gstack/docs/gstack-2/TEST-EVIDENCE.md

GStack 2.0 test evidence

Candidate checkpoint — 2026-07-20

These results describe the working tree at the documentation checkpoint. They are deliberately narrower than the P0 release matrix. Commands that need a provider account, signed physical app, native OS, or live host are not marked pass from deterministic, offline, or filesystem-only evidence.

Command / probe	Observed result	What it proves / does not prove
Earlier focused macOS bun test test/gstack2-*.test.ts candidate run	Exit 0: 136 pass / 0 fail, 1,128 assertions across 15 files. Log: /tmp/gstack2-test-command-candidate-final2.log.	Historical focused checkpoint retained rather than overwritten.
Current macOS direct bun test --timeout 60000 test/gstack2-*.test.ts	Exit 0: 218 pass / 0 fail, 2,229 assertions across 20 files.	Current routing, runtime, privacy, installer, upgrade, parity, browser-provider setup, and adversarial-harness surface is green. The test:gstack2 wrapper's generated-cleanliness precheck remains a commit-time gate, so this row records the direct underlying suite rather than relabeling that wrapper.
bun test --timeout 30000 test/gstack2-skills.test.ts test/gstack2-skills-routing.test.ts after regeneration	Exit 0: 3 pass / 0 fail, 81 assertions.	The pinned corpus/parity test and both 25-scenario structured-routing tests are green. This remains structural/fixture evidence, not specialist live execution.
bun run scripts/gstack2/run-parity.ts, 2026-07-17 rerun	Exit 0: 4,681 checks passed; 55 modules, 16 sections, 25 scenarios, 16 regressions, 78 assets.	Current source/render/provenance/contract/asset/fixture parity is green. It is deterministic parity, not live-host behavior.
Earlier regenerated structural parity checkpoint	Exit 0: 2,403 checks passed with the then-current 55/16/25/16/45 inventory.	Historical candidate checkpoint before later thin-prelude and asset coverage; superseded by the current 4,681-check rerun.
Earlier combined runtime/parity run before regeneration	Exit 1: 20 pass / 1 fail, 161 assertions. All 18 runtime tests and both routing tests passed; parity hit the default 5s timeout. A separate long-timeout attempt reported 29 stale-generation checks.	Preserved as history: generator inputs had changed. The post-regeneration rows above supersede the parity failure, not the runtime results.
bun run scripts/gstack2/semantic-parity.ts and bun test test/gstack2-semantic-parity.test.ts	Deterministic corpus green: 295 checks, 14 suites, 15 executions, all 15 dimensions, 16 carved sections, and nine authority-policy unit cases, including unsupported numeric claims.	Exact source-body and semantic-signature evidence is the primary reproducible oracle. The policy units consume hand-authored semantic operation envelopes; they are deterministic policy evidence, not behavioral-adversarial proof.
Retained Claude Haiku live semantic transcripts	Three retained samples, all classified REGRESSION. Two samples used now-obsolete visible-wrapper prompts. The current office-hours sample also penalizes independently sampled omissions even though those details remain byte-preserved in the candidate source.	Live-model sampling is supplemental and currently not green. These results must not be retried until favorable, cherry-picked, or used to overrule deterministic source loss. See SEMANTIC-PARITY.md.
Installed-host adversarial v1	FAILED. The immutable v1 slash-invocation artifact is retained.	Unfavorable activation/classifier evidence; not a behavioral pass. See eval overview.
Installed-host adversarial v2	FAILED. QA passed, while debug, review, and ship were false negatives from the v2 read-only-Git warning classifier.	All four dispatchers activated, but the top-level run failed and must stay failed. Immutable artifact: 2026-07-17T04-09-01-809Z-3d23a270.json, SHA-256 7ab15ea575cb9a634b7d00212dd9d74902b1188281ae6a503a32ccf382facbf5.
Installed-host adversarial v3 offline harness	18 pass / 0 fail, 111 assertions.	The classifier fix and offline harness are covered. This deterministic result does not override the failed paid live run.
Installed-host adversarial v3 paid live run	FAILED 3/4. Debug, QA, and ship passed; review failed compound inspection. This was the one-shot run, with zero retries.	Unfavorable evidence is immutable and must not be retried or relabeled. Artifact: 2026-07-17T19-48-45Z-v3-live-gpt-5-4.json, SHA-256 fcffdf2b0ee7bb9ac1351e246546af2cd352779bda7b1f8dc4a08f51fc66ef2f.
Six-skill catalog measurement after regeneration	Six names/descriptions total 982 characters, about 246 four-character token-equivalents; baseline correctly parsed catalog was about 1,100.	77.6% reduction, above the 75% gate. Re-measure if frontmatter changes. The buggy 4,214 baseline estimate is not used.
bun test ios-qa/daemon/test	95 pass / 0 fail, 229 assertions.	Covers daemon regressions including malformed device JSON, hardware-UDID/CoreDevice selection, bounded proxy timeout, and expected-bundle mutation header. It is not a signed-app live pass.
