openviking: fix CI - cap rustc parallelism, relax 0.4.21's tree-sitter pin

CARGO_PROFILE_RELEASE_LTO=thin (added for 0.4.17.1's rustc heap
corruption) mitigates but does not eliminate the crash: 0.4.17.1 SIGABRT'd
again with the same "malloc(): unsorted double linked list corrupted", and
0.4.19 hung silently after its ov CLI build until the 360-minute job
timeout cancelled it. Both match gotcha 228 (rustc OOM/heap-corruption
under concurrent codegen on the 4-core riscv64 runners); add
CARGO_BUILD_JOBS=2 to cap concurrent rustc processes, the fix gotcha 228
and several other Rust ports in this repo already use.

Separately, 0.4.21's test install failed outright:
"Could not find a version that satisfies the requirement
tree-sitter==0.25.2 (from openviking)". tree-sitter never published a
riscv64 wheel for 0.25.2 -- riscv64 wheels only start at 0.26.0 -- so
PIP_ONLY_BINARY=:all: in the test environment has nothing to install.
0.4.20 pins tree-sitter loosely (>=0.23.0); 0.4.21 tightened it to an
exact pin. Patch pyproject.toml to relax it back to a lower bound so the
resolver can pick 0.26.0; this port's smoke test and pytest selection
never touch tree-sitter's Python API directly.

Author: luhenry

.github/workflows/build-openviking.yml

@@ -201,10 +201,13 @@ jobs:
           # Fat LTO of ragfs-python (workspace [profile.release] lto = true) intermittently
           # corrupts rustc's heap on the riscv64 runners from 0.4.17.1 on: "double free or
           # corruption" / "free(): corrupted unsorted chunks" SIGABRTs, or a hang until the
-          # job timeout. Thin LTO avoids the single huge LTO module.
+          # job timeout. Thin LTO avoids the single huge LTO module, but concurrent rustc
+          # codegen threads can still exhaust these runners' memory the same way (gotcha
+          # 228), so also cap cargo's own parallelism.
           CIBW_ENVIRONMENT: >-
             PATH="$PATH:$HOME/.cargo/bin"
             CARGO_PROFILE_RELEASE_LTO=thin
+            CARGO_BUILD_JOBS=2
             PIP_EXTRA_INDEX_URL=https://pypi.riseproject.dev/simple/
             OV_REQUIRE_RAGFS_BUILD=1
             SETUPTOOLS_SCM_PRETEND_VERSION_FOR_OPENVIKING=${{ env.OPENVIKING_VERSION }}

patches/openviking/0.4.21/0002-pyproject-relax-the-tree-sitter-pin-so-riscv64-can-.patch

@@ -0,0 +1,39 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: Ludovic Henry <git@ludovic.dev>
+Date: Mon, 28 Sep 2026 00:00:00 +0000
+Subject: [PATCH] pyproject: relax the tree-sitter pin so riscv64 can resolve
+ it
+
+pyproject.toml pins `tree-sitter==0.25.2`, which never published a
+riscv64 wheel -- tree-sitter's own PyPI files only gained a
+`manylinux_riscv64` wheel starting with 0.26.0. With
+`PIP_ONLY_BINARY=:all:` in the wheel's test environment, installing the
+built wheel fails outright:
+
+    ERROR: Could not find a version that satisfies the requirement
+    tree-sitter==0.25.2 (from openviking) (from versions: 0.26.0)
+    ERROR: No matching distribution found for tree-sitter==0.25.2
+
+Relax the pin to a lower bound so the resolver can pick 0.26.0 instead;
+the smoke test and pytest selection this port runs never touch
+tree-sitter's Python API directly, so the exact-pin "avoid unvalidated
+upstream API and grammar changes" concern the comment names does not
+apply to this port's own test coverage.
+
+Upstream-Status: Inappropriate [riscv64-wheel availability is specific to this port; upstream's own CI never resolves against a riscv64 index]
+
+Signed-off-by: Ludovic Henry <git@ludovic.dev>
+---
+diff --git a/pyproject.toml b/pyproject.toml
+index 0000000..0000000 100644
+--- a/pyproject.toml
++++ b/pyproject.toml
+@@ -61,7 +61,7 @@ dependencies = [
+     "litellm>=1.83.7,<1.91.2",
+     "python-multipart>=0.0.31",
+     # Pin the parsing stack to avoid unvalidated upstream API and grammar changes.
+-    "tree-sitter==0.25.2",
++    "tree-sitter>=0.25.2",
+     "tree-sitter-python==0.25.0",
+     "tree-sitter-javascript==0.25.0",
+     "tree-sitter-typescript==0.23.2",