, not `model`, and that is deliberate. Keeping `model` synchronous keeps isActive, the teardown writes, the feedback sinks and the active-changed true/false pairing meaning exactly what they meant yesterday — and, more to the point, it makes a late completion structurally unable to resurrect a dead capture. A straggler can only ever publish a calibration, and nothing is parsed while `model` is null. Teardown, which is where this sort of change actually bites. Both stop() and the unplug path end the claim before they close anything: ending burns the token, so a read that lands afterwards publishes nothing and says so in the log. They then wait, bounded at 500 ms and normally already over, for the read to let go of the connection they are about to close — closing a descriptor with a transfer in flight pulls it out from under the kernel, the same rule the pad-audio borrow follows. It cannot deadlock: the reading thread blocks on the EP0 transfer and on the hand-off's own monitor, never on anything a teardown holds. If a pad has stopped answering entirely the wait elapses and teardown proceeds regardless, which is the same exposure the feedback writes already carry and better than an interface that never comes back. Tested where it is testable. MotionCalHandoff is the piece that carries the hazard and it is pure, so it has its own test: nothing is visible until the read lands, a read that outlived its claim publishes nothing, a re-claim never inherits the previous pad's calibration, and a doubled end still refuses every outstanding token. Mutation-checked both ways — deleting the token check fails 3 of them, deleting begin's clear fails the fourth. Not covered: DsCapture's own claim/teardown ordering is not unit-testable in this module — there is no Robolectric, and the class builds a main-Looper Handler and needs a UsbManager — so it is argued in comments rather than pinned. The on-glass re-verification f6de620f owes is unchanged and still owed. Gate: `:kit:compileDebugKotlin`, `:kit:testDebugUnitTest` (61 cases across the module, 0 failed) and `:app:compileDebugKotlin` green, with the four new cases confirmed present in the JUnit XML rather than assumed from a green build.	2026-08-07 16:10:22 +02:00
enricobuehler	f6de620f34	fix(client/android): a captured Sony pad's gyro turned the wrong amount ... G14/G16 leg 3. This supersedes the nominal constant 0e40b374 shipped, which was always labelled a stopgap. Measured on glass 2026-08-07: a DualSense over USB into an Android phone, streaming to a Linux host, flat and face up, arrived as |accel| = 0.811 g where 1.000 was owed. The parse forwarded the pad's raw i16s verbatim, and raw device units are not wire units. 0e40b374 rescaled acceleration by the nominal 10000/8192 and deliberately left gyro alone, because a constant provably cannot fix gyro: the same still average showed this unit's accel calibration is near-identity (~1% off) while its gyro's emphatically is not — a near-identity gyro calibration would imply 1024 LSB per deg/s, i.e. ±32 deg/s full scale, which no controller has. That scale is per unit, and the only thing that knows it is the pad. So the client now asks. HidUsbLink grows a GET_REPORT path — EP0, the exact mirror of the SET_REPORT it already had — and DsCapture reads the pad's IMU calibration feature report ONCE, while claiming it: 0x05 / 41 B on a DualSense or Edge, 0x02 / 37 B on a USB DualShock 4. DsDevice.MotionCal then applies hid-playstation's own arithmetic per axis, which is the same math the host's contract test (crates/pf-inject/tests/motion_contract.rs, SonyImuCalibration) reads from the other end: gyro raw × speed_2x × 20 / (|plus−bias| + |minus−bias|), accel (raw − (plus − range/2)) × 20000 / range. Long arithmetic, because the gyro multiplier overflows an Int, and clamped, because both are >1 multipliers and a full-scale flick would otherwise wrap the i16 into a motion in the opposite direction. Reading the blob also removes acceleration's residual ~1% factory bias that the nominal constant left behind. Once at claim and never per report. EP0 is independent of the interrupt endpoints so the read is safe alongside the reader thread, but a blocking control transfer in the report path would wreck capture latency, and the calibration is fixed for the life of the connection anyway. The capture logs the derived resolutions, which is the discriminator for whether a blob was read at all: a real pad declares ≈16 LSB per deg/s, the fallback reads back as exactly 20. A pad that refuses, answers short, or declares zeroes (a clone, a broken unit) keeps today's behaviour per axis — nominal accel, gyro straight through. Nothing here ever zeroes motion: slightly mis-scaled beats silent. Not covered. The axis frame is still untouched: this leg puts gravity on Y where the Apple leg put it on Z, so at least one client's frame is wrong, and settling it needs the bare-metal Linux reference reading G16 step 1 calls for. Rescaling is frame-independent, so it stands however that resolves — remapping is not, so it stays out. Bluetooth's grouped plus/minus layout is not implemented either: this path is USB-only by construction (Android exposes no raw path to a Classic pad), and a half-used generalisation would be a latent bug rather than a feature. Gate: `:kit:compileDebugKotlin` + `:kit:testDebugUnitTest` green, 16 DsDeviceTest cases run 0 failed, and the five new ones were confirmed present in the JUnit XML rather than merely compiled. Non-vacuity checked by mutation — perturbing the gyro conversion fails 6 tests, including all four new ones that assert a number. On-glass re-verification owed, on the rig that measured the defect (DualSense → USB → phone → 192.168.1.21): at rest |a| = 1.00 g exactly via ~/gyroscope.py, and a nominal 90 deg yaw integrating to ~90 deg via ~/integrate.py — the same 90 deg that read ~62.7 deg before this change.	2026-08-07 15:47:33 +02:00
enricobuehler	0e40b374e7	fix(client/android): DualSense acceleration arrived ~18% short ... G16 leg 2. A DualSense over USB to an Android phone, streaming to a Linux host, flat and face up: |accel| = 0.811 g where 1.000 is owed. Magnitude is frame-invariant, so this is unambiguous regardless of the separate axis question below, and it came from a 27-second static average — no sampling error in it. `DsDevice` said so plainly: "Gyro/accel stay in raw device units". It read the i16s out of the pad's report and forwarded them verbatim. But raw device units are not wire units — the wire is fixed at 10000 LSB/g and the pads' native resolution is the 8192 that hid-playstation calls DS_ACC_RES_PER_G. 8192/10000 = 0.819 predicted against 0.811 measured. Acceleration is now rescaled on both the DualSense and DualShock 4 parse paths, clamped because the multiplier is >1 and a real near-full-scale slam would otherwise wrap the i16 into an impossible acceleration in the opposite direction. Two things deliberately NOT done. Gyro is left alone. It is almost certainly low by the same mechanism, but it cannot be corrected with a nominal constant the way acceleration can: the still average shows this pad's accel calibration is near-identity (~1% off), while the gyro's emphatically is not — a near-identity gyro calibration would imply 1024 LSB per deg/s, i.e. ±32 deg/s full scale, which no controller has. Fixing gyro means reading the pad's calibration feature report and applying its own numbers, which also removes acceleration's residual 1% bias. `HidUsbLink` can SET_REPORT but has no GET_REPORT path yet, so that is a real change rather than a constant, and it is owed. I tried to pin the gyro factor by integrating the on-glass rotations instead: a nominal 90 deg yaw integrated to ~88.5 deg through the Apple client (correct) and ~62.7 deg through Android. Directionally consistent, but the readout samples at 5 Hz and a ~1 s rotation is badly undersampled, so that ratio is not a constant anyone should ship. Recorded, not used. The axis frame is also left alone. This leg puts gravity on Y where the Apple leg put it on Z, so at least one client's frame is wrong — but Android forwards the pad's own axis order un-remapped, which makes its reading evidence about the hardware rather than about us, and resolving it needs the bare-metal reference reading G16 step 1 calls for. Every bare-metal Linux box was unreachable (Deck down, HTPC down, .25 is another KVM guest). Rescaling does not touch axis order, so this fix stands however that resolves. Gate: `:kit:compileDebugKotlin` and `:kit:testDebugUnitTest` green, JNI libs built clean at the API-28 floor across 3 ABIs. On-glass re-verification owed: re-run the at-rest reading and expect 0.99-1.00 g.	2026-08-07 15:23:10 +02:00
enricobuehler	9e9bb9f466	fix(client/apple): acceleration was upside down — measured on glass ... G16, first result. A DualSense paired to an iPhone, streaming to a Linux host, lying flat and face up: hid-playstation decoded z = −0.99 g where a DualSense owes +1.00. Vector magnitude was 1.006 g, so the scale was already correct — this is purely direction, and it was wrong for every accelerometer sample the Apple client has ever sent. The cause is a convention mismatch, not a sign typo. Apple reports acceleration as the gravity VECTOR, which points down: a device face-up on a table reads z = −1. An accelerometer physically measures proper acceleration, and at rest that is the +1 g normal force pushing UP — which is what a DualSense's report, and therefore our wire, carries. The two are exact negatives. Both branches were affected, because `m.acceleration` follows the same Apple convention as the gravity/userAcceleration split, so reading the "raw vector" was not an escape from it. `rotationRate` is a true angular rate and needs no flip. The same session confirmed that independently: rotating the pad clockwise seen from above produced a negative yaw, which is correct under the right-hand rule about an up-pointing Z. That asymmetry — accel wrong, gyro right — is itself evidence for this diagnosis rather than a blanket frame error, and it is why the fix is three negations at one site instead of a remap. The sweep predicted this ("Apple accel plausibly INVERTED — CoreMotion gravity -1 g vs DS +1 g up at rest") but could not confirm it without hardware. It is now measured, and the mechanism is confirmed in the code rather than inferred from the number. Method, for whoever repeats it: the readout is python-evdev on the host reading the virtual pad's own motion node, dividing by the axis `resolution` the kernel publishes, so it prints deg/s and g. That is downstream of the calibration blob — the same layer a game reads — which is what makes a sign error visible to a human at all. Two things this does NOT establish. The host was a KVM guest, so the DualSense could not be attached natively for a side-by-side reference reading; the test stands on the DualSense convention being a fixed property of the hardware, which is decisive for the at-rest sign but weaker for the gyro axis ORDER. And the fix itself is unverified on glass: confirming it needs a rebuilt client on the device, so someone should re-run the same at-rest reading and see +1.00. Gate: `swiftc -parse` clean. A full typecheck needs the gitignored PunktfunkCore.xcframework assembled first and has not been run.	2026-08-07 14:56:27 +02:00
