build(deps): GPUI Kit 0.7.1 on gpui-fast v0.1.0; composer dictation via gpui-component speech

[Tryanks]

Summary

Two things, one lock: GPUI Kit 0.7.1 requires gpui-pre 0.3.8, and the speech module that replaces crates/voice ships in gpui-component 0.7.1.

 Cargo.toml [patch.crates-io] gpui-pre* → git gpui-fast   (1b381adb → 3755a552 = v0.1.0, zed a1b7107)
 crates/ui/Cargo.toml
-  gpui-base 0.7.0, gpui-kit-assets 0.7.0, gpui-component-macros 0.7.0, gpui-wry 0.7.0
+  gpui-base 0.7.1, gpui-kit-assets 0.7.1, gpui-component-macros 0.7.1, gpui-wry 0.7.1
+  gpui-component 0.7.1 (features = ["speech"], optional)   # enabled by `voice` (desktop only)
-  voice = ["dep:tcode-voice"]            # macOS 26 only, Swift shim
+  voice = ["dep:gpui-component"]         # macOS 13+, Windows; Linux has no system recognizer
-crates/voice/                            # Rust + Swift SpeechAnalyzer shim, SDK-26 build.rs
+crates/app/Info.plist                    # usage descriptions embedded into the unbundled binary

Upstream API changes needed three call-site fixes: wrap_line gained an indent argument (markdown inline flow), RequestFrameOptions gained signal fields and WgpuRenderer::gpu_specs returns Option (iOS and Android backends).

gpui-fast v0.1.0 is now on crates.io, but gpui-kit still pins gpui-pre, and a [patch] cannot redirect to a differently named registry crate, so the git patch stays as gpui-fast's compat README documents; only the lock moves.

Composer dictation now runs on the kit's session state; Tcode keeps only editor bookkeeping:

mic click → SpeechState::toggle
  Started          focus textarea, anchor = cursor
  Partial / Final  replace anchor..anchor+last_len with the whole transcript
                   (abort if the editor value is not what we last wrote)
  Escape           stop (graceful, Final within the kit's 3 s timeout)
  typing / submit / destination switch / teardown
                   cancel at once, text already in the editor stays

The kit's SpeechButton/SpeechWaveform read gpui-component's theme global, which Tcode does not initialise, so the mic stays Tcode's own button over SpeechState::levels(); the site names that gap.

Evidence

• Before: mic button only on macOS 26 with an SDK-26 toolchain; transcript_edit unit test owned the old chunk arithmetic.
  After: cargo nextest run --workspace --locked → 889 passed, 6 skipped. New dictation_revises_the_draft_and_cannot_overwrite_edits_or_submission drives the real composer through the kit's SpeechRecognizer/AudioInput seams: UTF-8 anchored insertion, longer/shorter/empty revisions, Final without duplication, typing cancels, Escape → Stopping, submit while Stopping ignores late results.
• fmt, clippy -D warnings, machete, and RUSTFLAGS=-D warnings checks for iOS sim, wasm and Android NDK all pass locally; the mobile/web lock graphs contain no gpui-component, cpal or oboe.
• Desktop app launched with a throwaway TCODE_DATA_DIR; mic button present and legible in both themes, wide and narrow. otool confirms __TEXT,__info_plist in the dev binary.
• Not verified: live recognition on this Mac (the TCC prompt could not be accepted from automation), Windows and Linux runtime, and SFSpeechRecognizer vs SpeechAnalyzer quality. crates/voice stays in history and the kit's recognizer seam allows bringing it back if a real gap shows.

Merge Danger

Door: two-way

Revert restores the Swift engine and the previous lock. No data or protocol changes.

Blast Radius: desktop

Every desktop build takes gpui-component (+39 lock entries, cpal needs libasound2-dev on Linux, already in CI). Windows dictation sends audio to Microsoft's online service; macOS stays on-device. Any GPUI rendering difference from the 84 gpui-fast commits reaches all platforms; the full test suite and smoke launch showed none.