Focused DebugBridge/template build tests	33 pass / 0 fail in the candidate run, including Swift debug compilation/XCTest and Release symbol absence.	Static/build evidence for debug-only bridge wiring; still not an installed physical-app journey.
Earlier physical-iOS setup attempts	Incomplete, not passes: the preflight recorded 9 pass / 0 fail / 1 deploy check skipped and 29 assertions; one target returned signing_unavailable; the legacy iPhone10,6 target returned device_not_wired / CoreDevice error 1011; a later authorized phone reached tunnel setup but its first session-acquire socket closed.	Immutable historical setup/partial evidence. The later pass does not erase or relabel these attempts.
Physical-device smoke, 2026-07-20	PASS: 12/12 harness checks and all five required live iterations. Release symbols were absent; signing, safe in-place install, launch, CoreDevice bootstrap, boot-token rotation, five session acquire/release cycles, ten screenshots, accessibility reads, coordinate taps, bundle checks, state cleanup, tunnel shutdown, and temporary-workspace cleanup passed.	Closes the physical-iPhone evidence layer for the tested wired paired iPhone17,1; it is not a claim about every iPhone or iOS release. Redacted artifact: ios-physical-device-2026-07-20T17-49-19-302Z.json.
Context.dev contract (gstack2-runtime-context.test.ts)	22 pass / 0 fail, 139 assertions.	Persists explicit host/local-browser/none choices without consent, rejects private/credential URLs and request material plus private DNS, proves zero lookup/fetch before mode+consent, validates documented endpoint paths and exact failure taxonomy, and makes search typed unsupported without network.
gstack context smoke --url https://www.context.dev --json	PASS: the official /web/scrape/markdown endpoint returned ok: true and credit metadata.	The verified key entered through protected stdin/environment worked. The temporary secrets file was mode 0600, the isolated home was removed, and the safe output did not contain the key. Committed redacted artifact: evals/context-dev/live-smoke-2026-07-17.json.
Standard installer matrix	PASS: 510/510 checks, 18 install cases, two removal cases, skills CLI 1.5.19.	Project/global installs pass for seven hosts, including Kimi Code CLI through the standard .agents/skills path; selected-skill and opt-in compatibility-alias cases, copies, and hashes pass. This remains installer/filesystem evidence. Committed artifact: evals/installation/install-matrix.json.
Standard Agent Skills install + actual runtime-absent Codex invocation	PASS. Root --list returned exactly six. The selected source was time-attack/gstack/skills --skill qa; despite skills 1.5.19's pre-filter display counting hidden aliases, exactly one skill (qa) installed byte-identically. Codex exited 0 with gstack absent, reported NEEDS_SETUP and one approval prompt, changed no files, and created no runtime/browser/external-service activity.	Closes the actual-host runtime-absent judgment gate for Codex without overstating seven-host UI coverage. Artifact: standard-codex-runtime-absent-2026-07-17.json.
bun test --timeout 30000 test/gstack2-runtime-install.test.ts	Exit 0: 41 pass / 0 fail.	Managed allowlist, hashes, spaces, source/internal-link rejection, production-only frozen dependencies, deterministic exact Sharp/ngrok platform closure, native-load rollback smoke, atomic browser-choice rollback/recovery, provider-aware stable launchers, wrapper neutrality, and state-preserving uninstall pass.
bun test --timeout 30000 test/gstack2-runtime-setup-ux.test.ts	Exit 0: 29 pass / 0 fail.	Explicit managed/installed/later browser selection fails closed before network, local options do not mutate, signed previews remain provider-identical across same- and cross-release capability additions, developer-source provider switches replace the exact retained set, visible extension Chromium remains managed-only, and launcher/doctor paths honor the persisted selection. This is deterministic local evidence, not native Windows or every installed Chrome-family build.
Deterministic clean macOS arm64 managed runtime bundle audit	110 components, 1,829 files, 450,044,315 bytes, 50 capability launchers.	This is a platform-specific bundle measurement, not a universal byte count; platform-native package payloads differ. Setup includes the Sharp/ngrok closure and excludes the development-only Claude Agent SDK. The Hugging Face sidecar is outside the bundle and its package is development-only, so production setup installs neither its inference runtime nor model weights; the L4 capability reports unavailable. The standard skill installer remains Markdown-only. Committed artifact: evals/runtime-bundle/darwin-arm64.json; reproduce with bun run scripts/gstack2/audit-runtime-bundle.ts --output evals/runtime-bundle/darwin-arm64.json.
