vendor

Third-party crates FastTrackStudio patches, kept in one place so the fix lives once instead of once per repo.

Every consumer repo (architect, task, fasttrackstudio) points a [patch.crates-io] entry at this repo and pins its own rev.

Each is the published crates.io release named in the table, verbatim, plus the one fix. Do not casually rebase onto upstream main — upstream has drifted (see the notes below), and a rebase silently mixes API changes into what is meant to be a minimal fix. To move to a newer release, start from that release's published source and re-apply the fix; the fleet (facet 0.50 / vox 0.10 / phon 0.2) is on rc.7 as of 2026-09.

crate	upstream	pinned to	why
phon	bearcove/phon	0.2.0-rc.7	Def::Scalar opaque shapes
phon-jit	bearcove/phon	0.2.0-rc.7	nightly probe breaks under nix
styx-format	bearcove/styx	5.0.0-rc.5	angle brackets in bare scalars
facet-core	facet-rs/facet	0.50.0-rc.7	chrono display truncates sub-second precision
vox-phon	bearcove/vox	0.10.0-rc.7	bogus post-decode length check aborts every real schema exchange
tao	tauri-apps/tao	0.35.3	iOS release builds crash at launch on a released scene configuration
anyrender_vello	dioxuslabs/anyrender	0.12.0	window-renderer fixes found at 2560x1440 + the FTS blitz compositor/rasterizer (studio)
anyrender_vello_hybrid	dioxuslabs/anyrender	0.8.0	set_size resizes a wrong-sized surface (studio)
baseview	RustAudio/baseview	0.3.3	macOS frame timer in common run-loop modes, so live resize doesn't freeze (studio)
dioxus-core	DioxusLabs/dioxus	rev f717a8e1	Callback::__point_to no longer unwraps a dropped owner (studio)
dioxus-native	DioxusLabs/dioxus	rev f717a8e1	iOS simulator takes vello-hybrid (studio)
dioxus-native-dom	DioxusLabs/dioxus	rev f717a8e1	stale ElementId→NodeId mappings after subtree removal (studio)
raw-window-metal	rust-windowing/raw-window-metal	1.1.0	ObjC class name collides across plugin bundles (studio)
wgpu-hal	gfx-rs/wgpu	29.0.4	iOS Simulator has indirect draw/dispatch (studio)
winit-uikit	rust-windowing/winit	0.31.0-beta.2	proxy wake-up drains events queued during it (studio)
branches	fereidani/branches	0.4.6	its nightly path uses core::intrinsics items current nightlies dropped (studio)
world-sys (world/)	mmorise/World	—	the WORLD vocoder, BSD-3-Clause (studio)

Each of the crates marked (studio) carries its own UPSTREAM.md or FTS-PATCH.md with the full story. They moved here from studio's in-tree vendor/ (2026-10) so a studio CI run caches them as compiled dependencies instead of building and testing them as workspace members; studio patches them in from a tag of this repo.

tao — the scene configuration it returns is already released

Symptom: a release iOS build of a Dioxus app crashes on every launch, before any app code: EXC_BAD_ACCESS in objc_retain, under -[UIApplication _connectUISceneFromFBSScene:transitionContext:]. A debug build of the same code runs. iOS 27 requires the scene lifecycle, so every app with UIApplicationSupportsMultipleScenes takes this path.

tao's application:configurationForConnectingSceneSession:options: returned Retained::as_ptr(&config) and then dropped config — the only reference — so the caller received a pointer to an object it had just released. In debug an autorelease pool happened to hold it; in release the objc runtime's autorelease elision does not, and UIKit retains freed memory.

Fix (src/platform_impl/ios/view.rs, one line): return Retained::autorelease_return(config), the +0 autoreleased reference an Objective-C method returning an object owes its caller.

phon — Def::Scalar opaque shapes

Symptom: every task-server list RPC aborts the process.

vox wraps each RPC reply payload as a phon opaque field. Upstream's derive has no branch for Def::Scalar opaque shapes (Uuid, chrono types, url::Url) — ref_of falls through to DeriveError::Unsupported. Because the encode thunk is extern "C", the panic crosses a non-unwinding boundary and aborts rather than unwinding.

Fix (src/derive.rs, ~130 lines): a last-resort branch encoding display/parse scalars as a Primitive::Bytes UTF-8 Display run, decoded via the shape's parse vtable. All pre-existing branches keep precedence, so previously-working shapes encode byte-identically.

Upstream status: not fixed as of the published 0.2.0-rc.7: derive.rs has no Def::Scalar, has_display, has_parse, or Display fallback; the ref_of chain still ends in Unsupported.

Forward-ported to rc.7 (2026-09): the published rc.7 crate verbatim plus the same four hunks in src/derive.rs (129 added lines, nothing removed). The only adaptation is the ref_of hunk's neighbours, which rc.7 spells SchemaId::from_raw(..) where rc.5 had SchemaId(..). Covered by tests/display_scalar.rs (its own test target, with its own dev-deps, because the publish strips the upstream tests' dev-deps); it fails on unpatched rc.7 with the lowering error.

phon-jit — nightly probe breaks under nix

Symptom: the macOS desktop build fails to compile stencils.