enricobuehler	cbfa03b7ad	docs(host/pads): the SC2's bInterval is already 1 kHz — don't "fix" it to 250 Hz ... Working G14/G18 turned up two sweep findings that do not survive contact with the code. Neither is implemented; one is now guarded. The 2026-08-07 sweep read the Triton (Steam Controller 2) usbip endpoint's `bInterval: 1` as 125 µs — an 8 kHz duplicate storm — and the plan's G14 says to raise it to 4 "like the Deck". That reading assumes a high-speed device, where bInterval is the 2^(n-1) × 125 µs exponent. Both Triton devices declare `UsbSpeed::Full`, and on a full-speed device the field is a plain frame count in milliseconds: 1 means 1 ms, which is the 1 kHz the existing comment claims. Raising it to 4 would mean 4 ms — a 4× cut to the motion rate a passed-through SC2 delivers, in the name of fixing a problem it doesn't have. The endpoint now carries the reasoning so the next reader doesn't repeat it. G18's first bullet ("bound/rate-cap the host's rich-input channel; motion is unbounded") is stale rather than wrong — it was true of the tree the sweep read. Current main already routes rich input, motion included, through a 1024-deep `sync_channel` whose `offer()` helper `try_send`s and drops on full, ending the loop only on Disconnected. That is the same bounded-queue pattern the mic plane adopted for security-review S6. Nothing owed. G14's remaining bullet — DS/Deck neutral accel should read 1 g on the up axis instead of 0 g free-fall — is deliberately NOT done here. Which axis is up is precisely what G16's on-glass session measures: `switch_proto` documents the wire as z-up and its neutral ships +Z, but the Deck's kernel negates Z/RZ, so guessing would leave one backend confidently disagreeing with another. A wrong constant is worse than the current obviously-unset 0. Gate: fmt, build, clippy --all-targets -D warnings, and the test suites — green.	2026-08-07 14:00:30 +02:00
enricobuehler	77797a9e20	feat(client/pads): stop streaming gyro into a session that cannot receive it ... G8 of the gyro program, SDL-client half. The `Welcome` has always carried the backend the host actually RESOLVED, which is not necessarily the one the client asked for — Auto lands on Xbox 360 for anything not Sony/Valve/Xbox, and a Switch Pro on a Windows host folds to X360 too. No client read the field. So a player with an 8BitDo, or a Switch Pro on Windows, got a controller whose gyro did nothing, with nothing anywhere saying why: the client shipped ~250 Hz of Motion datagrams and the host parsed and discarded every one. `GamepadPref::has_motion()` answers whether a backend has a motion plane at all. The SDL client checks it on the first gyro sample: it logs one line naming the resolved backend and pointing at the fix (pick a DualSense-class controller type), then stops sending. Once per slot, not per sample — this path runs at the pad's sensor rate. `Auto` deliberately answers true. It means "unknown" — an old host that omitted the echo, which may well have resolved a DualSense — and suppressing motion on unknown would silently break working gyro, a worse failure than sending datagrams nobody reads. The predicate is an exhaustive match so a new backend has to state its answer rather than inherit one, and a table test pins both halves: a false negative kills working motion, a false positive keeps the void open, and both are silent. Owed: the plan wants this surfaced as a one-line UI hint, not just a log line. Apple already stores `resolvedGamepad` and Android needs the plumb; neither is done here, and both want their own gate. Gate (Linux CI image): fmt, build, clippy --all-targets -D warnings, and the test suites — green, with the new capability test observed running.	2026-08-07 13:52:27 +02:00
enricobuehler	ce5047f3ad	fix(host/pads): the Windows driver stops halving motion and stops serving torn reports ... G6 + G15 of the gyro program. G6 — the UMDF gamepad driver's input path. Its timer ran at 8 ms and completed one pended READ_REPORT per tick, so a game could observe at most ~125 Hz while clients stream motion at ~250 Hz: every other sample was overwritten in the slot before anything read it, and the ones that survived carried up to 8 ms of extra latency. For gyro, a dropped sample is not a dropped frame — it is rotation that never reaches the game. The timer now ticks at 2 ms (about a real DualShock 4's Bluetooth cadence). Only the cheap half runs on every tick: read the input slot, complete one pended read. The channel handshake and the health marks stay on their historical ~8 ms, because they cost more, nothing wants them faster, and `driver_heartbeat`'s documented "+1 per ~8 ms tick" is what the host reads as liveness. The same slot is a single unqueued buffer that both sides touch without a lock, so a driver read landing mid-copy handed the game a report that was half the previous frame and half the next. For a button that is a one-tick glitch; for motion it is a spike in angular velocity, which an integrator turns into aim movement. `PadShm` gains an `input_gen` seqlock (v2.3, carved from reserved space inside the v2 legacy region): the host takes it odd, fences, writes the 64 bytes, and stores it even; the driver samples it either side of its read and retries once. The old code's own comment called this out as a known residual — it is now closed rather than documented. Version posture matches the ring's, with one simplification: no capability stamp is needed, because an old host never writes the field and a constant 0 is indistinguishable from "no write in flight", so a new driver against an old host behaves exactly as it does today, and an old driver ignores the field entirely. The Steam Deck write path had neither the seqlock nor even the trailing Release its DualSense sibling carried; all three Windows backends now publish through one `publish_input`. G15 — motion-cadence observability. The host already computed the measurement a "gyro feels floaty" report needs (client inter-arrival percentiles), but kept ONE global accumulator, so two motion-capable pads in a session interleaved into each other's gaps and produced a number describing neither. It also sat at `debug` behind a `tracing::enabled!` check, so a field log arrived with nothing in it and the only way to get the measurement was to ask for a re-run. Now per-pad and always on, summarized at `info` when the session ends — the moment a field report is being written. It costs one subtraction and one array increment per sample: percentiles come from a fixed log2 histogram instead of a growing sorted Vec, so there is no allocation, no per-window sort, and no way for a client streaming as fast as the link allows to make the instrument expensive. Percentiles are reported as bucket upper bounds (`_le`), which is a factor-of-two answer to a question whose answers are orders of magnitude apart. Gaps of 500 ms or more are counted as stalls rather than folded into the percentiles — an interruption is not a cadence, and averaging it in would report a healthy feed as a terrible one. Gates. Windows CI runner .133, the drivers workspace on the real WDK: cargo build, clippy -D warnings (which enforces the unsafe-audit lints), and fmt — all green, against a source whose SHA-256 matches this commit's. Linux CI image: fmt, build, clippy --all-targets -D warnings over pf-inject / punktfunk-core / punktfunk-probe / pf-client-core / pf-driver-proto / punktfunk-host, and the test suites including the 5 new motion-cadence tests — all green. Not measured on glass. G6's stated gate is a sensor-rate reading (SDL testcontroller or Steam's calibration screen) that matches the client's send rate; that is still owed, and a driver change only a compile has seen deserves it before anyone trusts the number.	2026-08-07 13:44:54 +02:00
enricobuehler	4834c2ee51	fix(host/pads): DualShock 4 gyro ran 40× fast, and no pad ever stopped turning ... Phase 1 of the gyro program (design/gyro-program.md, G1-G5) — the five correctness fixes under it. Gyro aim integrates angular velocity over time, so each of these is not a cosmetic wrongness: a wrong scale is every rotation being the wrong size, a wrong clock is every rotation being integrated against a fictional dt, and a stale sample is rotation that never happened. G1 — the DualShock 4 calibration blob. A Sony pad does not assume a motion scale, it reads one out of a fixed calibration feature report. Ours declared 0.5 LSB per °/s and 8192 LSB/g while the wire delivers 20 and 10000, so every DS4-type session decoded gyro 40× too fast and acceleration 1.22× hot — since the backend shipped. The blob now states the wire's own units (the DualSense blob's numbers, deliberately: both pads consume the identical wire sample). Its interleaved per-axis order is NOT a bug and stays: the virtual pad declares BUS_USB, where interleaved is the correct layout; grouped is Bluetooth's. The same blob lives a second time in the UMDF driver, which is a separate WDK workspace that cannot depend on pf-inject — one wrong table in two files, where fixing one reads as fixing it. Both are fixed, and the DS4 feature reports now live in dualshock4_proto beside the DualSense's rather than in the Linux backend, so there is one canonical copy to point at. Field hosts keep the old blob until they update the host package. G2 — the gate that would have caught it. Nothing pinned any backend's declaration against the wire, so tests/motion_contract.rs now applies the CONSUMER's arithmetic (the kernel's, and SDL's, which differ) to each backend and asserts the result lands back on the wire constants — for the DualSense and DS4 blobs, and for the Deck and Switch Pro rescales. It also parses the driver's Rust source and re-derives the units from THAT, so the two copies cannot drift. Verified non-vacuous both ways: re-introducing the old blob fails with "declares a fractional 32/64 LSB per °/s", and reverting only the driver's copy fails with "the UMDF driver's DS4_FEATURE_CALIBRATION has drifted from pf-inject's". The wire units themselves move to punktfunk_core::input::gamepad, referenced by the client's capture scale, the Deck/Switch rescales, and the probe — whose at-rest vector said 16384 (a driver's number, not the wire's) and now says 1 g. G3 — real sensor clocks. The DualSense advanced its sensor timestamp by +1 raw unit per report (0.33 µs — a frozen clock) and the DS4 by a flat +188 (~1 ms) regardless of the real 4-8 ms cadence. Anything integrating rate × dt off that field got nonsense. All four backends now stamp elapsed monotonic time in their own units via a shared SensorClock, anchored to the pad's first report so an irregular publish loop cannot make it drift, and truncated to the field width — which reproduces the wrap real hardware does. G4 — motion is level-triggered and had no watchdog. merge_frame preserves the last sample and the heartbeat re-emits it, so a feed that stops leaves the pad rotating forever — and with G3's honest clock, at a dt that keeps growing. Rumble and the pen plane each have an idle timeout; motion now has one too, at 100 ms. Angular velocity only: acceleration is kept, because gravity is legitimately persistent and blanking it reads as free-fall. The SDL client parks its gyro at zero when a slot closes, which is the case we can flush rather than wait out. (The Apple half of this rides in PR #88.) G5 — a pad returning inside the 300 ms replug grace keeps the same device and skips the create path, so a different controller inherits the previous one's touch contact and rotation — and a pad with no gyro never sends a sample to correct it. sweep() now reports re-claims separately from drops, and the manager clears the rich plane on one. Rich fields only: rumble and hidout dedup deliberately survive a removal. Gates (Linux, CI image): fmt, build, clippy --all-targets -D warnings over pf-inject/punktfunk-core/punktfunk-probe/pf-client-core, and the test suites — 110 pf-inject unit + 6 contract + 29 pf-client-core gamepad, all green. Not yet verified on glass; the on-glass sign/scale session is G16.	2026-08-07 12:47:26 +02:00
