Move onde-mistralrs to 0.9.5 and release 1.3.4

.agents/skills/forked-dependencies/SKILL.md

@@ -69,8 +69,8 @@ tqdm and mio entries.
 
 | Crate | Fork / source | Ref | Purpose | Consumed where |
 |-------|---------------|-----|---------|----------------|
-| mistral.rs (12 crates) | `setoelkahfi/mistral.rs` | branch `fix/all-platform-fixes` | All-platform fixes; **published** as `onde-mistralrs-*` | onde deps (registry) |
-| candle-core, candle-nn | `setoelkahfi/candle` | branch `onde/candle-0.11.0-27f20fea` (rev `27f20fea`) | Post-0.11.0 candle APIs; **published** as `onde-candle-{core,nn}` | mistral.rs fork deps (registry) |
+| mistral.rs (12 crates) | `setoelkahfi/mistral.rs` | branch `fix/all-platform-fixes` | All-platform fixes; **published** as `onde-mistralrs-*` (0.9.3 on crates.io; 0.9.5 synced with upstream, not yet published) | onde deps (git until 0.9.5 is published, then registry) |
+| candle-core, candle-nn | `setoelkahfi/candle` | branch `onde/candle-0.11.0-27f20fea` (rev `27f20fea`) | Post-0.11.0 candle APIs; **published** as `onde-candle-{core,nn}`. Upstream mistral.rs is now on candle rev `66a8cf18`, but the fork still compiles against this snapshot (host, iOS); `reduce.metal` changed in between, so watch Metal runtime behavior | mistral.rs fork deps (registry) |
 | tqdm | *none — upstream* | crates.io `tqdm 0.8.1` | Crossterm 0.29 (→ mio 1.x) is upstream as of 0.8.1 | mistral.rs fork dep (registry) |
 | sysctl | `setoelkahfi/sysctl-rs` | branch `feature/watchos` | watchOS (`target_os = "watchos"`) support | `[patch.crates-io]` in onde **and** mistral.rs fork |
 | mio | `setoelkahfi/mio` | branch `feature/visionos` | visionOS support for the `mio 0.8.11` line | **RETIRED** — patch removed from onde once it consumed onde-mistralrs 0.9.3 (crossterm 0.29 → mio 1.x). Fork kept for history/upstream PR. |
@@ -98,6 +98,13 @@ tqdm and mio entries.
   `-flash-attn`, `-metal-compile`, `-code-exec`, `-sandbox`, `-macros`,
   `-vision`, `-audio`, `-mcp`. The rename is a `package = "onde-..."` override on
   each workspace dep entry, so Rust source still `use`s `mistralrs_*` unchanged.
+- **Paged attention:** upstream #2493 made Metal paged attention the default.
+  onde turns it off on iOS, tvOS, visionOS and watchOS with
+  `.without_paged_attn()` (GGUF and UQFF builders in `engine.rs`, and the
+  validation load in `hf_cache.rs`); macOS keeps the default. Android is CPU
+  and unchanged.
+- **0.9.5 API change:** `with_device_mapping` now takes `DeviceMapMetadata`,
+  so the tvOS memory cap uses `.with_max_model_len(512)` instead.
 - **Publishing + bump flow:** see [[sdk-regressions]] (the 12-crate publish
   order, the candle Step 0, and the tqdm/core2 publish-verification snags).
 

.agents/skills/sdk-regressions/SKILL.md

@@ -422,7 +422,7 @@ target first.
 ### Step 0 — check if the candle snapshot needs an `onde-candle-*` publish
 
 Since 0.9.x, upstream mistral.rs pins `candle` to a git rev past the latest
-released candle (e.g. rev `27f20fea`, 19 commits past `0.11.0`, adding
+released candle (e.g. rev `27f20fea`; upstream has since moved to `66a8cf18`, which 0.9.5 does not need to compile, 19 commits past `0.11.0`, adding
 `QTensor::indexed_gemv`, `gemv_fused_shared_lhs`, `BarrierPool::execute_chunked`).
 `cargo publish` strips git pins, so publish verification resolves candle from
 crates.io and fails with E0599 "method not found" if the pinned rev has APIs the
@@ -456,6 +456,18 @@ candle-metal-kernels = { version = "0.11.0" }   # unchanged at rev -> official
 
 ### Step 1 — check if `onde-mistralrs` needs a new publish
 
+> Testing before publish: keep onde on the registry form
+> (`version = "0.9.5", package = "onde-mistralrs{,-core}"`) and patch by the real
+> package name from the command line, e.g. `cargo check --config
+> 'patch.crates-io.onde-mistralrs.git="https://github.com/setoelkahfi/mistral.rs"'
+> --config 'patch.crates-io.onde-mistralrs.branch="fix/all-platform-fixes"'`
+> (repeat for `onde-mistralrs-core`). A `[patch.crates-io]` rename of `mistralrs`
+> is silently ignored. `cargo package -p <the 12 crates>` in the fork verifies
+> the packaged crates against an overlay of their unpublished siblings. Don't
+> commit the Cargo.lock these produce. Since 0.9.5, Metal
+> paged attention is upstream's default (#2493), so onde calls
+> `.without_paged_attn()` on iOS/tvOS/visionOS/watchOS.
+
 First, see how far the fork has drifted from upstream:
 
 ```bash

CHANGELOG.md

@@ -1,3 +1,16 @@
+## 1.3.4
+
+### Inference engine
+
+* Moves `onde-mistralrs` and `onde-mistralrs-core` from `0.9.3` to `0.9.5` on every target. The new crates 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. Its KV block pool does not fit in memory on iOS, tvOS, visionOS or watchOS, so model loading (and the validation load in `hf_cache`) now calls `without_paged_attn()` on those targets. Desktop and Android are unchanged.
+* On tvOS the auto device map override is replaced by `with_max_model_len(512)`.
+
+### Packaging
+
+* No public API changes: the published `onde` crate surface is identical to `1.3.3`.
+* All version sources (root, Kotlin, Dart, React Native, and the root and per-SDK lockfiles) are aligned on `1.3.4`.
+
 ## 1.3.3
 
 ### Cargo workspace