punktfunk/docs/releases/v0.19.0.md

Wire-compatible with 0.18.x — existing pairings and clients keep working. The new stylus plane is capability-negotiated (`HOST_CAP_PEN`) and rides an additive datagram that older peers never send and never have to parse, so a 0.18 host and a 0.19 client (or the reverse) pair exactly as before and simply fall back to pen-as-touch. WIRE_VERSION stays **2**; the embeddable C ABI moves to **13** for one new entry point (`punktfunk_connection_send_pen`); the Windows display-driver protocol is unchanged at **6** (compat floor **3**).

## Pen, stylus & tablet — a whole new input plane

Punktfunk streams now carry a first-class, pressure-sensitive stylus, end to end and on every platform. It is its own wire plane — not mouse events in disguise — so pressure, tilt, hover and barrel buttons survive the trip from glass to host.

- **The wire (P0).** A state-full `RICH_PEN` datagram carries batches of up to 8 samples — sub-pixel `f32` position, `u16` pressure, hover distance, tilt + azimuth, an eraser tool, and two barrel buttons — ordered by a wrapping sequence number and dropped whole when stale, so a late batch never rewrites the present. A `PenTracker` coalesces the stream and `HOST_CAP_PEN` advertises the plane; toward a host without the bit the client keeps its pen-as-touch fallback.
- **Linux host (P1).** A per-session uinput virtual graphics tablet injects the full stylus — pressure and tilt included — and tears down with the session. `HOST_CAP_PEN` goes live wherever uinput is available.
- **GameStream / Moonlight (P2).** The host ingests Moonlight's `SS_PEN`/`SS_TOUCH` pointer packets and advertises the matching `featureFlags`, so a capable Moonlight client sends real stylus onto the same plane.
- **Windows host (P3).** `PT_PEN` + `PT_TOUCH` synthetic pointer injection drives the Windows pen stack directly — pressure and barrel buttons reach applications as genuine pen input.
- **iPad (P4).** Apple Pencil (including Pencil Pro) capture maps onto the stylus plane — pressure and tilt from the panel, with the cursor-capability fix so the pencil and the pointer coexist.
- **Android (P5).** Active-stylus capture (S Pen and friends) forwards onto the pen plane with pressure and tilt.

## Touch injection, fixed

- **Windows native touch actually injects now.** `PT_TOUCH` pointer ids must be contiguous injector slots, but the host was passing the raw wire touch ids straight through — Windows rejected them **silently** and Moonlight-native multitouch landed nothing. Wire ids are now compacted into slots, and the first rejection is surfaced with a warning instead of vanishing.
- **NaN pressure no longer drops a finger.** VoidLink's finger touches arrive with a NaN `pressureOrDistance`; the GameStream path discarded the whole packet. It now tolerates NaN and injects the touch.

## Cursor & pointer polish

- **The forwarded cursor is scaled to the video fit** on the SDL/Linux and Apple/macOS clients. A high-DPI host pointer was drawn against the client's own backing scale and came out roughly 2× too large; it now renders true-size against the streamed video.

## Windows host fixes

- **No more `0x57` on display-config isolate.** When a supplied CCD path is doomed (a Steam Deck's live-deactivate always is), the isolate escalates to a keep-only supplied config instead of failing the whole apply with `0x57`.
- **Cursor exclusion is reported adapter-wide.** The IddCx declare's cursor exclusion is not a per-target property; the virtual-display driver now reports `cursor_excluded` across the adapter.
- **Secure-desktop cursor stood down.** The IddCx hardware cursor is held down while the Windows secure desktop (UAC / Winlogon) is up, ending the duplicated/stuck pointer there.

## Gamepad

- **A virtual Steam Deck pad no longer degrades to a DualSense because of an unrelated Steam controller.** To stop Steam Input from double-driving two *identical* controllers, the host downgrades a requested virtual Steam pad to a DualSense when a matching physical Valve controller is present — but the gate matched *any* `28DE` (Valve) device. So plugging a physical Steam Controller 2 (`28DE:1302`) into the host dropped a client's virtual Steam Deck (`28DE:1205`) to the wrong pad — the passthrough came up as a DualSense. The gate now keys on the exact VID+PID: distinct Steam controllers coexist (Steam Input drives them side by side fine), and only a genuine same-identity duplicate — a physical Deck alongside a virtual Deck — still degrades.

## Android input — mouse & keyboard regressions

Two regressions from 0.18.0's mouse-&-keyboard overhaul, felt most on Android TV boxes:

- **Mouse back/forward stays in the stream.** A mouse's back/forward buttons were synthesized by the reader as `SOURCE_MOUSE` key events and leaked into Android navigation (yanking you out of the stream on a TV); they're now swallowed while streaming.
- **Hardware typing no longer pops the IME.** The soft keyboard is gated on `imeShown`, so typing on a physical keyboard against a text-input host doesn't summon the on-screen IME.

## Release & CI

- **Release notes now ship with the release.** Notes are authored in-repo at `docs/releases/vX.Y.Z.md` and seeded into the release body at creation, and stable releases announce to the Discord `#releases` channel. This file is the first of them.
- **iOS `.ipa` attached.** The iOS build exports an App Store-signed `.ipa` to the unified release and to the run artifacts (archival/TestFlight, not a direct sideload).

## Versions

WIRE_VERSION **2** (unchanged — 0.18.x hosts and clients interoperate) · C ABI **13** (adds `punktfunk_connection_send_pen`; embedders recompile) · Windows display driver protocol **6**, compat floor **3** (unchanged — pen injects through the Windows pen stack, not the display driver). Windows drivers ship separately, as always.