enricobuehler	c6b183450a	Merge pull request 'docs: the 4:4:4 story catches up with the code that ships it' (#87) from worktree-docs-support-matrix-444 into main ... Reviewed-on: #87	2026-08-07 09:27:32 +00:00
enricobuehler	3a35773b70	Merge origin/main; 4:4:4 has no software floor, and the ABI is 17 ... Reconciling with #85 (FFmpeg is gone from the client). Two of my claims were true against the pre-merge tree and false against this one. Note 4 said the desktop clients need no 4:4:4 decode probe "because every rung can display full chroma — swscale converts for the software rung". There is no swscale any more. The CPU floor is openh264 + rav1d, it is 4:2:0 8-bit by contract and has no HEVC at all, so it refuses a 4:4:4 stream rather than converting one. The client still advertises the bit unprobed, which was the point of the original fix, but the honest reason is different: full chroma is a hardware path (Vulkan RExt, NVIDIA today), and what catches a box whose hardware 4:4:4 fails is note 2's codec reconnect, not a downgraded picture. Note 13 repeated the same wrong premise and now names both halves of its warning. The C ABI is 17, not the 14 I read before the merge. Textual side of the conflict: #85 rewrote the Codecs column and notes 1-3 of the same table while leaving note 4's stale 4:4:4 text alone. Theirs kept in full; only the 4:4:4 column and note 4 are mine.	2026-08-07 11:25:17 +02:00
enricobuehler	5aa1ca392e	Merge pull request 'fix(host/encode): the Windows host build stops failing on unused split-encode helpers' (#86) from worktree-fix-winhost-clippy-dead-code into main ... Reviewed-on: #86	2026-08-07 09:22:57 +00:00
enricobuehler	b245cb6b29	Merge pull request 'FFmpeg is gone from the client — native decode M0–M10' (#85) from worktree-native-decode-m0 into main ... Reviewed-on: #85	2026-08-07 09:21:45 +00:00
enricobuehler	f49f22a292	fix(host/encode): the Windows host build stops failing on unused split-encode helpers ... The Windows host job died in its clippy step: eight items in pf-encode's split-encode policy (`SPLIT_AUTO`..`SPLIT_DISABLE`, `resolve_split_mode`, `max_forced_split_mode`, `clamp_to_engines`) were reported as never used, and `-D warnings` turns that into a build failure. Nothing about the encoder was wrong — the items simply have no reader in one particular build of the crate, and nothing was telling the compiler that. `codec.rs` compiles on every platform, but the split policy only ever has a caller on Linux (the libav NVENC path reads it unconditionally) or on Windows with the `nvenc` feature (the direct-SDK backend). A featureless Windows build of pf-encode has neither, so every item in the cluster is genuinely dead there. Gate them on the union of their callers' cfgs, the way `forced_split_width` next door already is. The step lints pf-encode itself WITH `--features nvenc,amf-qsv,qsv`, where the items are live, which is why this was invisible there; the failure came from the next command in the same step, `clippy -p pf-vdisplay`, which pulls pf-encode in as a plain default-features dependency. Same item-level `dead_code` trap this crate has now hit five times. Verified: default-features pf-encode reproduces all eight errors before the change and none after (macOS default-features exercises the identical "cluster has no caller" arm as featureless Windows — the two remaining errors there, `vbv_frames_env` and a redundant closure call, are pre-existing and macOS-only; both items have real Windows callers). Linux default-features and Linux + nvenc `--all-targets` both stay clean, so the callers still see the policy. `cargo fmt` clean.	2026-08-07 11:11:15 +02:00
enricobuehler	f1e7ec3535	docs: the 4:4:4 story catches up with the code that ships it ... The support matrix said the desktop clients' Full chroma switch "has no effect today" and that only the Apple client asks for 4:4:4. Both stopped being true in July: `clients/session/src/main.rs` advertises VIDEO_CAP_444 whenever the setting is on, deliberately with no client-side probe, because every desktop decode rung can display full chroma — the Vulkan presenter samples the 2-plane 4:4:4 pool formats and swscale converts for the software rung. So Linux, Windows and Apple all ask; Android is the one that genuinely doesn't implement it. The other half was HDR. `9f72a3b6` gave the Windows IDD-push capturer a packed 10-bit BT.2020 PQ RGB output, so NVENC encodes HEVC Main 4:4:4 10 and the two compose — the matrix still said "4:4:4 and HDR together is refused", and hdr.md still called PyroWave the only exception. Linux is the side that keeps the trade: handshake.rs resolves the depth back to 8 for a 4:4:4 session, so full chroma wins and the stream is SDR. Three cells move ❌ → ⚠️ rather than ✅ on purpose. The client half is unconditional, but the host half is not: HEVC 4:4:4 means an NVIDIA host, or PyroWave on any vendor. The notes say which, and point at the stats overlay's `4:4:4→4:2:0` tag — this negotiation is the one that fails loudly. Also: C ABI version 13 → 14; PyroWave's ≈8K 4:4:4 block-index ceiling now has a note; and the roadmap no longer calls Intel 4:4:4 a hardware limit, which the matrix and vaapi.rs both contradict — VCN can't, VAAPI hasn't. Spot-checked and left alone as still accurate: the Linux client clipboard stub, VAAPI declining 4:4:4, Android having no 4:4:4 at all, and the wire / driver / gamepad-channel versions.	2026-08-07 10:58:58 +02:00
enricobuehler	bbbcf321e5	Merge origin/main into worktree-native-decode-m0 ... main moved 93 commits while this branch ran. Two conflicts, both where main's new work sat next to M10's excision: packaging/flatpak/io.unom.Punktfunk.yml — main added the vendored gamescope WSI layer (the only route to HDR on a Deck) and, before it, a vulkan-headers module. Took both: this branch predates them and deletes neither. But the headers module's stated consumer was pf-ffvk's bindgen over FFmpeg's hwcontext_vulkan.h, and M10 deleted pf-ffvk — so it now reads as dead weight to the next person. It is not: the WSI layer IS a Vulkan layer, compiles against those headers, and builds after it, so module order is the dependency. Rewrote the rationale to say so, including why dropping it would be expensive to discover — flatpak.yml has no pull_request: trigger, so a manifest break reaches main invisibly and a tag then ships no Linux flatpak. Also recorded that the native decoder needs nothing from there: pf-vkdecode reaches Vulkan through ash, which is pure Rust bindings, no bindgen, no C headers. crates/pf-console-ui/src/screens/settings.rs — main restructured the gamepad settings into TABS, which removed the per-row section headers; this branch had left Some("Video") untouched from the merge base and added the pre-M10 decoder migration next to it. Git could not tell those apart. Took main's structure (no header, its deliberate change) with this branch's migration layered on: a stored `vulkan`, `vaapi` or `d3d11va` names no preset in the tabbed list and would render as "—", then silently rewrite the user's preference on the next save. Gates on the merged tree, Linux container: fmt clean; cargo check --workspace --all-targets clean; clippy --workspace --all-targets -D warnings clean; tests green across pf-vkdecode (187), pf-client-core (163), pf-console-ui (58) and punktfunk-host (447 of 448 — the one failure is the pre-existing gamestream::stream::tests::sender_delivers_batches, a UDP-loopback EINTR under qemu that fails identically on a pristine HEAD).	2026-08-07 10:50:32 +02:00
enricobuehler	3608de25ed	Merge pull request 'fix(ci): runner hygiene stops eating its own jobs' (#84) from worktree-ci-runner-hygiene into main ... Reviewed-on: #84	2026-08-07 08:37:42 +00:00
enricobuehler	75dfab1d35	fix(host): a reconnecting session inherits its launch instead of starting it again ... Two defects, found while tracing M8's codec-fallback reconnect and recorded verbatim in d5e23146 as out of scope there. A client retry re-sends Hello::launch verbatim, and the host launched unconditionally. Steam and Epic URIs hide it — the launcher focuses the running copy — but a gog:/custom: target really did start a SECOND COPY of the game. The client cannot fix it by dropping the field: on Linux the per-session gamescope is re-adopted through pf-vdisplay's display registry, whose reuse key includes the launch command, so a retry without it orphans the running game. And the retry minted a fresh launch_stamp, so procscan refused to adopt a game started more than 2 s before it — the game was minutes old, so a reconnected session had no game-exit detection for the rest of its life. Both are now answered by a launch registry (launchreg.rs): one record per (client fingerprint, library id), written at launch time and INDEPENDENT OF THE TERMINATION POLICY. That independence is the point. The existing fingerprint-keyed reclaim only exists under GameOnSessionEnd::Always — under the default Keep, arm_grace is never called, so nothing was recorded at all in exactly the configuration the defect was reported in. The design correction that matters: at launch time the host knows NOTHING about the game's processes — that is the premise of the whole lease design. So identity flows BACKWARDS from the watcher, which publishes the concrete ProcRefs it adopted, and the registry's liveness is Scanner::alive over that recorded set, re-verified by (pid, start). Never a re-scan by spec: a later scan would find a copy the player started since, and adopting that is what procscan's rule 1 forbids. The published set is never cleared on exit either — the last thing the watcher saw is what makes a quit game read Gone rather than "no opinion", which is how it becomes relaunchable at once instead of being suppressed for the window. On rule 1: an adopting session inherits the older floor, so its own find() admits what the ORIGINAL session's lease already admitted for its whole life. That is the correct reading of "the same launch, continued" and not a new exposure — rule 1 forbids adopting processes that PREDATE the launch, and these postdate it. The match rule is pure and total (covers()): liveness is authoritative where it has an opinion, and only Unknown falls through to the tie-breakers — a live holder, or a 90 s in-flight window for a re-dial while the launcher is still working. Gone beats both, deliberately: a title that crashed on startup must relaunch at once. Both race orders are