[setoelkahfi]

Moves every target from onde-mistralrs{,-core} 0.9.3 to 0.9.5 and bumps the SDK to 1.3.4. No public API changes.

The 0.9.5 crates are published on crates.io. They merge upstream mistral.rs 0.9.5 into the fork and add visionOS and watchOS arms to the Metal embedded_library! macro, so --features metal builds for tvOS, visionOS and watchOS again.

Upstream made Metal paged attention the default, and its KV block pool does not fit in memory on iOS, tvOS, visionOS or watchOS. Model loading in engine.rs and the validation load in hf_cache.rs now call without_paged_attn() on those targets. On tvOS the auto device map override is replaced by with_max_model_len(512). Desktop and Android are unchanged.

The branch has two commits:

• the dependency move and the mobile paged-attention change
• the 1.3.4 version bump across the root crate, Kotlin, Dart, React Native, their lockfiles and the changelogs

Checked against the published crates, with no "patch not used" warnings:

• host: cargo check, cargo clippy -- -D warnings, cargo test inference:: (46 passed), cargo check --workspace
• cargo check for iOS, iOS-sim and Android
• cargo +nightly check -Z build-std for tvOS, visionOS and watchOS

cargo test hf_cache:: runs no tests because that module has none.

sdk/dart/rust/Cargo.lock and sdk/react-native/rust/Cargo.lock still pin onde-mistralrs at 0.9.3. Cargo ignores them in workspace mode and they were already stale, so I left them alone.

Not included: the 1.3.4 tag, which triggers the five publish workflows.

[sigit-code-review]

This change updates all targets to use onde-mistralrs 0.9.5 (forked from upstream mistral.rs 0.9.5), adds visionOS and watchOS Metal support, and disables paged attention on iOS, tvOS, visionOS, and watchOS to avoid memory issues. On tvOS, the device mapping override is replaced with with_max_model_len(512) due to an upstream API change. The version is bumped to 1.3.4 across the root crate, Kotlin, Dart, React Native, and changelogs. Documentation is updated to reflect the new dependency versions and behavioral changes. No public API changes are introduced. The changes are well-documented and the migration to the new paged attention default is handled with target-specific cfg guards. No findings from previous reviews remain unresolved, and the update appears to be robust and well-tested as described.

---

Automated review by siGit Code Review · commit 77523af · see the review dashboard

[sigit-code-review]

This change bumps the onde SDK and all wrappers to 1.3.4 and moves all targets from onde-mistralrs 0.9.3 to 0.9.5, syncing with upstream and adding visionOS/watchOS Metal support. It disables paged attention on iOS, tvOS, visionOS, and watchOS to avoid memory issues, and replaces the tvOS device mapping override with with_max_model_len(512) per the new API. The changelogs and all version sources are updated. No public API changes are introduced. The change is internally consistent and matches the described migration and upstream changes. No critical or warning-level issues are introduced by the diff.

---

Automated review by siGit Code Review · commit 3f2bc56 · see the review dashboard