Port pinned CI suite selection, setup, and vendor execution

Parent #66; continues #703/#722 after the native outcome and child-lifetime work (#723).

Crash and core processing checkpoint (2026-09-21)

Full logical Bun parity remains incomplete. 2e2b71edde is committed as Chris Breuer without coauthors and pushed to origin/main.

Each Home child in home test --bun-corpus-ci now follows the pinned crash path, including setup, primary and vendor phases. A nonzero child with live remap waits 500 milliseconds, fetches /traces, and retains parse failures, remap failures, mapped stacks and fetch failures. PID and aggregate trace/core counts are durable; a trace or new core cannot count as a passing file.

The exact coredump-upload default is BUILDKITE=true on Linux. The coordinator probes kernel.core_pattern, rejects empty, relative and piped patterns, snapshots its directory around each child, warns when a signaled main PID lacks its core, and invokes GDB per new core with the original 240-second bound. Only termination, current-thread and backtrace lines are retained; missing or failed GDB remains an attached diagnostic.

A real failing Home test retrieves one mapped trace from a live service. Controls cover all trace shapes/truthiness, core applicability/differencing and exact GDB behavior. The first broad run retained 26/27 because its control JSON encoded a literal \n while asserting a newline; fixing that fixture changed no production path. The final build passes 23/23 steps and 27/27 harness tests, followed by a fresh direct 27/27 run on the rebased published tree. Elysia remains 208 discovered entries, 129 selected files and one original exclusion (9,665 bytes). Validators, formatting, Pickier, shell/whitespace and setup-sync checks pass. No original corpus-case credit is claimed.

Immutable planning, pinned ordering, real checkout/setup/remap/primary/vendor controls, durable errors and owned shutdown remain intact. Actual Buildkite Linux core generation/GDB and encrypted upload, native Linux Docker execution, complete primary execution, #725, and remaining runtime/platform gaps stay open. #724/#66 remain open; this is not a full-suite or 100% parity claim.

Implementation and verification boundary

Current publication / remaining work (2026-09-20)

Full logical Bun parity remains incomplete. Native phase/context source 64481fe09 is committed as Chris Breuer without coauthors on origin/main. Home SHA256 d74f86c7953200c0434f4e8a89894dd21626359c505935faf50b662e6d31be50; Zig 0.17.0-dev.2163+89ff10d56.

Native vendor checkout and install/build are now separate phases. Checkout verifies the pinned tag without installation hooks; later preparation requires the recorded revision and refuses a changed checkout before hooks. Native child launches carry live remap and ready Docker settings through primary/vendor tests and package preparation. The default CI coordinator still needs to connect selection, services, setup and execution in the original order.

23/23 build steps, 21/21 native harness tests, seven launcher units, four service/child-lifetime units and 20 validator controls pass. Private CLI checks retain checkout-only success, changed-revision rejection with later operations unstarted, and guarded install/build success. Five real pinned remap HTTP checks pass. The first harness result (20 pass, one fail) is retained: the new private fixture lacked the required bunfig.toml. Its config correction changes no original corpus test or production launch behavior; the original queued runtime batch stopped before execution.

Fresh Elysia checkout completes five operations, then native install/build completes four, at unchanged deadlines. All 283 resulting source hashes match the previously verified pinned setup. All 129 selected original Elysia files complete once with 1,516 registered passes, zero failures, registered skips or TODOs. The original WebSocket exclusion stays separate. This is the first vendor execution on the new build; the prior binary's port-3000 failure remains retained and #726 stays open for the separate bind-error compatibility audit. Setup/service outcomes earn zero case credit, and repeated regressions are not new distinct coverage.

Remaining: #724 default full CI orchestration and complete applicable primary/vendor execution; actual Linux Docker startup and native platform verification; crash processing; #725 optional legacy DuckDB; independent native build ownership and remaining runtime features. The 50 pinned setup inputs still match; this batch did not repeat primary root/test installs. No full-suite or 100% parity claim.

Implementation and verification boundary · Retained evidence

Publication commit: b83c54f4d.