Upstream's build.rs calls nightly_available(), which probes rustc +nightly. The nix-pinned toolchain has no rustup nightly, and the stable fallback cannot compile the stencils either because become (explicit tail calls) is parse-gated, not feature-gated.

Fix (build.rs only, ~30 lines): use PHON_JIT_NIGHTLY_RUSTC when set, otherwise set RUSTC_BOOTSTRAP=1 and always compile the tail-call stencils. Sources and stencils are untouched.

Upstream status: not fixed. build.rs on main still calls nightly_available(), with no env override and no cfg gate.

Note for a future re-vendor: upstream main now gates the arm64-macOS path on DEP_WEAVY_JIT == "1" and adds a phon_jit_native cfg. That is not in rc.5 and does not help the aarch64-apple-darwin build this fix exists for.

styx-format — angle brackets in bare scalars

Symptom: a saved keybind config cannot be re-parsed.

facet-styx emits keys <C-s> as a bare scalar, and styx-parse then rejects > because it is the attribute separator. The formatter produces a file its own parser refuses.

Fix (src/scalar.rs, one line): add '<' | '>' to the set can_be_bare rejects.

Upstream status: not fixed — scalar.rs is unchanged since 2026-01-16 and angle brackets are still absent from the reject set. Upstream's own compliance corpus documents the bug (compliance/corpus/01-scalars/bare-angle-brackets.styx: "< can appear in bare scalars, but > cannot").

Strictly only > is fatal; quoting both is the safe superset.

rc.5 is still the newest published styx-format, and facet-styx 5.0.0-rc.7 depends on ^5.0.0-rc.5, so this entry needed no forward-port for the rc.7 fleet.

vox-phon — bogus post-decode length check aborts every real schema exchange

Symptom: any peer-to-peer schema exchange over vox fails with "schema decode consumed N of M bytes" — a daw-proto batch type as small as one enum variant already hits it ("consumed 159 of 361 bytes"), so this is not an edge case, it is the common case.

parse_schema_bytes (src/schema.rs) decodes each schema in the closure with phon_schema::schema_from_bytes, then rejects the result unless its second tuple element equals the encoded slice length. That element is not a buffer position — phon_schema 0.2.0-rc.6 changed schema_from_bytes to return (Schema, usize) where the usize is the owned-allocation byte count charged against the new DecodeLimits budget (string/Vec bytes it had to copy for DecodeLimits accounting). It has no relationship to the serialized length: struct/enum tags, field-name strings, and list-length headers all add encoded bytes without charging the allocation budget, so the two numbers diverge for every schema with any structure. schema_from_bytes already guarantees full-slice consumption itself — it rejects trailing bytes internally before returning Ok — so the check in vox-phon was both wrong and redundant.

Fix (src/schema.rs, one line removed from parse_schema_bytes): drop the if consumed != slice.len() { ... } block; keep the schema_from_bytes call, ignoring its second element.

Upstream status: not filed yet. This surfaced from bumping architect (which pulls vox) in the daw repo; not yet checked against bearcove/vox main.

Upstreaming

All are filable against the bearcove/facet-rs repos, which are active. Doing so is a deliberate decision, not a chore to run unprompted.

facet-core — chrono display truncates sub-second precision

Symptom: every timestamp on every vox RPC arrives rounded to the second, and an activity feed ordered by time puts a checkpoint before the root creation it followed.

chrono's types are registered as Def::Scalar / UserType::Opaque, so their only encode path is the display fn in each vtable — there is no structural representation for a format to fall back on. Upstream 0.50.0-rc.5 writes the DateTime types with to_rfc3339_opts(SecondsFormat::Secs, true) and the naive types with %Y-%m-%dT%H:%M:%S / %H:%M:%S. Still so in the published rc.7. All four drop the fraction, so every facet format inherits the loss: facet-json, facet-toml, and vox.

It is silent and one-directional. The value round-trips without error and comes back rounded, which is why it survived: nothing fails, and two events in the same second simply stop being orderable.

Fix: SecondsFormat::AutoSi for the RFC3339 types, %.f for the naive ones. Both write nothing when the fraction is zero, so a whole-second value formats exactly as it did before and only the previously-lossy case changes — existing files keep parsing, and the diff on stored data is empty unless precision was actually being lost. NaiveDateTime's parse format gains %.f for the same reason (its display now emits one; NaiveTime's parse already accepted it).

Forward-ported to rc.7 (2026-09): the published rc.7 crate verbatim (its Cargo.toml.orig kept as Cargo.toml.upstream) plus the same nine-line src/impls/crates/chrono.rs change and test target.

Covered by facet-core/tests/chrono_precision.rs, which is its own test target: the crate's main target pulls in type_name.rs, whose facet-testhelpers dev-dependency the publish strips, so main cannot build from the published source at all.