fix(runtime): consolidate patch correctness and retire unsafe shortcuts

[f0rr0]

Latest update — 41c74d8 (2026-09-08)

Not merge-ready. The Windows/MSVC exception-recovery blocker remains. The new execution-stack guard is a tested review candidate, not enabled in consumer packages.

This update pushes the completed resource-budget cleanup and the actual stack-safety candidate code:

• Centralize existing transport/link-time budgets, preserve their values, and document ownership, tradeoffs, PostgreSQL/PGlite comparisons and the distinction between linear and execution stacks. No new environment flag or larger production stack.
• Add four candidate patches and focused regression probes: Wasmer VM accounting/EH guard, fixed Wasmer API query, PostgreSQL early check, and Rust import registration. Each passes clean application against its documented base.
• Keep those patches outside production series and dependency selection. Published Wasmer 7.2.1 lacks the new API; Windows/browser contracts and recovery margins are not qualified. This PR does not pretend to deliver the new guard to downstream consumers yet.
• Remove ordinary-host-call stack snapshots from the candidate. Its successful fixed query has no allocation, lock, arbitrary callback or host-stack switch. Missing accounting and the 64 KiB engine floor fail terminally; PostgreSQL retains a separate 512 KiB early-recovery reserve and the engine retains its 1 MiB default.
• Preserve code and evidence without importing binaries, producer trees, the benchmark framework or historical result archives.

Candidate evidence, not exact-head release qualification

• Wasmer VM: 62/62 debug and 62/62 release tests; baremetal compile check.
• 1,000-cycle scalar ABI/nested-stack probe; 100 serialized-AOT deep throw/catch and rethrow cycles, followed by expected terminal stack failure.
• Real PostgreSQL memory/directory recovery: repeated SQLSTATE 54001, subsequent queries, PL exception catches, savepoint rollback and directory reopen.
• 144 successful benchmark children against the retained previous guarded direct-binding candidate, with identical guest/AOT and SDK dependency versions. RTT improved 2.6% memory and 3.6% in the directory follow-up. The initial +8.2% directory temporary-INSERT warning did not reproduce at that magnitude: the eight-pair follow-up measured +1.7%; COPY measured −2.0%. Small overheads remain uncertain; no blanket non-regression claim.
• PR preparation: pinned Moon repo:docs-policy sdk-contracts:all passed (9 tasks, 5 cached), profile-budget test passed, both pq_flush ABI tests passed, regression-probe Rust formatting passed, all four candidate patches passed git apply --check, and staged whitespace checks passed.

Before activation: settle/consume the upstream VM/API contract (coordinate Wasmer #6913), qualify recovery margins and error cleanup across supported targets, resolve the Windows/browser contract, then integrate pins/imports/artifact invalidation together. No release build or exhaustive qualification was rerun merely to publish this review update.

---

Summary

Previously published checkpoint: 0f4a996. Not merge-ready: Windows/MSVC Wasm exception recovery is blocked in the consumed Wasmer runtime. The known formatting and Postmaster source-root fixes are pushed. No Windows target/test is removed or waived, and the repeat optimization remains held outside this PR.

Consolidate the completed patch-correctness work on current main, without importing the research harness, historical result archives, old binaries or producer trees.

• Recover Native trusted-session, startup/cwd cleanup, host process-boundary and COPY/cancellation fixes while preserving main's ABI 10 and public SDK APIs.
• Replace WASIX host-longjmp symptom handling with live guest-local recovery and typed outcomes. Integrate the matching Rust/browser host ABI, bounded protocol/tool output, entropy and fail-closed lifetime/transport changes.
• Remove 13 rejected embedded WASIX patches, including LIKE, int4 comparator shortcuts, semaphore reset and scalar atomics. Preserve the seek, parser, WAL-sync and JSONB optimizations already merged on main.
• Use the fixed strict-memory AOT profile llvm-opta-ro_ftable throughout production, caches, manifests and all eight carriers; reject incompatible legacy overrides and artifacts.
• Clean the ordered Native (22), embedded WASIX (31), Postmaster PostgreSQL (7), and browser-host (27) patch series, with patch-local motivation and focused checks. The Postmaster runtime has separate ordered Wasmer (9) and libc (8) series. Do not claim the inherited runtime bundle is fully qualified or upstream-ready.
• Align generated-header installation/fingerprints, ABI export policies, Moon cache inputs and the existing core-only fast-path packaging check.