Earlier declared Linux Dev Container plus bun run test:gstack2 inside it	Exit 0: 136 pass / 0 fail, 1,127 assertions across 15 files. Log: /tmp/gstack2-devcontainer-gate-candidate-final4.log.	Historical container checkpoint retained rather than overwritten; the current 150-test container result is recorded in the native-CI row below.
scripts/gstack2/runtime-install-smoke.sh in the clean Linux arm64 container	Pass: production-only frozen dependencies installed with the development Agent SDK and Hugging Face/ONNX runtime absent; the managed Anthropic SDK, Sharp, and ngrok imports passed; prebuilt capabilities rebuilt; setup/doctor/version/design/PDF passed; a local-browser journey and Sharp full-page screenshot passed; uninstall preserved state.	Proves a source copy with spaces can build and complete the managed runtime lifecycle without Git history, an executable local-model stack, or Darwin-only iOS artifacts. It is not native Windows evidence.
Runtime lifecycle and external-effect matrix	Pass: real filesystem/subprocess tests cover clean install/uninstall, paths with spaces, source symlinks and internal-link rejection, macOS read-only reporting, interrupted-pointer rollback, crash-journal repair, last-known-good launcher recovery, and an actual local Git push that executes at most once across resume.	Closes the named local filesystem/recovery/idempotency gates. Aggregate cancellation evidence is recorded separately below.
bun test design/test	101 pass / 0 fail, 381 assertions.	The retained design suite is green and is also included in the later uninterrupted broad pass.
bun test make-pdf/test plus combined-fixture render	189 pass / 0 fail, 398 assertions; four PDF pages rendered to PNG and visually inspected with no detected layout defect.	Retained strict PDF tests and a live internal render are green on this macOS host. Cross-platform visual equivalence remains a separate platform claim.
Diagram suites	51 pass / 0 fail / 1 skip, 120 assertions; the opt-in paid lane recorded two skips.	Offline Mermaid/SVG/PNG/Excalidraw coverage is green. Skipped paid-provider cases are not live evidence.
Isolated local-browser journey plus stop regression	Navigation, snapshot, screenshot, and status passed; stop returned success and left no observed process leak. browse/test/stop-ack-before-shutdown.test.ts is 2 pass / 0 fail.	Provides one real local Chromium journey and a focused regression for acknowledging stop/restart before delayed shutdown. The complete browser suite is included in the later broad pass.
Live local-browser SIGINT aggregate	PASS. An isolated local-file journey completed goto, status, and snapshot; SIGINT produced an in-memory authenticated 503, reduced five owned processes to zero, and left no listener, root credential state, credential-shaped file, or device session. The unrelated browser daemon survived. Focused regression: 34 pass / 0 fail, 62 assertions.	Closes the requested aggregate process/listener/state/credential cleanup gate without claiming device interaction. Artifact: evals/browser/cancellation-2026-07-17.json.
Native CI run 29615621805, commit a8a5fa1a	PASS: installer 470/470; macOS 150/0/1,189; Ubuntu 150/0/1,189; native Windows 150/0/1,145; Dev Container 150/0/1,188. All native installer-discovery steps found exactly six skills.	Complete native matrix evidence. Sanitized artifact: evals/ci/native-2026-07-17.json. Runs 29608904265, 29611504979, 29611757175, 29612056517 (cancelled), 29613668419, 29614448170, and 29614899434 remain superseded diagnostics.
Earlier bun test strict singleton checkpoint	Exit 0: 6,240 pass / 226 expected skips / 0 fail, 25,449 assertions across 383 files. Log: /tmp/gstack2-full-singleton-final-committed.log.	Historical broad checkpoint retained rather than overwritten.
Current bun test through the strict singleton runner	Exit 0: 6,255 pass / 226 expected skips / 0 fail, 25,509 assertions; 384/384 shard headers and terminal single-file summaries.	One uninterrupted current-head macOS broad pass. Expected skips are provider, credential, Poppler-environment, paid, or model-sidecar gates declared by their tests; external/live evidence remains separate.
Earlier local bun run test:windows singleton checkpoint	Exit 0: 2,815 pass / 57 expected skips / 0 fail, 8,586 assertions across 213 files. Log: /tmp/gstack2-windows-singleton-final-committed.log.	Historical curated-lane checkpoint retained rather than overwritten.
Current local bun run test:windows through singleton shards	Exit 0: 2,829 pass / 57 expected skips / 0 fail, 8,648 assertions; all 214 selected files completed.	The current curated Windows-safe subset is locally green. Native Windows is independently green in run 29615621805.
Forbidden production-scope audit	Production dependencies and the 110-component managed bundle contain no cloud-browser provider, Hugging Face/ONNX inference runtime, model weights, or alternative physical-iOS backend. Transformers remains development-only for retained tests; CoreDevice/devicectl is the sole physical-iOS path.	Deterministic dependency/bundle/backend evidence. It does not substitute for native-platform, signed-device, or live-provider behavior.