handled and neither is relied on. Teardown-first takes the Running arm; handshake-first (a fast re-dial on a half-open connection) takes the holders>0 arm, and the old teardown then sees superseded() and does nothing — without which, under Always, it would arm a grace the new session had already passed its chance to reprieve, and the reaper would kill the new session's game. Two tradeoffs taken deliberately: a custom: command with no detection hints stays Unknown forever, so that reconnect trades game-exit detection for not double-spawning; and IN_FLIGHT_WINDOW is a fixed 90 s rather than sharing disconnect_grace_seconds, because the two have opposite failure costs — grace being wrong leaves a game running, this being wrong silently swallows a launch the player asked for. Gates: fmt clean; clippy -p punktfunk-host --all-targets -D warnings green in the Linux container; 418 passed, +9 exactly the new tests. One failure, gamestream::stream::tests::sender_delivers_batches, is pre-existing and environmental — a UDP-loopback EINTR under qemu at stream.rs:1697, outside every hunk in this change (the last is at +448), and it fails identically on a pristine HEAD. I reproduced both the failure and its location myself rather than taking it on report. ⚠ OWED: the Windows leg is COMPILE-UNVERIFIED. cargo check --target x86_64-pc-windows-msvc dies in ring's C build on macOS and xcheck.sh does not cover punktfunk-host. The Windows edits are small restructures of existing branches plus a bool assignment, reasoned through but seen by no compiler. Run it on .133 before this merges. I narrowed that exposure by inspection afterwards, and it is smaller than the blanket warning suggests. The change presents exactly two things to a Windows compiler that a Linux one did not already see. launchreg gates only alive_count (lines 227/231), whose cfg(any(linux, windows)) arm calls Scanner::system().alive(procs) — the identical call gamelease.rs:563 already makes in code that compiles on Windows today. And the Windows launch arm at native/stream.rs:1666 reads only ungated bindings the Linux arm type-checks thirty lines below it (adopt_launch:1658, spawned_now:1663, launch_claim:1463) and calls only the pre-existing library::launch_title. No new type, no new signature, no Windows-only API. That is an argument, not a compile. The run on .133 is still owed.	2026-08-07 10:14:21 +02:00
enricobuehler	138a1f1b2f	fix(host/windows): the staging-dir SID checks document their unsafe blocks ... clippy's undocumented_unsafe_blocks (deny) flagged the three blocks that 81039581 introduced: the SAFETY comment sat outside the closure, so IsValidSid/EqualSid inside it read as undocumented, and from_raw_parts shared a comment that only covered the GetLengthSid line above it. Windows host clippy is the only leg that lints this cfg(windows) code, red since.	2026-08-07 10:11:04 +02:00
enricobuehler	39cfb7234c	fix(ci/docker): a cache-hit builders job stops failing on a login it never uses ... The LAN-registry docker login only serves the Push step (Reconcile and Tag-for-release authenticate via curl -u), but it ran unguarded — so a hit=true leg landing on a host with a misconfigured docker daemon failed at login with nothing to push (run 16044/16013 f44 leg). Gate it like Build/Push.	2026-08-07 10:11:02 +02:00
enricobuehler	c3cdee9bf5	fix(ci/prune): the 2-minute image prune stops deleting images mid-pull ... docker image prune -af --filter until=2h keyed on image CREATION time, so a base image built days ago that merely had no container at that instant was "aged" — including one a job had just pulled and not yet created. Measured 2026-08-07: three job failures, each coinciding with a prune tick to the second ("No such image: …punktfunk-rust-ci:latest", every step cancelled), plus a 4-7 GB re-pull of every idle base image within minutes. The routine tick now retires only what this host actually accretes — per-SHA app tags older than 2h (their creation time IS the local build time) — then sweeps dangling layers, which cannot touch a tagged image. The blanket -a prune survives only in the near-ENOSPC burst guard, where one re-pull beats every concurrent job dying. docker-reclaim.{sh,service,timer} are the hourly leak reclaimer that so far lived hand-installed on home-runner-1 only; home-runner-2 went without it and re-accumulated 176 leaked volumes (~60 GB) until jobs died of ENOSPC on 2026-08-06/07. Checked in so both hosts install the same files from here.	2026-08-07 10:10:59 +02:00
enricobuehler	dee97e893c	fix(vkdecode): the address the driver keeps is now the address we keep ... The AV1 use-after-free fix (cdd1f3ef) stabilised the wrong half. NVIDIA was measured retaining pColorConfig, so StoredParamsAv1 boxed the colour and timing blocks — but OwnedStdAv1SequenceHeader kept the Std struct ITSELF inline, so the pStdSequenceHeader we handed vkCreateVideoSessionParametersKHR was a stack address inside ensure_parameters, dead the moment it returned. The fix worked because of WHICH pointer that driver happened to hold. A driver retaining the outer one instead — no more of a spec violation than retaining pColorConfig was — reproduces the original bug exactly: plausible pictures, wrong content, no error and no counter moved. The same shape was in the shipping codecs, one step further from evidence: the H.264 and H.265 create paths pointed pStdSPSs/pStdPPSs/pStdVPSs at function-local Vecs, and both Add paths handed over the wrapper's inline std field and then moved the wrapper. Those are spec-legal — the object stores copies — and have never misbehaved on the fleet. They are fixed anyway, because that is precisely what was true of H.264/H.265 before the same class of bug was found in them, and a correctness argument that reduces to which vendor we tested is not one. So: the Std struct is boxed inside each owning wrapper (one level out from what _color_backing already did), and the contiguous create-time arrays are now fields of the stored parameters, assembled at their final address. Identical bytes at identical offsets — only where they live changed. The line drawn deliberately, in prose at session.rs:29: Std DATA is pinned; the VkVideoSessionParametersCreateInfoKHR chain itself is not. Retention there would be a different and far more extreme class of driver bug, and pinning it needs a self-referential struct over lifetime-parameterised builders. ⚠ NOT hardware-verified. No GPU has run this — the fleet is unreachable and the 250/250 parity that proved this code bit-exact cannot be re-run. That is why the change is constrained to address stability alone, and why it ships five CPU-only tests instead: three capture the pointer handed to Vulkan, perform the real move, and assert it survives — each verified FAILING first, with genuinely differing addresses, not a tautology. Two more pin the create-array ownership; those fail before the fix as compile errors rather than assertions, because the pre-fix bug there is a dangling pointer and asserting on it is UB. Also: caps.rs claimed the borrow checker pins a profile chain between wire() and its last use. False at exactly one site — decoder.rs took a raw *const, ending the borrow, leaving nothing but inspection to stop a future editor moving the chain before create_query_pool. Correct today, guarded by prose, which is how the first bug shipped. It is now compiler-enforced: the pointer write and the create call live inside one helper that takes the profile by reference, so the borrow is held across both by the signature. An audit cleared the chains otherwise — no entry point we pass one to retains it. Gates: fmt clean; clippy -D warnings over pf-vkdecode AND pf-client-core in the Linux container (its only real consumer, which cannot build on macOS at all — wol.rs uses deps its manifest gates to linux/windows, so workspace clippy has never passed there and does not now); 187 lib tests green on Linux, up from 182.	2026-08-07 09:57:39 +02:00
enricobuehler	93b8528d09	Merge pull request 'fix(encode): NVENC split-frame encode never engaged for HDR — engage it, measured' (#83) from worktree-nvenc-s1-split-reconfigure into main ... Reviewed-on: #83	2026-08-07 07:57:29 +00:00
enricobuehler	b5205fef52	Merge pull request 'fix(client/apple): audio stops crackling on lossy, bunching Wi-Fi' (#82) from worktree-audio-wifi-distortion into main ... Reviewed-on: #82	2026-08-07 07:37:19 +00:00
enricobuehler	515a3c2912	feat(pf-encode): wire split arbitration on Windows too ... The last coverage gap, and only worth building once S1 proved it possible: the Windows backend drives NV_ENC_DEVICE_TYPE_DIRECTX, and an in-place splitEncodeMode change had never been tested there. It works (071358cb), so the arbiter is now ungated from Linux-only to the union of both direct-SDK backends and wired into windows/nvenc.rs: the submit stamp, the feed hook on AU completion, apply_split_mode, split_key, arm_split_arbiter, and set_send_spread_us. Same gates as Linux, and they are correctness conditions rather than preferences: opt-in while it earns trust, an operator PUNKTFUNK_SPLIT_ENCODE pin always wins, a cached verdict short-circuits, >=2 engines, never H.264, and the sub-frame trade is only entered when the host has actually reported a send spread to price it with. The one Windows-specific difference is that `async_rt` is a real possibility here (opt-in two-thread retrieve) and the arbiter refuses it, because under pipelined retrieve the submit->AU span includes queue depth and the comparison would be noise. ⚠ Two more instances of the same item-level dead_code trap, caught by the Windows run and not by reasoning -- that is now 4 and 5: - `clear_split_verdicts` is called only by the Linux on-hw test, so it is dead on Windows; gated to `all(test, target_os = "linux")`. - The arbiter methods first landed inside `impl Encoder` rather than the inherent impl (the anchor I used, supports_chunked_poll, is a trait method), which the compiler caught as "not a member of trait Encoder". Verified .158 (RTX 4090 / Ada, driver 610.88, D3D11): clippy --features nvenc --all-targets -D warnings clean, and 2 on-hardware NVENC tests green including S1 re-run with the arbitration code in place (engines=2 latched, DISABLE->TWO_FORCED accepted, zero IDRs, reverse accepted). Verified .21: clippy clean with AND without the nvenc feature, 65 unit tests, 25/25 NVENC on-hardware. fmt clean.	2026-08-07 09:28:34 +02:00
enricobuehler	8c994965d4	docs(licensing): say what the licence gate cannot see ... The plan's M10 checklist named "the about.toml carve-out that puts FFmpeg outside the automated licence gate". There is no such stanza — I looked, on this branch and on origin/main. The carve-out is structural, which is worse: cargo-about walks the CARGO graph, so a native library reached through a permissively-licensed -sys crate is invisible to it. ffmpeg-sys-next is WTFPL and passes the gate cleanly while the LGPL libavcodec it link-imports is never harvested at all. So about.toml's own claim to be "exactly the regression guard we want against a copyleft dependency silently entering the linked set" was overstated: it did not catch FFmpeg entering and would not catch the next one. The comment now says so, and says where the LGPL obligations are actually discharged instead. The one genuinely good piece of news is recorded too: since M10 the client links no FFmpeg, so for every client artifact the crate graph and the linked set coincide and the gate finally means what it appears to mean. The gap is the host's alone. Gate: cargo about generate about.hbs --fail — passes.	2026-08-07 09:27:30 +02:00