This is intentionally a substantial correctness integration, not a small isolated performance patch: guest/host ABI and Native process ownership must land together. It excludes optional compiler/runtime experiments and the obsolete pre-N-API Node filesystem path.

Follow-up implemented at 51789da

• Preserve PostgreSQL's Native fsync=on default instead of silently adding -F; explicit fsync=off remains available. Fresh Linux ICU direct and installed-broker checks verify both settings, identity/login policy, session reset and nondefault-database settings.
• Fix the ICU collation import failure without permitting subprocess execution, and move the signal-mask declaration to its type-owning header to fix Windows runtime/extension compilation.
• Replace the Postmaster monoliths with 8 Wasmer + 7 libc logical patches. Strict application produces exactly the previous source trees. Separate Wasmer0009 disables unsafe nonvolatile-memory assumptions in all four compiler routes and invalidates old artifact identities.
• Postmaster source/shell units and both freshly compiled product/CLI strict-memory identity tests pass. The corrected-profile hosted Postmaster producer/qualification at 51789da also passed (job); the subsequent libc0008 full carrier remains subject to the new head's CI.
• SDK header mirrors, Clippy and the build graph failures are corrected. The repository-pinned Moon toolchain runs locally; SDK contracts and graph tests pass.

Additional follow-up at 8e45be7

• Initialize the WASIX host-selected role as the real PostgreSQL session principal during startup. Enforce catalog login/admission, role/database defaults and login triggers; prevent RESET ROLE/DISCARD ALL from restoring bootstrap-superuser rights. Delete the late SET ROLE workarounds from Rust direct/proxy and browser direct/worker paths.
• Preserve typed startup SQLSTATE/details for Rust SDK callers and the exact original error bytes for proxy clients, using the existing decoder.
• Make libc sigsetjmp evaluate both arguments once while keeping setjmp in the live caller. C/C++ O0/O2 host checks and actual C/C++ Wasm O2 execution pass; the old macro fails the negative control. The new helper/header was tested over the retained EH sysroot on a freshly built strict Wasmer runner, not mislabelled a full sealed-carrier rebuild.
• Add inherited memcmp alignment, byte-order and protected-page boundary checks; no new speedup is claimed. Update ordered-series source locks and expose the newly established POSIX signal-mask limitation truthfully.
• Reuse the active AOT-enabled serializer executable instead of rebuilding a shared bare xtask path which another worktree can replace. No new public environment flag or research framework.

Hosted failure fixes prepared at 94d674d

• Initialize and consistently propagate the Postmaster libc source root before capability validation. A no-build behavioral test executes the real script prefix/inventory with default/custom roots and verifies missing-libc rejection.
• Correct the single Biome formatting failure. The full 263-file JS formatting path list now passes locally using the exact pinned Biome version.
• These fixes plus profile/experiment, timeout and Windows-blocker documentation are pushed at 0f4a996. All material producers finished before the push: iOS extensions and Linux/macOS AOT passed on 8e45be7; Windows serialized/validated/installed AOT successfully but crashed during SQL-error recovery in its Rust regression suite. Neither the failed parent run nor its successful individual jobs are represented as exact-head green qualification.