Required final command ledger

Append exact exit codes, aggregate counts, environment/version metadata, and artifact/log locations for all of the following before changing status to DONE:

bun run gen:gstack2
bun run test:gstack2
bun run scripts/gstack2/run-parity.ts
bun run scripts/gstack2/semantic-parity.ts
bun run build
bun test
bun run test:windows
bun test design/test
bun test ios-qa/daemon/test
host UI/process launch for six installer-verified hosts
passing live installed-host adversarial gate (live v3 one-shot is retained failed evidence)
runtime-absent judgment through an actual host invocation (Codex passed; broader host UI coverage remains)
macOS + Linux + native Windows + Dev Container
local browser live journey + cancellation/leak cleanup (passed)
PDF strict + diagram suites
Context.dev verified-key public-page smoke
physical signed-iPhone five-check loop + Release symbol check (passed on 2026-07-20; retain earlier failures separately)
upgrade/fail/recover/rollback/uninstall end-to-end

The focused results above should remain in the ledger even after a later pass; record the newer SHA/time beside the newer result rather than overwriting the history.

Baseline evidence — 2026-07-16

Scope: detached base worktree /tmp/gstack2-baseline.e2qk7F, commit bb57306d98c97011b0919c6132705a15b1579781, captured 2026-07-16. These are baseline observations only; no result below was produced by the GStack 2.0 implementation.