enricobuehler	b27135308f	fix(scripts): the xcframework never absorbs a Homebrew libopus ... On a Mac with brew's opus installed, audiopus_sys found it via pkg-config and statically linked it into the aarch64 slice — a lib built for the RUNNING macOS (minos 26.0, tripping the script's own version guard) and existing only for the host arch, so the x86_64 slice silently fell back to the vendored build and the two slices shipped different libopus builds. Force the vendored CMake build for every slice (OPUS_NO_PKG_CONFIG=1), with the CMake policy floor modern CMake (>=4) needs to accept libopus's old cmake_minimum_required.	2026-08-07 09:23:37 +02:00
enricobuehler	64c92da356	fix(client/apple): the jitter ring deepens on the sessions that actually starve ... The shared JitterPolicy grew an adaptive target floor — clustered genuine underruns raise the live target a step at a time up to max_target_ms, a long quiet spell relaxes it back — and the three Rust rings all run it via note_read. The Apple ring is the one hand-written mirror, and it mirrored the shed half but not the growth half: its target was pinned at the 20 ms base forever. On Wi-Fi that bunches arrivals (power-save is the classic; the field MacBook report is the symptom), 20 ms is regularly shorter than one delivery stall, so the ring re-primed through every stall for the whole session — crackle that never got better, on exactly the client where a Moonlight with a deeper buffer sounds fine on the same host and network. The ring now carries the full mirror of note_read: 3 underruns inside a 5 s window grow the target 10 ms (capped at COREAUDIO's 70), 30 s of quiet gives a step back, and the write-side hard trim follows the grown target (including the Rust policy's target+quantum guard, which the mirror also lacked). New tests pin the mirror to the Rust suite's expectations — growth, relax, the cap — plus the field scenario end to end: bunched 60 ms deliveries with every fourth burst 30 ms late converge to a silence-free tail instead of crackling forever.	2026-08-07 09:23:29 +02:00
enricobuehler	81d257c7fa	fix(client/audio): the in-core decoder conceals lost packets like every other client ... A field report: game audio on a MacBook (M1) crackles over Wi-Fi against a host that plays clean to other clients. The Apple client is the one client whose Opus decode lives in core (punktfunk_connection_next_audio_pcm — AudioToolbox has no multistream path), and that decoder only ever decoded packets that ARRIVED. The Linux, Windows and Android decode loops all feed an AudioGapTracker and synthesize libopus packet-loss concealment for every packet the wire lost; the in-core path had the tracker sitting unused in the same crate. So on Apple every lost 5 ms datagram — at ~200 packets/s over Wi-Fi, a steady trickle — landed in the playout ring as a hard time-domain gap: a click per loss, sustained crackle under real loss. The redundant-plane recovery (0xD2) hides single losses when the host grants it, which is exactly why the survivors are the burstier gaps that need concealing most. The decode now runs through the same accounting as everyone else: concealed frames land in front of the arriving frame in one contiguous buffer (the embedder just writes it to its ring), a DTX marker advances the accounting without being decoded, and the output buffer is pre-sized for a full concealment run so the borrow-until-next-call pointer can never dangle. Unit-tested against real libopus: gaps, duplicates, DTX-after-loss, and the 50 ms cap.	2026-08-07 09:23:15 +02:00
enricobuehler	071358cbf7	test(pf-encode): S1 on WINDOWS/D3D11 — passes; Windows arbitration is buildable ... Everything the split-encode programme rests on had been proven only on Linux/CUDA. The Windows backend drives NV_ENC_DEVICE_TYPE_DIRECTX, so none of it transferred by assumption -- and if the driver refused an in-place split change there, Windows arbitration would simply not be buildable. RESULT on the RTX Windows box (RTX 4090 / AD102, driver 610.88, D3D11): engines=2, latched by query_caps (WP1.1's probe, validated on Windows hardware rather than inferred from Linux) DISABLE -> TWO_FORCED via nvEncReconfigureEncoder, resetEncoder=0: ACCEPTED, ZERO IDRs, and the reverse likewise. So the foundation now holds across three platform x arch x driver combinations: Linux/CUDA Blackwell 610.57.04, Linux/CUDA Ada 610.43.03, Windows/D3D11 Ada 610.88. ⭐ UNBLOCKS ALL FUTURE WINDOWS ON-HARDWARE TESTING. pf-encode's nvenc test binaries were believed unlinkable on Windows ("NvEncodeAPICreateInstance unresolved", recorded as pre-existing and worked around by only ever running clippy there). They link fine given the SDK import library: RUSTFLAGS='-L native=C:\Users\Public\nvenc -l nvencodeapi' `-L` alone is not enough -- without a `-l` nothing pulls the archive in, which is why the earlier attempt still failed. ⚠ This is TEST-BINARY-LOCAL and must stay that way: production deliberately dlopens NVENC rather than link-loading it, and an unconditional link-load is the known crash class on non-NVIDIA Windows hosts. ⚠ Box note: the RTX Windows box answers on .158, not the .173 in its memory entry, and `Administrator@` there resets the connection right after SSH2_MSG_SERVICE_ACCEPT in a way that reads like the host being down -- the working login is "Enrico Bühler"@192.168.1.158.	2026-08-07 09:23:05 +02:00
enricobuehler	5a1ec6198e	docs(ci): the Windows host's FFmpeg tree is no longer shared with the client ... windows-host.yml called FFMPEG_DIR "the same BtbN lgpl-shared x64 tree the Windows CLIENT links against". Since M10 the client links no libav* at all and windows.yml sets no FFMPEG_DIR, so the sentence pointed a reader at a link that does not exist. The provisioning script still fetches the tree — for the host alone — which is the part worth saying out loud, because the next person to read it will wonder why a client-provisioning step still mentions FFmpeg.	2026-08-07 09:16:32 +02:00
enricobuehler	0430d907bb	fix(pf-encode): gate forced_split_width to Linux — WP4 broke the Windows build ... The verification gap flagged in 01294e3a was real. `.133` came back up and the WP4 commit failed Windows clippy: `forced_split_width` is used only by the libav NVENC path (`enc/linux/mod.rs`), but it was added to `codec.rs`, which compiles everywhere -- so it is dead code on Windows and `-D warnings` rejects it. Third time this crate has hit the same item-level dead_code trap (see `subframe_env_forced`, and the arbiter items in `nvenc_core`), and the third time it was caught by actually running the Windows check rather than by reasoning about it. The comment on the gate says so, since the pattern is clearly not self-evident from the code. Verified .21: clippy -D warnings clean both WITH and WITHOUT the nvenc feature, 65 unit tests. Verified .133: Windows clippy --features nvenc --all-targets -D warnings clean, zero errors, zero dead_code. fmt clean.	2026-08-07 09:01:05 +02:00
enricobuehler	5c05246098	feat: M10 — FFmpeg is gone from the client ... cargo tree -p punktfunk-client-session finds no ffmpeg. The host still does, which is the whole point: pf-encode keeps libavcodec unconditionally and no host workflow, packaging script or licence file was touched. Deleted: crates/pf-ffvk, video_vulkan.rs, video_vaapi.rs, video_libav.rs, the libavcodec half of video_d3d11.rs, the av_log machinery, ffmpeg::codec::Id as the decoder's vocabulary (the quic CODEC_* wire constants now serve, which is why the evidence table was keyed on them), DecodedImage::VkFrame and ::Dmabuf, the presenter's AVVkFrame lane, and the ffmpeg-fallback feature with everything behind it. DrmFrameGuard collapses from an enum to a newtype, which removes an unsafe impl Send. Roughly 25,000 lines. Then the CI, packaging, licensing and docs work the plan's §6 lists: the Windows workflows lose FFMPEG_DIR, PF_FFVK_VULKAN_INCLUDE and their PATH prepend; the MSIX loses its DLL wildcard; the client .deb stops emitting libav sonames on its own because depends come from dpkg-shlibdeps; arch, flatpak and nix drop the dependency; and the README's "FFmpeg 7 or 8" contract narrows to the host. Three defects reached users' machines in the first cut, and none was in the deletion itself. All three desktop Settings UIs offer vulkan, vaapi and d3d11va as stored decoder values, so those strings sit in shipped settings files today. Refusing them by name — which is the correct rule for a stale pin — would have bricked every upgraded client whose owner ever touched that dropdown. They now migrate onto the native rung for the same hardware family, at decoder construction AND at each dialog's lookup, because a legacy value that matches no preset displays as "Automatic" and silently rewrites the user's preference on the next save. M9's evidence filter was deleted on the argument that with no libavcodec twin below, barring an unproven rung removes hardware decode rather than moving down one rung. That is true on Windows and false on Linux for Intel and every unknown vendor id, where prefer_vulkan_first is false and the order is native-vaapi → native-vk: a rung that has decoded nothing anywhere sitting above one that is 250/250 on three drivers. Every Intel Linux desktop would have moved from libavcodec VAAPI, shipping for years, onto pf-vaadec by default — and a rung that constructs and then produces wrong pixels leaves only by the error-streak demotion, which this codebase already documents as not tripping on the B580's strobing. The filter is restored as a narrow, pure, testable rule: an unproven rung yields to a proven one, and to nothing else. Windows deliberately passes no rung below, because that vendor family is the one with a measured wrong-pixel report against Vulkan decode, and trading no evidence for evidence of corruption is the wrong direction. And the notices still said FFmpeg was bundled. The root file is what both desktop clients include_str! and what the MSIX ships, three lines under the new card saying no FFmpeg is bundled; Apple's Acknowledgements said it too, on iOS, tvOS and macOS. The generator now emits four per-client files scoped by transitive closure — 0 FFmpeg mentions in each, verified — while the root file keeps it for the host. That also ends the standing false attribution of ffmpeg-next, GTK4, windows-rs and the NVENC SDK to an iPhone. Windows has no reachable box, so it was compiled instead: a cross clippy at -D warnings on x86_64 and aarch64-pc-windows-msvc with the C toolchain stubbed so build scripts run without linking. That gate immediately caught an include_str! path one directory too deep, which nothing else could have. Gates: container clippy -D warnings, 160 tests, workspace check, both Windows targets clean, client ffmpeg count 0 and host 2. The four decode crates are untouched, so the hardware rungs' 250/250 stands. ⚠ Owed and unrun: no GPU has executed any of this milestone. M8's on-glass software check, M7's D3D11 and VAAPI AV1 hardware legs, and M9's field bake all still want hardware, and the bake window and criteria remain the user's.	2026-08-07 08:58:47 +02:00