The 4,754-path discovery inventory is pre-exclusion. Home's current corpus_runner.runGateWithOptions selects that inventory (with optional start/limit) and runs each path. The pinned 4982b91e3702094330f3be3883354c52b8c01323 CI runner also applies platform/build modifiers and expectations, performs root/test package installs and registry/service setup, and adds vendor suites. Therefore a run of the 4,754 paths alone cannot prove completion of the upstream CI suite.

Source evidence: scripts/runner.node.mjs in the pinned Bun checkout, especially getTestExpectations, getTestModifiers, getRelevantTests, getVendorTests, and runTests. The pinned vendor manifest currently adds Elysia tag 1.4.28, including its explicit ws-connection exclusion. Home's copied vendor.json is identical. Home's expectations.txt differs by four removed historical exclusions (bunx and three HTTP/2 files); those additional coverage decisions must stay visible, not silently replace the pinned reference or hide failures. All 4,754 discovered paths exist and the 110 declared mirrored test dependencies are present, but this is not proof that full CI setup is equivalent.

Required work:

1. Port the actual pinned modifier/expectation/filter/shard selection contract into Home. Preserve the original reference metadata and record every exclusion with its reason/modifiers. Separately retain any deliberate additional Home coverage. Excluded, skipped, TODO, unstarted, and unsupported entries never count as passing implementation credit.
2. Audit and port the real root/test installation and registry/service preparation, including environment, configuration discovery, ownership and cleanup. Execute with Home's native commands; do not substitute system Bun, fake services, dummy configs, or reduced dependency graphs. Keep setup failures explicit in the durable journal.
3. Prepare the pinned Elysia vendor checkout and dependencies, discover its actual tests and applicable original skips, and execute it in Home with the correct project/configuration context. Retain resolved revision, source/snapshot integrity, selected inventory and complete outcomes, alongside the primary corpus.
4. Verify selection against the unchanged pinned runner across supported platform/build contexts, and verify production setup/execution. Use the original deadlines and one attempt per unchanged build; no retries or deadline extensions to obtain green results.
5. Execute the complete applicable suite, including vendor tests, and preserve all failures for feature porting. Independent native build ownership, default CLI fallback removal, native Windows/Linux/ASAN execution, and unresolved runtime features remain part of #66; passing this driver comparison alone is not full parity.

This issue is a concrete remaining prerequisite, not a redefinition of #66 as only runner infrastructure or a smaller test subset.

Native services and explicit vendor preparation checkpoint (2026-09-08)

Full logical Bun parity remains incomplete. Published as Chris Breuer without coauthors to origin/main: native service/preparation source a2e2863fc, verified evidence c9fc9bf63. Home SHA256 91d6b1024f9d4c02c6b7f9089b476d8a0646294f4ed7392e5e3ce549e52db763.

Native service lifetime and explicit vendor preparation are implemented. home test --bun-corpus-remap launches the original installed remap service with its original arguments and 5,000 ms startup deadline, drains output throughout its lifetime, and owns shutdown until parent stdin closes or the service exits unexpectedly. home test --bun-corpus-prepare-vendor elysia selects and journals clone/fetch/checkout/revision/install/build, retains separate Git/Home executable identities, and preserves the original 180-second Git/install and 60-second build deadlines. A failed install leaves build visibly unstarted. These operations earn zero registered-case or passing-process credit.

23/23 native build steps and 21/21 harness tests pass, along with four service lifetime units, four journal units, 18 validator controls and five private CLI controls. The private CLI evidence retains invalid readiness, the five-second deadline after output EOF, unexpected service exit with parent stdin open, install failure and build failure. The real pinned remap service passes five HTTP protocol checks through the native command, with complete captures and unchanged dependency hashes. Fresh native Elysia preparation completes all seven steps in 49.8 seconds; its 283 resulting source hashes match the earlier independently verified setup/lockfile normalization. Formatting, targeted Pickier and 50 pinned setup-input checks pass.

