Phased C Rust migration

[dddiggory]

This pull request contains changes generated by a Cursor Cloud Agent

  

---

Note

High Risk
High risk because it changes core distance/math routines and parts of HNSW planner logic to call into a new Rust static library via FFI, which can impact correctness, performance, and cross-platform builds.

Overview
Introduces a new Rust staticlib (rust/) and wires it into the extension build/link process (Makefile + Makefile.win), making Rust a required build dependency and updating CI to install a stable toolchain.

Moves a growing set of performance-critical kernels to Rust via rust_ffi.h/ffi.rs, and switches C call sites (notably bitutils, halfvec/halfutils, and parts of hnsw cost/initialization) to use vector_rust_* implementations, adding SQL-callable parity wrapper functions like vector_rust_* for validation/debugging.

^{Written by Cursor Bugbot for commit 53d127d. This will update automatically on new commits. Configure here.}

[cursor]

Cursor Agent can help with this pull request. Just @cursor in comments and I'll start working on changes in this branch.
_{Learn more about Cursor Agents}

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

Bugbot Autofix prepared fixes for both issues found in the latest run.

• ✅ Fixed: Dead code: Default assignments immediately overwritten by Rust
  • Removed the dead code assignments to Default implementations that were immediately overwritten by Rust implementations.
• ✅ Fixed: Unrelated kernel misused for ratio cap check
  • Replaced the fragile misuse of vector_rust_hnsw_should_update_element_max_distance_kernel with a direct ratio > 1.0 comparison.

Or push these changes by commenting:

@cursor push 8537f92ebf

Preview (8537f92ebf)

diff --git a/src/halfutils.c b/src/halfutils.c
--- a/src/halfutils.c
+++ b/src/halfutils.c
@@ -307,11 +307,6 @@
 	 * Could skip pointer when single function, but no difference in
 	 * performance
 	 */
-	HalfvecL2SquaredDistance = HalfvecL2SquaredDistanceDefault;
-	HalfvecInnerProduct = HalfvecInnerProductDefault;
-	HalfvecCosineSimilarity = HalfvecCosineSimilarityDefault;
-	HalfvecL1Distance = HalfvecL1DistanceDefault;
-
 	HalfvecL2SquaredDistance = HalfvecL2SquaredDistanceRust;
 	HalfvecInnerProduct = HalfvecInnerProductRust;
 	HalfvecCosineSimilarity = HalfvecCosineSimilarityRust;

diff --git a/src/hnsw.c b/src/hnsw.c
--- a/src/hnsw.c
+++ b/src/hnsw.c
@@ -218,7 +218,7 @@
 HnswShouldCapRatioAtOne(double ratio, bool useRust)
 {
 	(void) useRust;
-	return vector_rust_hnsw_should_update_element_max_distance_kernel(true, true, 1, ratio);
+	return ratio > 1.0;
 }
 
 FUNCTION_PREFIX PG_FUNCTION_INFO_V1(vector_hnsw_clamp_ratio);

[cursor]

Dead code: Default assignments immediately overwritten by Rust

Low Severity

In HalfvecInit, the four function pointer assignments to *Default (lines 310–313) are immediately overwritten by the Rust wrapper assignments (lines 315–318). The Default assignments are now dead stores that serve no purpose — the Rust versions unconditionally replace them before the SIMD dispatch check.

 

[cursor]

Unrelated kernel misused for ratio cap check

Low Severity

HnswShouldCapRatioAtOne calls vector_rust_hnsw_should_update_element_max_distance_kernel(true, true, 1, ratio) to check if ratio > 1.0. This repurposes a completely unrelated kernel with carefully chosen constant arguments to produce the desired boolean result. The coupling is fragile — any change to the distance kernel's logic would silently break the ratio check.