enricobuehler	01294e3a53	refactor(pf-encode): WP4 — one split policy, shared with the libav path ... The libav NVENC path carried its own inline copy of the split decision and had already drifted from the direct-SDK selector: it hard-coded a 2-way split regardless of engine count, and had no depth rule at all. That is the drift the shared resolver was extracted to prevent, and the copy quietly reintroduced it. Routing it through `resolve_split_mode` needed the policy to MOVE. `nvenc_core` is gated on `feature = "nvenc"`, but the libav path is precisely the build where that feature is OFF (`PUNKTFUNK_NVENC_DIRECT=0`, and the featureless packages -- the packaging gap this project has been bitten by before). So resolve_split_mode / max_forced_split_mode / clamp_to_engines, plus a new `forced_split_width`, now live in `codec.rs`, which is always compiled and already owned SPLIT_FORCE_PIXEL_RATE. That means the NV_ENC_SPLIT_ENCODE_MODE values had to be hand-written as plain constants, since the SDK enum does not exist without the feature. They are therefore pinned: `nvenc_split_constants_match_the_sdk` (feature-gated, the only place both are visible at once) asserts all five against the real enum, so the copies cannot rot. ⚠ Only the FORCED outcomes are actionable on the libav side -- libavcodec's `split_encode_mode` AVOption is its own vocabulary and our DISABLE is the NVENC enum's 15, which would be meaningless there. DISABLE/AUTO both map to "leave the option unset", which is exactly today's behaviour (unset = the driver's auto). `engines = 0` ("not probed") maps to 2-way, preserving what that site always did; a 3-NVENC part gets the wider split only on the direct-SDK path, which is the one that actually probes. ⚠⚠ VERIFICATION GAP: .133 went down mid-change (no ping), so the WINDOWS leg is UNVERIFIED. This matters more than usual -- the Windows backend imported resolve_split_mode from nvenc_core and that import had to move too, which a grep caught rather than a compiler. Re-run before trusting it: cargo clippy -p pf-encode --features nvenc --all-targets -- -D warnings Verified .21: clippy -D warnings clean BOTH with and without the nvenc feature (the featureless build is the whole point of the move) and with nvenc,vulkan-encode; 65 unit tests incl. the new constant-parity test; 25/25 NVENC on-hardware; punktfunk-host clippy clean. fmt clean.	2026-08-07 08:16:35 +02:00
enricobuehler	38554c1c6e	feat(client): M9's code half — native first, FFmpeg behind an off-by-default feature ... `ffmpeg-fallback` on pf-client-core, default off on the crate. With it off the libavcodec rungs are not compiled, pf-ffvk leaves the dependency graph, and no ladder or demotion arm names them; with it on each sits exactly where it sits today, directly below its native twin. That is the switch which makes M10 a deletion rather than a redesign. The bake window and the regression criteria are the user's, per the plan, and nothing here claims the M9 gate is met. The hard part was not the feature, it was honesty. Two of the four native rungs have never decoded a frame on any hardware — native VAAPI at all, and native D3D11VA's AV1 leg — and making those the default would assert evidence that does not exist. So admission is per rung and per codec: a pair with hardware evidence joins `auto` always; a pair without it joins only when nothing proven is left below it (a build with no FFmpeg twin, where the alternative is not a proven rung but the CPU) or when the user asks with PUNKTFUNK_NATIVE_FIRST=1. Pins bypass it, so a lab run can still reach any rung. The shipping default therefore changes in exactly three ways, all evidence-backed: AV1 `auto` takes native Vulkan (250/250 bit-identical on an RTX 5070 Ti), Windows H.264/H.265 `auto` takes native D3D11VA above its FFmpeg twin (parity on two GPUs plus a 30-minute soak), and a failing Vulkan rung on Windows demotes to native D3D11VA first. Everything unproven is byte-for-byte as it was. The evidence state is written where it cannot rot: a table in video.rs's module docs, the same facts in code as `native_evidence()`, a test asserting them in both feature states, and a per-session log line carrying the rung, the codec, whether hardware has verified that pair and the evidence string — at WARN when it has not. A support engineer reading a log can now tell proven from assumed without asking anyone. Termination needed a new guarantee. With the FFmpeg twins gone, two native rungs in opposite per-vendor orders could hand a session back and forth forever, so a rung once entered is never re-entered and the walk is monotone to software. The never-delivered fall-through still works: with the feature on it is unchanged, and with it off it is redundant, because the next candidate already IS the rung below. ⚠ ffmpeg-next remains a hard dependency of pf-client-core, deliberately. What is left off-feature is three type-level residues — the codec-id vocabulary, the AVVkFrame guard that is pf-presenter's public import, and a pixel-format in one signature — every one of them an M10 §6 line item. Deleting them here would mean deleting the presenter's FFmpeg lane, 55 call sites, in a milestone whose gates cannot run a GPU. No libavcodec decoder is opened in a default build. ⚠ video_d3d11.rs was gated item by item rather than wholesale, and nothing in this tree compiles it — it needs a Windows check before anyone trusts it. Gates: both feature states, container clippy -D warnings and 158/159 tests, workspace check. The four decode crates are untouched, so the hardware rungs' 250/250 stands.	2026-08-07 07:00:29 +02:00
enricobuehler	d5e23146c0	feat(client): M8 — the software rung is openh264 and rav1d, and swscale is gone ... The ladder's last rung no longer runs FFmpeg. H.264 decodes through openh264, AV1 through rav1d, and HEVC is refused outright: no permissively licensed software HEVC decoder exists, so an HEVC session that exhausts its hardware rungs now tears down and re-dials advertising HEVC-less caps, and the host picks H.264. The plan calls that a first-class path; it is one. swscale is deleted, and with it the BT.601 default that its correction code existed to undo. Colour on the H.264 lane now comes from the same pf-bitstream planner every hardware rung submits from — openh264 reports no VUI at all — and AV1's comes per-picture from the sequence header. One colour source, one CSC: the old default is unrepresentable rather than merely fixed. Frames reach the presenter as three tightly-packed planes through the planar CSC pass, which had to be un-gated from the pyrowave feature and its device probe, since the last rung must exist on devices that failed that probe. rav1d rather than the dav1d crate, deliberately and against the plan's literal wording: dav1d-sys is system-deps-only, so it would add a system library and a .pc file to every client package — in the milestone family whose excision checklist exists to delete exactly that. rav1d is the same decoder, same licence, statically linked. The cost is honest: no-asm builds on both decoders, and software throughput is still unmeasured. The colour test is the milestone's exit criterion, so it is built to fail. Three fixtures, and a mutation check: hardcoding the swscale default turns the red bar to [255,24,0], and swapping Cb/Cr turns red to blue — a silent error no metadata assertion could catch. Review then disproved the range half of it numerically: with eight saturated bars, decoding the full-range fixture with the wrong range gives max error ZERO, because a mismatch only pushes values outside [0,1] where the shader clamps. A mid-tone was added; the wrong range now costs 11, well past the tolerance. The exit criterion I set was satisfiable by a test that proved nothing. Two blocking defects, both emergent rather than local. Software AV1 on a 10-bit stream never reached its typed refusal: rav1d is built 8-bit-only and returns ENOPROTOOPT, which the send loop turned into a generic error, so the pump's typed downcast missed and every AU failed identically — a permanent freeze on precisely the shipping case, since AV1 is advertised only where hardware AV1 exists and hardware AV1 plus HDR is Main 10. The shape is now read from the sequence header before any byte reaches the decoder, exactly as the H.264 leg reads the active SPS. And the new Reconnecting phase was the first state that is not streaming, not connecting, and still holding a live stream — which opened all three guards that had made a second launch impossible. Pressing A assigned over `stream` where every other site shuts down first, and StreamState has no Drop, so the old pump was detached: a second live session still submitting to a Vulkan device that gets destroyed underneath it. Nothing about the reconnect was wrong in isolation; the defect lived between a new state and three guards nobody re-examined. Start is now defensive and the retry raises the connecting modal, so the UI matches the state and B can cancel. Also closed: retry_caps was computed, tested and never applied, so a shape refusal could end a session reporting no codec available while a working retry existed; the retry inherited force_software sticky-true, landing an HEVC→H.264 fallback on software H.264 with working hardware H.264; it re-dialled with a stale mode; the CPU present arm had no survivable-failure handling where the pyrowave arm — same pass — has it; HEVC is no longer advertised when the decoder is pinned to software; and the software rung now feeds the recovery-point SEI it already had in hand to the re-anchor gate. ⚠ Two host-side gaps found while tracing, neither in scope here: Hello::launch is NOT idempotent (gog:/custom: targets spawn a second copy on a retry; the field is kept verbatim because dropping it orphans the gamescope display whose reuse key includes the command), and a reconnected session can never adopt a game predating its own launch stamp, so it has no game-exit detection. ⚠ OWED: the on-glass software run. ~200 lines of new Vulkan on a path that only runs because the GPU already failed, and no driver has seen it. The review's minimum check is sync validation enabled, a non-multiple-of-16 mode, a mid-session resize and demotion, and both colour matrices. Gates: container clippy -D warnings over four crates, 236 tests, workspace check. pf-vkdecode and pf-bitstream are byte-for-byte untouched, so the hardware rungs' 250/250 stands.	2026-08-07 06:18:23 +02:00
