fix(wasix): preserve core error recovery and limit SJLJ code growth

[f0rr0]

Summary

A PL/pgSQL error after COPY can close the connection or trap in __wasm_setjmp_test instead of returning its SQLSTATE. The WASIX libc sigsetjmp wrapper places the exception handler in a frame that has already returned when PostgreSQL calls siglongjmp.

• Expand sigsetjmp into compiler-visible setjmp at single-user core call sites, extending the existing side-module fix.
• Enable Wasm EH/SjLj at the backend relocatable link, where ThinLTO bitcode is lowered. Enabling it only at the final link is too late.
• Set that partial link's inline threshold to zero to limit LLVM-generated catch-state spills and code duplication. Frontend and final-link optimization settings stay in place.
• Add a client regression that alternates COPY, a PL/pgSQL error, and a successful query three times, checking SQLSTATE P0001, statement rollback, and connection reuse.

Fixes #227.

Includes the CI prerequisite fixes from #228: public Rust README documentation, source-only qualification, selected source uploads, and explicit WASIX consumer artifact dependencies. Merge #228 first to keep this PR focused on the five runtime/regression files.

Artifact size

Compared with the same correctness fix using the default partial-link inline budget:

Artifact	Fix only	This change	Reduction
Core WASM	13,014,266 bytes	10,125,440 bytes	22.2%
Linux x86-64 AOT	31,461,048 bytes	26,305,272 bytes	16.4%
Compressed AOT	6,687,370 bytes	5,574,761 bytes	16.6%
Portable runtime archive	4,040,700 bytes	3,573,676 bytes	11.6%

The correctness fix alone grows WASM from 12,125,389 to 13,014,266 bytes. Analysis traced most of that growth to local loads and stores generated while lowering SjLj catch-state PHIs. The inline budget mitigates this compiler behavior; it does not fix LLVM's spill-placement algorithm.

Query performance tradeoff

Warm Linux x86-64 AOT SDK measurements, with identical SQL results and verified embedded runtime hashes:

Workload	Fix only	This change	Change
SELECT 42	81.40 µs	81.36 µs	−0.1%
Aggregate	1.957 ms	2.034 ms	+3.9%
Join	6.688 ms	6.374 ms	−4.7%
Sort	5.846 ms	5.948 ms	+1.7%
JSON processing	21.294 ms	23.008 ms	+8.0%
PL/pgSQL	798.01 µs	795.28 µs	−0.3%

These are medians of batch-average call costs, including SDK handling and row decoding, from six processes per build in repeated ABBA order on one pinned CPU. They are not individual-request latency percentiles. JSON processing shows a consistent cost; the other measurements do not establish a general speedup. This is an explicit size/performance tradeoff. TCP, concurrent throughput, cold start, and p95/p99 were not measured.

Release Intent

• Package/API/runtime change: PR title uses fix:.
• Docs/CI/repository-only change: no release intended.
• Source/input/runtime asset change: PostgreSQL/toolchain pins are unchanged; portable and Linux AOT assets were regenerated through the owning production tasks, including source fingerprints and package metadata. Generated artifacts are not committed.

Verification

• Full Moon affected and cross-platform CI checks passed: pending PR CI.
• Selected local owner checks passed: oliphaunt-pgwire-server:format-check, :lint, :test (26 tests), and :test-integration (12 tests).
• Production liboliphaunt-wasix:runtime-aot dependency chain passed: source inputs, compiler output, strict portable packaging, and Linux x86-64 AOT packaging/verification.
• liboliphaunt-wasix:smoke passed all seven PostgreSQL regression tests against the rebuilt runtime.
• The new COPY/PLpgSQL regression fails against the published old runtime and passes against the rebuilt runtime.
• Additional local probes exercised nested catches/rethrows, subtransaction and savepoint rollback, trigger failures, extended-protocol recovery, 100 alternating errors/successes, reopen, and CREATE DATABASE failure cleanup for both WAL_LOG and FILE_COPY. No unreachable traps occurred.
• All 1,336 import/export names and signatures are preserved; PL/pgSQL loads successfully. Table capacity changes from 5,770 to 5,783.
• Product artifact sizes and warm query performance were measured as described above.

The runtime size, performance, and local behavior measurements above used the exact tracked diff on base 32ce7b29510b74333e799601b69a71fd28122e80. Windows/macOS/ARM/browser qualification, the full extension lifecycle matrix, and the full TrailBase suite remain unverified locally. The host-recovery paths remain in place.

The updated head 429fac6c92671eeabb94367454a170c79ac4e285 includes current main and #228 through a linear history. Local release-intent and release-metadata checks pass; owner checks and full fresh CI qualification are still in progress. The runtime patches and COPY/error-recovery regression are unchanged. Earlier runtime measurements do not qualify the updated producer inputs.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
oliphaunt-docs	Ready	Preview	Oct 1, 2026 8:51am UTC