The latest complete Elysia run is not green: all 129 selected files complete once with 1,515 registered passes, one failure, zero skips/TODOs. handle distinct union in test/validator/response.test.ts calls .listen(3000); the unrelated PredictHQ development server (PID 28404 at diagnosis) owns that port. Pinned Bun fails the same unchanged case in a separate one-file diagnostic (27 passes, one failure). The listener and original tests were not changed, and no retry or alternate port was used to obtain a pass. Both full-run source integrity and the original selection match. The original WebSocket exclusion remains separate. The first diagnostic verifier expected the same error literal; a separate retained-capture analysis confirms the conflict without rerunning either runtime. A private error-object comparison additionally found message/errno/address/port differences, now tracked in #726; Home's existing preservation of real OS failures must be retained while auditing compatibility.

572 committed compressed artifacts have verified decompressed hashes. Restored journals reproduce successful service/preparation, failed private controls, the failed complete Elysia run and failed pinned-Bun diagnostic separately. The first build was stopped below the unchanged 1 GB critical disk floor after macOS allocated more swap; verification did not run. A cleanup preflight and refused admission are retained too. Corrected cleanup preserved protected binaries and linked native inputs; the completed rebuild required 5 GB free to start. Removed 6,083,408,581 logical bytes of generated caches/dependencies, old vendor data and unused build outputs. Source and other active work are preserved.

Still open: connect these operations with the original CI ordering and selection gates, Docker coordination, live service settings and default unified coordinator (#724); free the original fixed-port prerequisite before claiming a successful new Elysia run; resolve the missing optional legacy DuckDB and earlier lifecycle failures (#725); audit bind-error compatibility (#726); execute the whole applicable primary/vendor suite and finish remaining runtime features, native Linux/Windows/ASAN verification and independent native ownership. Earlier root/test setup process success and the separate 18-case N-API DuckDB run remain evidence on their recorded earlier binary; they are not new execution on this one.

API and scope · Native service/preparation evidence · Earlier native setup evidence

Earlier native root/test setup checkpoint (2026-09-08)

Full logical Bun parity remains incomplete. Published as Chris Breuer without coauthors to origin/main: native setup source 8607e68dd, verified checkpoint 478cdf5c2. Home SHA256 65c97173d344a4a2cebe5f2db98c2929b34f94492a4f5929e904475f5444dc6e.

Native root/test setup execution is implemented. home test --bun-corpus-setup checks native platform expectations, records all 50 pinned input hashes and both original install steps before execution, then attempts both installs with separate original 180-second deadlines. Captures, input integrity and temporary storage cleanup are durable outcomes. Installs earn zero registered-case or passing-process credit. The shared setup launcher also implements and exercises the original 60-second build operation.

A real runtime-ownership gap was fixed: setup initially omitted the node command alias, so private lifecycle controls resolved the installed system Node and failed. Setup now gives bun, node and home owned aliases to the same Home executable, outside its fresh package cache. The first 18/19 harness result and separate diagnostic are retained. The corrected native batch passes 23/23 build steps, 19/19 harness tests, six launcher units, 14 validator controls and four actual setup CLI controls. Positive/negative controls retain both setup outcomes; platform mismatch and extra arguments fail before installation. Targeted Pickier and whitespace checks pass.

The first clean original setup execution through this command passes the original process gate: both installs exit 0 within their unchanged deadlines (234 root packages reported in 3.0 seconds; 2,898 test packages in 169.2 seconds; total command wall time 179.7 seconds). Both captures reach verified EOF, all 50 pinned inputs stay unchanged and temporary setup storage is removed. This does not establish complete dependency or feature support. The audit finds no missing required package manifests, but optional legacy duckdb@1.3.1 is absent after its compilation attempt. The binding-presence control fails, and its unchanged original fixture exits before registration on ABI 147. Its one successful process-exit record earns zero implementation or registered-case credit. The separate unchanged @duckdb/node-api file passes 18 registered cases, with source and N-API binding hashes unchanged. #725 stays open; no original test source, install deadline, dependency declaration or ABI gate was changed to obtain these outcomes.

Complete prepared Elysia regression on the new binary: all 129 selected files complete once with 1,516 registered passes, zero failures, zero registered skips/TODOs. All 283 post-install source hashes and complete captures/JUnit reports validate. The original WebSocket exclusion stays separate. Previous vendor setup is reused and repeated counts are not added as new distinct coverage. 490 committed compressed artifacts have verified decompressed hashes; restored setup, positive/negative controls, original DuckDB and vendor reports reproduce their distinct results. Removed 2,693,367,959 bytes of owned generated data while preserving the current executable, harness and parser archive.