Final focused validation

• Fresh portable guest/support/initdb, standard and ICU seeds, strict AOT packaging/artifact checks pass. Strict PostgreSQL replay: 22 Native / 31 WASIX / 7 Postmaster, no fuzz/rejects.
• Rust on those fresh assets: 211 library tests, 19 runtime tests, 7 proxy tests, 7 PostgreSQL regression tests pass with extensions enabled. This is a different feature scope from the initial tools-enabled test count below.
• The final tools-enabled configuration additionally passes 223 library tests, 7 public-API tests and 5 doctests.
• Real Chrome memory/IndexedDB recovery and COPY floor plus configured-role direct→worker→direct reopen, defaults, login triggers, reset and privilege/NOLOGIN rejection pass. TypeScript 343 tests / 36 files, binding and example typechecks pass.
• SDK contracts, graph unit tests, WASIX source/artifact guards, Postmaster canonical source/shell units and source lock verification pass. Rust formatting and Clippy pass for WASIX with/without extensions and xtask with/without the serializer feature.
• Final benchmark: 220 independent children, all correctness and actual durability/version oracles pass. Prior corrected candidate and final identity candidate retain matching RTT; final/strict-main RTT is +3.5% memory / +2.3% directory. Historical narrow-directory ~10% warning did not repeat, but final/prior wide temporary INSERT warns +5.1% memory / +4.2% directory. No blanket non-regression claim.
• Stock PostgreSQL 18.4 context still shows final narrow INSERT 1.41x memory / 1.64x directory, wide INSERT about 1.60x, and mixed OLTP 1.35x / 1.51x. Two stock-server process replicates and differing transports make this descriptive context, not paired cross-engine proof. The checked-in report contains the full nine-case table, scale, settings and provenance. Equal affinity was not exclusive CPU isolation; unrelated shared-host activity remained.

Initial 34f173f validation

• Strict PostgreSQL 18.4 patch replay: 22 Native / 30 WASIX / 7 Postmaster, no fuzz or rejected hunks.

• Fresh Linux standard Native library: ABI conformance, configured identity/login policy, six startup rejection cases, signal/mask/timer ownership, cancellation and COPY timeout/backpressure, five cwd cases and broad C smoke passed.

• Rust WASIX: 215 unit tests, 5 doctests, 7 public API tests passed.

• Browser TypeScript: 343 tests / 36 files and typecheck passed; six pinned-host source-contract tests passed.

• Real isolated Chrome: memory and IndexedDB both passed repeated syntax/division error recovery, PL/pgSQL exception catch, set-based INSERT, 1 MB COPY callback abort/drain/reuse; IndexedDB close/reopen preserved all 1,000 rows.

• Fresh portable guest/support/initdb built; standard and ICU seed initialization/profile probes passed. Strict AOT compiled, packaged and passed artifact validation; all eight actual generated carrier identities were checked.

• Shared protocol schema/generated views, compiled C bridge ABI, 64 MiB aggregate tool-output policy, SJLJ artifact guard, patch-review freshness, Postmaster source lock, focused libc/runtime checks, Rust formatting and documentation policy passed.

• assets verify-committed passed using the existing core-only build flag; full extension catalog validation remains the default outside that explicitly scoped path.

• Fresh Rust WASIX on the new portable guest and strict AOT: 18/18 runtime-smoke and 7/7 PostgreSQL regression tests passed. These exercise memory/directory storage, sync/async callback error and panic recovery/reuse, prepared parameters, INSERT/savepoints, rollback/cancellation, directory ownership/reopen, backup/restore and recovery after an abruptly killed child. These are focused cases, not exhaustive crash qualification. Rust COPY-specific coverage was not available in these suites; actual browser COPY coverage is listed above.

Boundaries and follow-up

• Windows/MSVC is a merge blocker. Job 101905963703 reaches an explicit panic in wasmer-vm 7.2.1's MSVC throw() implementation during savepoints_error_recovery_and_indexed_updates, exit 0xe06d7363. Official v7.4.0 and current upstream main still contain the unsupported MSVC exception stubs; an ordinary dependency upgrade is not a fix. Genuine engine exception support or a separately qualified compatible guest is required. Skipping tests, catching the panic or silently falling back to Native would not preserve the guest recovery/isolation contract.