enricobuehler	a20cd44ed4	feat(client): native VAAPI AV1 — the third rung, and two failure-path defects ... The libva AV1 layouts, the AuPlan conversion and the Linux rung's AV1 arm, completing AV1 across all three hardware backends. Pin-only. Layouts measured, not transcribed: the committed probe grew the AV1 structures and every size and offset it printed against libva 2.23.0 is a compile-time assertion. Three that a hand-count gets wrong — the picture buffer is align 8 because anchor_frames_list is a pointer, inserting seven bytes of padding; seg_info and film_grain_info carry their own padding tails inside the parent; and THREE of AV1's six bit-field unions are narrower than a word (one uint8_t, two uint16_t), so a u32 packer over any of them writes through its neighbour. This is the fifth way this program has had to spell "which pictures does this frame use", and it is unlike the other four: ref_frame_map is indexed by SLOT and holds actual VASurfaceIDs rather than indices into anything, ref_frame_idx is indexed by NAME and holds slots taken from the header — not from the plan's refs, where a lost reference leaves a hole and a hole is not a slot — global motion is picture-level, and there is no per-reference size field at all. Established from va_dec_av1.h and libavcodec's vaapi_av1.c, and stated in the module docs so the next reader does not re-derive it. Review verified the whole happy path — every layout assertion re-measured, every packer width and bit position, the reference convention, the num_elements buffer shape — and found both defects on FAILURE paths, neither reachable on the vendored vector. A conversion refusal permanently desynced the ledger. The mutation block sat after the tile walk, so any tile-shape refusal left the planner holding a picture with no ledger slot — and the resulting UnresolvedReference fires before that block too, so it never repaired. Every later access unit hard-errored until a shown key frame: one lost packet costing a GOP. The arm's own doc already warned that skipping conversion would desynchronise the slot map; the refusal door did exactly what the skip door was written to avoid. The block is hoisted, and a tile-shape refusal on an already-damaged plan is now concealed rather than refused. Fixing that exposed a sharper edge: the conversion can release a slot and reassign it to the refused picture in one call, so the binding would still hold the PREVIOUS picture's surface — a wrong reference rather than a missing one, which nothing downstream could notice. The caller now clears the binding unconditionally on the refusal path. And a damaged frame's surface was never written yet was bound as a reference and left in pending, so a later clean show_existing_frame would claim it with damaged = false and ship uninitialised GPU memory to the presenter — on several drivers another client's framebuffer. The justification quoted half of va_dec_av1.h; its next sentence gives the remedy, which is to point the problematic index at an alternative buffer. Damaged frames now submit as they do on the other two arms, with live surfaces substituted for invalid entries and reported as a bitmask — preferring a reference that really decoded over the decode target, and keeping libavcodec's deliberate all-invalid map on a shown key frame. Film grain is refused rather than decoded wrong: libva wants two surfaces, one ungrained for prediction and one grained for output, and libavcodec allocates a second frame for exactly that. The gate now sits after the mutation block so a grained frame costs itself rather than the GOP, and stays per-AU rather than per-sequence because a stream that merely DECLARES the tool decodes here perfectly. ⚠ Residual, flagged not fixed: a picture decoded from substituted references can still be shown by a later show_existing_frame. It is decoded memory now rather than uninitialised, and it is what the H.264/H.265 arms do, but tracking "this was concealed" through to display needs new session state. Gates: macOS fmt/clippy/125 tests/cargo-doc, container clippy -D warnings over seven crates and 548 tests, workspace check. pf-bitstream's diff is comment-only — verified — so the Vulkan rung's 250/250 stands untouched. Nothing here has decoded a frame: no VAAPI hardware is reachable.	2026-08-07 04:22:23 +02:00
enricobuehler	ef40890c80	feat(client): native D3D11VA AV1 — wired, and four defects it exposed ... The AV1 arm of the native D3D11VA rung, parity-required because today's FFmpeg d3d11va rung already decodes AV1 Profile 0 and the excision must not silently drop it. Pin-only, as that rung is today. decode() walks the temporal unit frame by frame; submit() splits into decode_into and present, because AV1 decodes frames that are never shown. The proven H.264/H.265 body is byte-for-byte unchanged — review diffed it against HEAD mechanically and found only a rename plus one refusal arm — and the VideoProcessorBlt hand-off is untouched. That mattered more than anything else here: those two codecs are hardware-proven, .173 is powered off, and no gate that runs could have caught a regression in them. Every descriptor value comes from libavcodec's dxva2_av1.c read verbatim, not from symmetry with the other codecs: three buffers and no qmatrix (AV1 transmits none), NumMBsInBuffer zero on all three, ConfigBitstreamRaw 1, surface alignment 128, pool +8, and the session sized from the SEQUENCE header's max frame size — sizing from the frame would rebuild the decoder and drop every reference the first time a stream legally resized downward. Two places where following the H.264/HEVC pattern would have been wrong. libav pads the bitstream buffer and grows only its descriptor's DataSize, never a tile's, because a tile's size is exact — charging padding to the last record is corruption, not filler. And the committed tile records were one per tile GROUP spanning the whole OBU, header and frame header included, where libav emits one per TILE addressing the payload past its tile_size_minus_1; the vendored vector is single-tile, so the old tests passed either way. Review then found four more defects in the already-committed conversion, each confirmed against libavcodec AND Chromium's D3D11 AV1 accelerator: Tile widths and heights were the coded minus-1 where the field is a superblock COUNT — every tile declared one superblock short, on every frame, with a comment asserting the opposite of the truth. StatusReportFeedbackNumber must be zero for AV1. Both reference implementations disable it specifically for this codec — libav's note reads "breaks decoding on some drivers (tested on NVIDIA 457.09)", Chromium's "it crashes :|" — while both set it for H.264 and HEVC, which is why this rung's proven codecs never showed it. It would likely have presented as a hang or a rejected submission rather than bad pixels, sending the next session after the tile records instead. frame_refs[].Index is an index INTO RefFrameMapTextureIndex, not a surface index; the neighbouring line already filled that map correctly. Measured: 1636 reference entries on the vendored vector where the two differ. qm_y/u/v need the 0xFF "no matrix" sentinel — 0 is a valid matrix index, and 274 of 274 frames transmit no quantiser matrix, so every one was being dequantized against matrix 0. Also closed: the slot leak the Vulkan rung had already found and documented (a frame refreshing no slot is never reported removed, so nine of them exhaust the ledger); a tile-grid check that could not fire, replaced with libav's own cols*rows guard; per-reference sizes now taken from the reference's own header via RefState rather than the current frame's; and the render size clamped against the decoded picture in both rungs, since AV1 permits a render size larger than the frame. The parity leg was rewired through the real decode path — it previously called the internals directly, so its hidden-frame assertion described the harness's own counter rather than production withholding anything. Gates: macOS fmt/clippy/383 tests, container clippy -D warnings over four crates and 499 tests, and on Windows .133 (.173 is powered off) clean checks plus 97 pf-dxvadec tests. All 8 Vulkan gpu_parity legs re-verified bit-exact on the RTX 5070 Ti after the shared-code change. No AV1 frame has been decoded through this rung anywhere: it needs .173 back.	2026-08-07 03:01:49 +02:00
enricobuehler	185332a866	fix(vkdecode): H.264 and H.265 session parameters own what the driver keeps ... The same use-after-free the AV1 rung was just fixed for, closed in the two rungs that ship. session.rs and session_h265.rs handed their Std parameter sets to vkCreateVideoSessionParametersKHR and dropped the backings when the call returned; NVIDIA 610.57.04 was measured retaining such a pointer to decode-record time, which is what made AV1 diverge on 250 of 250 frames. Nothing was known to be broken here — both rungs are bit-exact on four drivers — but that was luck rather than correctness: the freed blocks happen to still hold the right bytes in that window. The native Vulkan rung sits in the auto ladder above FFmpeg-Vulkan on shipping clients, so this was live code, and its failure mode is silent wrong pixels rather than a crash. StoredParams and StoredParamsH265 hold the parameters object together with every wrapper it points at, so an object whose backing is gone cannot be built. create_parameters_object takes the wrappers by value; the Add arm adopts them only after a successful update, so a failed update drops what it never stored; the Recreate arm replaces, destroys the old object, then drops its backings, written explicitly so the ordering survives later edits. The Add-vs-Recreate decision table and the VPS ledger are untouched — only ownership moved. params.rs still carried the refuted claim as a type-level contract, that Vulkan "copies all parameter data before returning" and keeping the wrapper alive across the call "is the whole obligation". Corrected to the measured truth. The tests are what stop this returning, and each was verified by sabotage: inlining the H.264 PPS box fails at pps pScalingLists, inlining the H.265 SPS DPB box fails at sps pDecPicBufMgr, and making either adopt drop instead of store fails both session tests. Two lessons are recorded in them. Pointer equality cannot be the assertion, because the Std struct carries pointers by value and a stale one compares equal — the read-back is the discriminator, so the tests clobber the dead stack first to make a dangling read deterministic rather than lucky. And the first H.265 draft read six of eight pointers and let the sabotage through, so it now reads every one with a labelled assert. ⚠ One site of this class remains, deliberately: the VkVideoProfileInfoKHR chains, where wire()'s borrow dies with its enclosing block while the object created from it lives on — three session creates, an image, a buffer, and a query pool built from a raw pointer into a stack chain. It spans six modules and all three codecs, and a profile is enums a driver resolves at create time with no per-frame deref, so the risk is materially lower. It wants its own pass with its own hardware verification. Gates: macOS fmt/clippy/196 tests, container clippy -D warnings, pf-vkdecode 182/182 and pf-client-core 140/140. On the RTX 5070 Ti, all 8 gpu_parity legs re-verified green after the change — H.264, H.265, Main 10 and AV1 all still bit-identical to libavcodec.	2026-08-07 01:04:17 +02:00