Still open: resolve legacy DuckDB’s missing optional binding and earlier deadline failures (#725), connect services and complete vendor preparation to the default unified CI coordinator (#724), execute the whole applicable primary/vendor suite, and finish runtime features, native Linux/Windows/ASAN verification and independent native ownership. Earlier native platform/selection checkpoints remain valid historical evidence; this setup component is not a full CI or parity completion claim.

API and scope · Complete checkpoint evidence · Earlier native platform checkpoint · Earlier original setup failure

Earlier original CI graph and install checkpoint (2026-09-08)

100% logical Bun parity is not complete. Published as Chris Breuer without coauthors to origin/main: original setup graph 47976822e, evidence 764208102. Runtime remains Home SHA256 19e41223c7fc2d2dcda763470fcc84570900db08b4d547340bdae5286d42bc98 from the preceding native build.

The original CI dependency graph is restored. The previous React-only root is replaced with the pinned root's 14 development dependencies and both complete type-package workspaces. The sync reads Git objects at Bun 4982b91e3702094330f3be3883354c52b8c01323; all 47 root/workspace files have matching Git blob IDs and SHA256 provenance. The normal corpus sync now includes these inputs, and a separate --check mode verifies them without writing or installing. Original source formatting is preserved, including the upstream bunfig comment's trailing space.

Clean native root install succeeds: 234 packages in 12.9 seconds. The clean test install fails at its original 180-second deadline. The active lifecycle is DuckDB's native source compilation. All 50 root/workspace/test-setup input files remain unchanged. Both installs use fresh caches and native Home aliases; 945,163,206 bytes of owned generated dependency files were removed first. All 14 root and 110 test dependency manifests are present afterward, but the DuckDB native binding is missing; manifest presence is not installation success. The service startup check remained behind its failed install gate and did not execute.

New linked prerequisite: #725. DuckDB 1.3.1's requested ABI-147 macOS arm64 prebuilt returns HTTP 404, so the unchanged lifecycle falls back to compiling C++. Home and pinned Bun report identical Node 26.3.0 / ABI 147 metadata. Isolated fresh lifecycle diagnostics reproduce the same 404 and actual source compilation under both runtimes; both hit 180 seconds (Home 180.009, Bun 180.013). This is a shared prerequisite failure, not a Home-only ABI/version mismatch. No lifecycle suppression, dependency edits, ABI spoofing, deadline extension or warmed retry is used to claim success.

81 committed evidence artifacts preserve all original inputs, complete outputs and both lifecycle controls; decompressed hashes and Git tracking are verified. Targeted Pickier passes after correcting a documentation blank-line warning. Zero new original test-case pass credit is claimed. Native host checks, clean test-install completion, service and setup orchestration, full applicable primary execution, remaining runtime features and independent native ownership remain open under #724/#66. The preceding Elysia checkpoint remains 129 completed files / 1,516 passing cases, separately verified rather than combined into a full-suite pass claim.

Setup and failure evidence · API and verification scope

Earlier prepared Elysia execution checkpoint (2026-09-08)

Published as Chris Breuer without coauthors to origin/main: implementation 9d6815afa, evidence 176f685f7. Home SHA256 19e41223c7fc2d2dcda763470fcc84570900db08b4d547340bdae5286d42bc98.

Complete pinned-selected Elysia execution: 129/129 files completed, 1,516 registered cases passed, zero failed, zero registered skips, zero TODOs. The run used no positional filter or start/limit and one attempt with original per-file deadlines. The selection matches pinned Bun exactly, including the separately recorded original ws/connection.test.ts exclusion; that excluded file earns no passing or skipped-case credit. All 129 captures reached EOF. Raw output and JUnit hashes, registered cases, selection completeness, all 283 post-install vendor source hashes and the Home executable hash validate.

The production home test --bun-corpus-prepared-vendor elysia command verifies Git HEAD against the manifest tag and executes each selected vendor file in its own project's cwd with ordinary config discovery, forced vendor test mode, correct validation paths and serial environment. Ordinary failures remain recorded and later files continue. This command explicitly requires prior checkout/install/build and journals setup_performed: false. The preceding Home 716f39bd… build performed the independently verified native install/build; the new 19e41223… build executed the suite. This is not a fresh end-to-end CI setup claim.