• The retained repeat experiment is held, not promoted: 184 correctness/settings-checked children show wide temporary INSERT 41.5% / 33.5% faster and logged wide 29.0% / 20.7% faster (memory/directory). A separate 72-child targeted+A/A recheck resolved the preconstructed-directory warning but retained a prepared-directory warning of approximately 6–7% against both baseline controls, predominantly in transaction body. Shared-host interference remains a limitation. Exact candidate artifacts and original baselines remain retained; no experimental framework is staged, and neither durability nor fill-chunk tuning is a demonstrated remedy for the warning.

• SQL statement_timeout is not a reliable CPU-bound deadline in current Rust WASIX. Actual baseline/candidate probes expose inherited libc timer-field and Rust unit/one-shot/delivery gaps. Native timeout/reuse and injected interrupt tests do not prove Wasm delivery. Public Rust/usage docs now explicitly distinguish accepting the GUC or timing out a caller future from stopping guest execution; issue Follow up patch consolidation: signal semantics, performance profiling, and qualification #201 records the coordinated remedy.

• No release-ready or blanket non-regression claim. The initial 324-child strict-profile comparison and final 220-child identity screen are separate evidence sets; the later one does not erase the earlier warnings. All accepted directory reports actually verify fsync/synchronous_commit/full_page_writes=on. No Native durability-change timing is inferred from the WASIX screen.

• Inherited Postmaster POSIX signal masks are not implemented. Pinned pthread_sigmask returns success/no-op, sigpending returns EINVAL, and handler sa_mask is ignored. The capability inventory now says unsupported. The macro fix does not repair delivery masking; correct remediation needs coordinated Wasmer/libc mask/pending delivery semantics and PostgreSQL stress tests. This is a high-priority known limitation, not an observed corruption or new-regression claim.

• Four completed prior/final ABBA profile captures collected 45,283 samples on unchanged strict guest/AOT payloads, using existing public Wasmer function extents for diagnostic address maps. The wide workload's repeat(chr(...),96) makes 96 one-byte copies per row; imported-copy self samples account for 24–27%, while direct statistics-counter samples are about 0.45%. This supports a bounded general repeat experiment, not causal attribution of the +5.1% warning. No new optimization is promoted. The user-authorized temporary profiling setting was restored to its original perf_event_paranoid=4 immediately after capture; no persistent setting changed.

• Hosted Native builds and the portable WASIX, Postmaster and iOS-extension producers passed at parent 51789da. The later 8e45be7 head also passed Native/extension and portable WASIX producers plus Linux/macOS AOT, but exposed the known formatting/root failures and Windows exception blocker. That parent run was cancelled by the push only after material producers finished, with app/E2E work remaining. 0f4a996 must qualify in its own run 34177975360; parent results are not relabelled as final-head results.

• Native now preserves fsync=on by default; durable-write latency can increase. Functional reopen tests still do not establish crash durability. No WASIX durability relaxation is introduced.

• Guest and hosts/AOT must be rebuilt together. Old profile manifests cannot be relabelled; terminal guest failures cannot be treated as recoverable SQL errors.

• Browser host runtime code was built before mail-header normalization. Every patch diff is byte-identical before/after that metadata-only step; the original build marker is retained and a local before/after equivalence receipt records both identities. This is not labelled an exact-release artifact.

• Local Moon affected-project resolution, SDK contracts and graph unit tests pass with the repository-pinned toolchain.

• Broad platform/ICU, extension, hostile-extension/resource, crash and release qualification remain separate work.

The complete prioritized remainder is issue #201: inherited Postmaster signal-mask semantics, a profile-supported repeat experiment and unresolved wide-INSERT warning, remaining platform/runtime qualification and upstream review. Optional compiler/ICU/ownership ideas are conditional research, not mandatory features. Rejected CRC/lazy-globals/unsafe shortcuts are not automatically revived.

See docs/internal/PATCH_CONSOLIDATION_20260907.md for the consumer implications and explicit retirement table. The original research state is preserved in a separate local recovery commit and is not staged here.

[vercel]

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

Project	Deployment	Actions	Updated
oliphaunt-docs	Ready	Preview	Sep 8, 2026 8:56pm UTC