Command ledger

Command	Evidence	Observed result
bun install	/tmp/gstack2-baseline-logs/bun-install.log	Completed; Bun 1.3.14 reported 297 packages installed.
bun run gen:skill-docs --host all	/tmp/gstack2-baseline-logs/gen-all.log	Reached the final GBrain token table and llms.txt summary (55 skills, 76 browse commands); no error marker is present.
bun run skill:check	/tmp/gstack2-baseline-logs/skill-check.log	Exit 1: claude/SKILL.md — generated file missing. The same run reported 54/54 outputs for each of the 10 hosts and fresh generated files, but the missing primary output keeps the command red.
bun run build	/tmp/gstack2-baseline-logs/build.log	Reached the final Node-compatible server bundle (server-node.mjs, 0.83 MB); no error marker is present.
bun test	/tmp/gstack2-baseline-logs/bun-test.log	Not green and incompletely terminated: 4,292 (pass) markers, 488 (skip) markers, seven explicit (fail) markers, and one unhandled between-tests error. The capture ends after browse/test/findport.test.ts without Bun's aggregate summary or exit record.
bun run test:windows	/tmp/gstack2-baseline-logs/test-windows.log	Exit 1 on shard 5/20. The runner selected 189 Windows-safe files and excluded 155. Completed shard summaries contain 593 passes, 14 skips, and one failed test before the run stopped.
bun test design/test	/tmp/gstack2-baseline-logs/design-test.log	Not green: 64 pass markers and one explicit failure. The log ends on daemon shutdown without an aggregate summary or exit record.
bun test ios-qa/daemon/test	/tmp/gstack2-baseline-logs/ios-daemon-test.log	Pass: 91 pass, 0 fail, 217 assertions across 10 files.

/tmp/gstack2-baseline-logs/git-status-after.txt is empty, recording no status entries after the runner. The logs are external audit artifacts and are not vendored into the repository.

Pre-existing bun test failures

The broad-suite log records these seven explicit failures:

1. gstack-decision-search --recent / --scope / datamark > --scope filters by scope expected output containing branch-call and received an empty string.
2. gstack-gbrain-detect > emits valid JSON even when nothing is configured expected status 0 and received 127.
3. gstack-gbrain-detect > reports gstack_brain_git: true when GSTACK_HOME has a .git dir attempted to parse empty output and raised Unexpected EOF.
4. gstack-gbrain-detect > reports gbrain_config + engine when ~/.gbrain/config.json exists attempted to parse empty output and raised Unexpected EOF.
5. gstack-gbrain-detect > malformed config returns null engine, does not crash expected status 0 and received 127.
6. gstack-gbrain-detect > detects a mocked gbrain binary on PATH and reports its version expected status 0 and received 127.
7. resolve-user-slug fallback chain > persists resolution to user_slug_at_<hash> on first call expected a hashed key and observed user_slug_at_local: persisttest.

Separately, test/gbrain-refresh-install-render.test.ts raised an unhandled between-tests error: Could not locate gbrain-refresh ok) branch. Because the log has no terminal Bun summary, the marker counts are evidence of what ran, not a claim that all files were reached or that the command had only those failures.

Other pre-existing failures

• skill:check: the primary claude/SKILL.md generated file was absent; the script explicitly exited with code 1.
• Windows-safe shard 5: Source-level guard: terminal-agent > lazy spawn: claude PTY is spawned in message handler, not on upgrade still looked for spawnClaude(, while the source excerpt used maybeSpawnPty(...). The shard reported 82 pass / 1 skip / 1 fail and the remaining 15 shards did not run.
• Design target: generateVariant Retry-After handling > HTTP-date: honors a future date with no extra leading exponential expected a delay of at least 2,500 ms and observed 2,246 ms.

What this evidence does not establish

• It does not establish a green baseline.
• It does not establish that GStack 2.0 preserves behavior; replacement contract tests have not yet been implemented.
• It does not turn incomplete logs into passes.
• It does not attribute these failures to later changes; every failure above was captured from the stated base SHA before GStack 2.0 work.