23/23 build steps, 17/17 native harness tests and 11/11 validator controls pass. Private controls exercise project config, custom test paths/extensions, serial environment and a real assertion failure followed by a successful file. Targeted Pickier and whitespace checks pass. 409 committed compressed artifacts have verified decompressed hashes; restoring the archived reports reproduces all 129 complete outcomes and 1,516 passing cases. Cleanup removed another 371,467,160 bytes of obsolete owned generated objects, preserving source work.

API and verification boundary · Complete execution evidence

Full Bun parity remains incomplete; #724 stays open. Remaining work includes native host detection and expected-agent checks, completion of original root/test installation (see the later graph-restoration checkpoint above), services, complete vendor preparation and unified setup/execution journaling, the complete applicable primary suite, native Linux/Windows/ASAN verification and remaining runtime/native ownership work. Earlier selection-only and 402-case checkpoints remain historical evidence; their counts are not combined into a whole-suite pass claim.

Earlier native selection and Elysia setup checkpoint (2026-09-08)

Published as Chris Breuer without coauthors on origin/main: source 4af7114c9, checkpoint 5b7c6cad3. Home SHA256 716f39bd45568c3cab84f080aafc32d9ec553452ba5d5945ec348e289e40bbc2.

• 23/23 build steps; 15/15 native corpus harness tests; 14/14 selection/discovery units; 10/10 outcome-validator controls; targeted Pickier pass. The selection API now implements executable/platform modifiers, original expectation parsing and OR matching, includes/excludes, Node-only and macOS x64 CI classification, filter/shard precedence, and stable modified-file priority. It rejects Home-only exclusions before launching children. A durable selection event partitions the whole inventory, retains original rule metadata and additional Home coverage, and exposes partial ranges separately from execution outcomes.
• 304,256 selection decisions match the unchanged pinned runner: 4,754 files across eight platform/build contexts and eight option combinations. This is a native selection comparison, not corpus case passes or native execution on all platforms. The macOS arm64 context selects 4,676 primary files, including four deliberately unexcluded upstream files; 78 exclusions remain explicit.
• Pinned Elysia discovery matches exactly: tag 1.4.28, revision 56310be9617b826f862c985eae95ae823d95f097, all 208 recursive entries in order, 129 selected files, original ws/connection.test.ts exclusion. Patterns use Home's actual ECMAScript regex engine and preserve the upstream explicit-extension precedence.
• Native Elysia install and build succeed, at the original 180-second install and 60-second build deadlines. Home installs 218 packages and runs the unchanged build script. All command aliases (bun, node, home) identify the same native Home binary. The only changed tracked file is bun.lock: installation synchronizes the stale root exact-mirror peer range with the existing package manifest. A separate pinned Bun install produces the identical resulting lockfile byte for byte. All other 282 tracked vendor files remain unchanged. The original strict integrity failure and its lock diff remain retained alongside this control; no tests, snapshots, or lockfiles were manually edited to obtain a pass.
• First failures are retained: two native regex interpreter traps (normal JSC startup was missing), a private execution-selection fixture missing its inventory manifest, and reference-driver Python/shard-accounting errors. The corrected native batch and comparisons pass. 54 committed compressed artifacts have verified decompressed hashes. Cleanup removed 2,159,035,152 bytes of obsolete outputs from this owned worktree and preserved the installed Home, newest harness, and native parser archive.

Still open: the default CLI does not yet run complete CI orchestration. Callers currently supply selection context. Native host detection and expected-agent checks, the original root/test dependency and workspace graph, registry/services, complete vendor checkout/preparation/execution, and the unified setup/outcome journal remain to be connected. The current mirrored root package is React-only and is not equivalent to the original CI root. None of the 129 vendor files has yet been executed through the production corpus driver. No new corpus-case pass credit is claimed for this checkpoint. Full applicable primary/vendor execution and all remaining runtime features remain part of #66; this issue stays open.

Selection API and verification boundary · First evidence · Verified evidence