enricobuehler	cdd1f3efce	fix(vkdecode): AV1 is bit-exact — the bug was a use-after-free, not the driver ... 250/250 frames bit-identical to libavcodec on NVIDIA 610.57.04, and all four other parity legs (H.264, H.265, Main 10, both four-byte-prefix twins) still green. session_av1 built the sequence header, handed pStdSequenceHeader to vkCreateVideoSessionParametersKHR, and dropped the backing the instant the call returned — on the documented assumption that Vulkan copies parameter data before returning. NVIDIA does not. It keeps the pointer and dereferences pColorConfig when a decode is RECORDED. The freed block became our own next allocation, whose bytes read back as mono_chrome = 1, and a monochrome frame skips exactly loop_filter_level[2..3] (AV1 7.14). That is the whole fingerprint two earlier rounds chased: luma bit-exact, chroma off by small amounts, and rewriting the chroma levels in the bitstream changing nothing — the driver read them correctly and then discarded them, because it believed the stream had no chroma. StoredParamsAv1 now holds the parameters object and its Std backing in one value, so an object whose backing is gone is unrepresentable. The road there is worth recording, because two well-evidenced conclusions were wrong before this one was right. A software oracle reproduced the divergence exactly by disabling chroma deblocking, and a GPU probe showed chroma levels [8,12] and [63,63] producing byte-identical output — which looked conclusive and was not. libavcodec's own Vulkan AV1 hwaccel is bit-exact on this same driver, which proved the hardware fine and the defect ours. ffmpeg never hits it: with VK_KHR_video_maintenance2 it uses inline session parameters and never creates a parameters object at all. The proof is direct rather than inferred: a throwaway Vulkan capture layer dumped both submissions and every byte of our AV1 picture info already matched libavcodec's, including the loop filter block; only the session parameters layer differed. Watching the block's address showed correct bytes at create and our next allocation at decode. Ruled out on hardware, so nobody re-tests them: filmGrainSupport, maxCodedExtent, maxDpbSlots/maxActiveReferences, VkVideoDecodeUsageInfoKHR, the tile-start sentinel, the setup slot's SavedOrderHints, a NULL pTimingInfo, and heap luck. Two earlier fixes are confirmed against libavcodec's captured wire bytes and kept: CDEF secondary strengths carry the coded value rather than the spec's in-place fixup, and LoopRestorationSize is log2-based. The refuted driver-ignores-chroma-levels claim is corrected everywhere it was written down, and that probe test now passes and points at the lifetime of everything a submission points at before blaming a vendor. ⚠ Adjacent and NOT fixed: session.rs and session_h265.rs drop their Std backings the same way, and those sets carry embedded pointers too. Both are measured bit-exact on four drivers, so nothing is known to be wrong — but the contract now rests on a driver behaviour measured FALSE for AV1 on a shipping driver. The SAFETY comments asserting it have been corrected; the structure is deliberately untouched pending its own pass. Gates: macOS fmt/clippy/336 tests, container clippy -D warnings, all green; 8/8 gpu_parity and 3/3 gpu_smoke legs verified on the RTX 5070 Ti.	2026-08-07 00:39:34 +02:00
enricobuehler	1062aa780f	test(pf-encode): measure the bits/frame curve — no crossover, split always wins ... WP0's real deliverable, and the hole every previous measurement in this programme had. All prior timings ran against driver-zeroed buffers, so rate control had nothing to code (~300 B/AU against an 833 KB quota) and only the PIXEL-proportional half of the encode cost was ever exercised -- while the 4K60 HDR field report was a BITS/FRAME problem at 6.8 Mbit/frame. Adds `pf_zerocopy::cuda::write_plane_from_host`, the exact mirror of the existing read_plane_to_host. No new loader entry was needed: cuMemcpy2DAsync_v2 was already in the table and CUDA_MEMCPY2D just needed the reverse memory types. Linux-only by construction (pf-zerocopy's `imp` is cfg'd to linux). ⚠ Two harness mistakes found and fixed by looking at bytes/AU rather than trusting the knob: - Pure per-pixel noise is INCOMPRESSIBLE, so a low bitrate target does not produce low bits/frame -- it OVERSHOOTS. At a nominal 50 Mbps the encoder emitted 719 KB/AU against a 104 KB quota, and the three lowest rows of the first sweep all sat at the same ~5.7 Mbit/frame. Sweeping nominal bitrate measures nothing. - So the sweep moves CONTENT DETAIL (block size) instead, and the x-axis is the bits/frame the encoder ACTUALLY produced, never the one requested. 4K60 HEVC 8-bit, real content, single-engine vs forced-2: bits/frame Ada 4090 Blackwell 5070 Ti 0.2-0.3 Mb 4567 -> 2381 1.92x 5549 -> 3552 1.56x ~1.1-1.2 Mb 5060 -> 2626 1.93x 5867 -> 4082 1.44x ~3.3 Mb 8478 -> 4455 1.90x 9286 -> 5862 1.58x ~9.6 Mb 16237 -> 8114 2.00x 16435 -> 9275 1.77x RESULTS. (1) Encode time scales strongly with bits/frame -- 4.6 ms to 16.2 ms across the range on Ada -- confirming the hypothesis' core claim. (2) There is NO CROSSOVER: split wins at every point on both architectures (Ada ~1.9-2.0x and notably flat, Blackwell 1.44-1.77x). So the arbitration's encode-side answer is essentially always "split", which makes the sub-frame handicap the only decision that actually matters -- exactly the part already built and unit-pinned. (3) It corroborates the field capture: at ~6.8 Mbit/frame these curves put single-engine 4K60 around 10-13 ms, and the field report was 10.3 ms on a 4090. That reads as real ASIC time, not the retrieve-queue inflation it might have been. ⚠ Caveat the data itself shows: cost is NOT monotonic in bits/frame alone. The 1px row lands at the HIGHEST bits/frame yet encodes FASTER than the 4px row on both boxes (Ada 10148 vs 16237 us) -- pure noise defeats motion estimation, which gives up early, where semi-structured content makes it search hard. Content structure is a real term, so "bits/frame" is a good axis but not a complete cost model. Verified .21: clippy -D warnings clean (pf-encode + pf-zerocopy), 64 unit tests, 25/25 NVENC on-hardware. Curves run on both Ada and Blackwell. fmt clean.	2026-08-07 00:30:09 +02:00
enricobuehler	50b3fd1012	fix(pf-encode): drop the 10-bit short circuit — measured wrong on Ada, twice ... WP1.3, and the measurement that justifies it. `resolve_split_mode`'s 10-bit rule sat ABOVE the pixel-rate arm and took no codec, so it (D1) vetoed 10-bit 4K120 -- the very case the pixel-rate arm exists for -- and (D2) applied an HEVC-Main10-on- Ada result to AV1 10-bit, which has no such measurement. Both fixed: the pixel-rate arm now comes first, and what remains is codec-scoped to HEVC and only applies BELOW that bar, where a second engine buys nothing anyway. The rule rested on one datapoint: 5120x1440@240 Main10 on Ada, forced-2 7.6 ms vs 2.8 ms single-engine -- split 2.7x SLOWER. Dropping the short circuit flips that exact configuration's behaviour, so it was re-measured on a 4090 (AD102, driver 610.43.03), 400 Mbps, sub-frame pinned off, via a new mode-parameterizable Main10 A/B test (PF_AB_MODE=WxHxFPS reproduces the original operating point). Ada 4090 single forced-2 ratio 3840x2160@60 4483 us 2178 us 2.06x split WINS 5120x1440@240 3689 us 2813 us 1.31x split WINS <- the veto's origin 3840x2160@120 4148 us 2189 us 1.89x split WINS Blackwell 5070 Ti 3840x2160@60 4216 us 2477 us 1.70x split WINS 5120x1440@240 4651 us 3894 us 1.19x split WINS Split wins for Main10 at every mode on BOTH architectures, including the config the veto came from. The original number does not reproduce. ⚠ Caveats, unchanged from the rest of this work: content is trivial (297-300 B/AU against an 833 KB CBR quota -- zeroed VRAM), so this is the pixel-proportional term and the bits/frame regime is still unmeasured; debug build; and the driver differs from whenever the original was taken. Also validated on Ada in the same session -- the whole spike set reproduces on a SECOND architecture and an OLDER driver (610.43.03 vs 610.57.04): S1a in-place split switch accepted with zero IDRs both directions; S1b takes effect (|C-B|=12 vs |C-A|=1921, the cleanest run yet); S1c pair flip passes; D5 confirmed (AUTO+sub-frame 4424 vs DISABLE 4409, 15 us apart -- and AUTO without sub-frame 2310 ~= TWO_FORCED 2314, so the arm stays); engines=2 with THREE_FORCED correctly clamped to mode 2; arbitration converged with exactly 1 keyframe. Verified: .21 clippy -D warnings clean + 64 unit tests; .133 Windows clippy -D warnings clean (the resolver signature grew a `codec` param, so both backends moved); Ada + Blackwell on-hardware as above. fmt clean.	2026-08-06 23:58:24 +02:00
enricobuehler	2366c4fe31	feat(pf-encode,host): price the HEVC sub-frame trade so arbitration can cover it ... The named next step after WP3's first increment. That increment deliberately REFUSED to arbitrate HEVC-with-sub-frame -- the fleet default, and the reported field case -- because engaging split there gives up sub-frame readback, whose whole value is that the send overlaps the encode. An encoder measuring only encode time would see split as ~2x faster, take it, and make end-to-end latency worse while reporting a win. This supplies the missing number. The real comparison is encode_1eng + send_of_last_slice against encode_2eng + send_of_whole_AU, so the challenger owes roughly spread x (slices-1)/slices. Split across the two sides that can each see half: - Host: new `Encoder::set_send_spread_us` (defaulted, forwarded by TrackedEncoder -- same trap class as set_wire_chunking, and unforwarded it would fail SILENTLY IN THE SAFE DIRECTION, which is the hardest kind to notice). The send thread is the only place a paced send is observed and the encode loop the only place the encoder can be touched, so it goes over an AtomicU32 like encoder_ceiling_kbps, EWMA-smoothed 3:1 per completed AU: one content spike must not flip a verdict that then gets cached. - Encoder: turns the raw spread into the handicap, because only it knows `slices`. SplitArbiter::with_handicap charges it to the challenger before the comparison. A unit test runs identical encode numbers with a cheap and an expensive send and asserts the verdict REVERSES -- with an expensive send the arm that looks twice as fast is a loss end to end, and the incumbent must hold. That is precisely the regression an encode-only arbiter ships. Gate now opens for HEVC+sub-frame only when a spread has actually been reported (and slices >= 2); with no hint it still refuses, so behaviour is unchanged until the host feeds it. Two mechanics this needed: - apply_split_mode became a PAIR flip (split + sub-frame), routed through resolve_split_subframe and restoring from `subframe_opened_with` so a session that never had sub-frame can never gain it. It also recomputes `subframe_chunks`, which reconfigure_bitrate does NOT -- spike S1c's finding; leave it stale and supports_chunked_poll keeps saying yes while numSlices never advances, so poll_chunk busy-polls its whole budget every AU. - The arbiter is now fed from BOTH completion points. A sub-frame session finishes through poll_chunk, so the incumbent arm of an HEVC experiment would otherwise never deliver a sample -- only the challenger, with sub-frame dropped, comes through poll. Verified .21: clippy -D warnings clean for pf-encode AND punktfunk-host with nvenc, 63 unit tests (1 new), 23/23 NVENC on-hardware green. Verified .133: Windows clippy -D warnings clean, zero dead_code. fmt clean.	2026-08-06 23:41:40 +02:00
enricobuehler	f34f311c69	Merge pull request 'fix(brand): capitalize "Punktfunk" in user-facing text; name the legal copyright holder' (#81) from worktree-docs-title-casing into main ... Reviewed-on: #81	2026-08-06 21:33:33 +00:00