fix(rust/ffi): forward original vendor_code as an error detail on the 1.1.0 layout

[fornwall]

Problem

In the ADBC 1.1.0 C error layout, AdbcError.vendor_code must hold the ADBC_ERROR_VENDOR_CODE_PRIVATE_DATA sentinel (INT32_MIN) to signal that private_data carries the extended error info. The Rust FFI exporter's export_error therefore stamps the sentinel over the driver's real vendor code on the 1.1.0 path (rust/ffi/src/types.rs), destroying it for every 1.1.0-layout consumer — only Rust-native and 1.0.0-layout callers can still see it.

This bit the adbc-spanner Rust driver (review finding UP-10): it keeps the numeric gRPC status code in vendor_code (e.g. 10 = ABORTED) as a documented contract, but that contract silently does not hold for 1.1.0 C callers.

Fix

When exporting through the 1.1.0 path and the Rust Error carries a meaningful vendor code (nonzero and not the sentinel), export_error now appends it as an additional error detail — the 1.1.0 rich-error-metadata mechanism, readable via AdbcErrorGetDetail:

• key: adbc.error.vendor_code (exported as adbc_ffi::ERR_DETAIL_VENDOR_CODE) — namespaced adbc.<object>.<field> naming the exact spec field it mirrors, following the existing exporter-synthesized detail-key precedent adbc.drivermanager.driver_load_trace in rust/driver_manager/src/search.rs
• value: the code as a decimal string in UTF-8 (textual detail values, like the driver-load trace and the postgres driver's PG_DIAG_* details; no -bin suffix since it is not binary proto)

The 1.0.0 path and the spec-mandated sentinel itself are unchanged; pre-existing details are preserved, the synthesized one is appended after them.

How the Go and C++ implementations handle the same collision

Neither forwards the vendor code as an error detail today, so the detail key introduced here has no cross-language precedent — but neither has a lossless answer either:

• Go (setErrWithDetails in the generated exporters, template go/adbc/pkg/_tmpl/driver.go.tmpl) makes a case-by-case trade:
  • 1.0.0 caller → writes the real VendorCode into vendor_code (same as Rust).
  • 1.1.0 caller, vendor code set but zero details → deliberately downgrades to the plain layout: the real code goes into vendor_code and private_data is left unpopulated, sacrificing the extended layout to keep the code visible.
  • 1.1.0 caller with details → sentinel + details win, and the VendorCode is silently dropped.
• C++ never faces the problem: the driver framework's Status (c/driver/framework/status.h) carries a status code, message and details but has no vendor-code field at all, and nothing under c/driver/ ever assigns a real vendor_code (the only occurrences are sentinel comparisons; likewise the legacy c/driver/common/utils.c helpers). C++-framework drivers simply never populate a vendor code.

So this PR's behaviour is aligned with Go in intent (don't let the 1.1.0 sentinel destroy the code) but strictly more information-preserving in mechanism: the code survives even when other details are present (Go's lossy case), and the extended layout is never sacrificed (Go's no-details case). The key is deliberately language-neutral so the Go exporter could adopt the same detail to fix its lossy branch if this is proposed upstream.

Tests

• test_set_error_out_extended_layout_forwards_vendor_code_as_detail — a 1.1.0 consumer reads the code back through the exporter's real ErrorGetDetailCount/ErrorGetDetail functions (made pub(crate) for the test), alongside a preserved pre-existing detail
• test_set_error_out_extended_layout_skips_meaningless_vendor_codes — no detail for 0 or the sentinel
• test_set_error_out_v100_layout_keeps_vendor_code_in_place — the 1.0.0 path still passes the code through vendor_code itself

cargo test -p adbc_ffi (11 passed), cargo clippy -p adbc_ffi --all-targets --all-features -- -Dwarnings and cargo fmt --all -- --check are clean. Workspace clippy/test failures in adbc_driver_manager are pre-existing on main (identical 29 dead-code errors with and without this change; they need the sqlite driver test lib).

Upstreaming

This is a fork-internal staging of a fix intended to be proposed upstream to apache/arrow-adbc (see the Go/C++ comparison above — upstream could pair this with the matching fix to the Go exporter's lossy details-present branch).

🤖 Generated with Claude Code

https://claude.ai/code/session_013bxmWF4Z9R4y1kaHEAPgS6