1.0.0 is published on npm under latest. Release checklist.
Verified registry bytes and clean installations; beta remains 0.1.0-beta.2.
Source d43e235 passed 10 × 188 = 1,880 sequential checks and 188 matching
four-worker checks, with zero retries/skips and the unchanged 100 ms bound.
All eleven raw reports were independently validated: 1,771 unique journey IDs,
maximum passing observation gap 47.7 ms.
Raw-report hashes and validation.
Historical 0.1.0/alpha/beta results below retain their original identities.
Beta.2 qualification. Source fd8f58a completed ten rounds of 180 tests and the separate 180-test four-worker comparison. All eleven raw reports matched, with 1,771 distinct journey IDs and zero retries. Beta.2 release notes · Independent validation.
Frozen candidate (31 October 2026): 1187fde.
The candidate/0.1.0-alpha.6 tag
remains at this commit. Candidate CI
passed three separate macOS 15 ARM allocations and Linux: 172 tests per job,
one worker, zero retries/skips, and both real package consumers in every job.
ARM maximum passing-window gaps were 87.7 / 65.6 / 50.7 ms; Linux was 23.9 ms.
The 100 ms limit is unchanged. Verified CI record.
The benchmark was measured at f9c492b: 108/108, with
input hashes matching this candidate. Earlier failures and measurements keep their
source identities. Short checks do not guarantee shared-runner latency.
Final source-bound reliability completed on 31 October 2026 UTC (1 November in Kyiv):
100 × 172 sequential executions + 172 four-worker checks, zero retries/skips.
Independent validation.
The release archive gate, owner sign-off and publication authorization remain separate.
Final reliability verified: The frozen 1187fde candidate completed 100 × 172 = 17,200 sequential test executions and 172 matching four-worker checks, with a 100 ms limit and zero retries/skips. All 101 raw reports were independently validated. Development still defaults to four rounds. See the protocol and retained results. Ordinary CI checks the real package and clean consumers without claiming final acceptance.
The historical alpha source candidate was 0.1.0-alpha.6, ruleset 0.4.0, schema 1.
Measurements use synthetic fixtures, not customer data. A passing regression test
may deliberately observe a failed or blocked journey; it does not mean that its
component is accessible.
| Check | Recorded result |
|---|---|
| TypeScript | Strict library, integration and benchmark checks passed |
| Unit tests | 39 passed, including verification-mode and resume-boundary checks |
| Browser suite | 1187fde: three ARM allocations 172/172 each; Linux 172/172. No retries/skips. Two added cases cover async context and budget expiry at dispatch; the earlier 130.3 ms failure is retained |
Worker isolation (1187fde) |
172/172 on four local workers; semantic signature matches all 100 sequential rounds |
| Canary isolation | Six cases: successful and exceptional paths for each scenario; five distinct markers |
| Observation lifetime | Coherent focus, detached replacement, changed locator bindings, absent targets, late sibling cleanup and final map disposal |
| Release evidence validator | 26 script tests passed, including four regressions for incomplete or conflicting guard hashes; no browser rounds run by these tests |
| CI guard | 52 actual subprocess exit-code cases; current guard record |
| Performance | Recorded f9c492b: 108/108 tests; median paired overhead: dialog -9.2%, manual tabs -6.8%, forms (insertText) -12.6%; samples and validation |
Final repeatability (1187fde) |
100 × 172 = 17,200 sequential executions; 16,261 distinct journey IDs across all 101 reports; zero retries/skips/failures. Proof and independent validation |
| Package consumers | 1187fde: both clean offline consumers passed in all four jobs. ESM/CJS declarations; vanilla: 6 browser tests + 3 guarded journeys; React: 3 browser tests + 1 guarded journey + expected exit-1 control. Archives and results; final acceptance remains false |
| Hosted platforms | Three distinct macOS 15 ARM runner IDs and non-overlapping job times verified at 1187fde; Linux passed too. Combined record. This is a bounded development check, not a guaranteed latency bound |
The interactive candidate review checked the frozen examples and existing HTML/JSON/JUnit reports. That interactive review did not run a full suite; owner sign-off remains separate from the completed final reliability series. The archive handoff distinguishes the preserved candidate from any later package rebuilt with updated documentation.
Thirty alternating measured pairs per workload, after five warmups per arm. The local runner was an Apple M3 (8 logical CPUs), macOS arm64/Darwin 27.2.0, Node 24.22.1, Playwright 1.64.1, Chromium 156.0.8078.4, on 31 October 2026. This benchmark ran independently of final reliability acceptance; it did not start a reliability series. The later final series is recorded separately above. The earlier October 31 result and October 29 result are preserved separately. Every pair also passed semantic equivalence checks; three known defects agreed on the first mismatch and dependent checks. Manual tabs include required observation windows, which dominate duration. These figures do not establish a universal speed advantage or replace checks in a consumer application.
Starting with beta.2, final repeatability requires ten complete sequential rounds for one frozen candidate and a matching four-worker comparison. The earlier 100-round reports below retain their original scope.
The full-suite protocol retains every raw run, rejects retries,
checks attachment identity, and prevents combining different source/environment
snapshots. The earlier alpha.6 local series at 0300552 completed all 100 rounds and its four-worker
comparison on 29 October 2026 at 22:31 UTC. Its identity is
191d2d89449e23a319e221949729f6d069d4c57acbb6a432056867bc159b0fed.
Twelve completed earlier local rounds and the hosted attempts are retained separately
and are not added to the new total.
The historical alpha.5 record accepted 45 rounds and excluded its interrupted 46th
round; it remains history rather than evidence for the changed runtime.
The historical proof and historical policy agree
on the source hashes at 0300552, 100 completed rounds and the final isolation signature.
The 101 raw reports contain 16,463 test executions, including 303 engine-only
tests without a journey attachment. All 16,160 journey IDs are distinct.
The canary tests inspect only KeyTrail's own JSON, messages and stdout/stderr. External Playwright artifacts are outside that guarantee. Budget, focus-loss and no-action-after-return regressions execute in each full round.
The source now resolves locators into a private browser map on every observation, reads focus and containment together, and clears stored nodes before returning. It waits for sibling reads even on failure and disposes the map at journey end. No target binding is cached; mechanism and regressions. The atomic-registration four-round proof validates its recorded runtime, before the consecutive-read change. Repeatability records remain tied to their own source hashes.
pnpm build
pnpm test:types
pnpm test:unit
pnpm test:guard
pnpm benchmark
pnpm verify:reliabilityOn macOS, prefix the long browser commands with caffeinate -is to prevent host
sleep. Run performance measurements without concurrent verification jobs. The default pnpm test:package packs the archive and installs it into two clean offline consumers; it
checks ESM/CJS imports and declarations, and runs the vanilla and React examples.
Only node scripts/check-package.mjs --release also requires fresh final evidence
and matching source hashes. Use the read-only readiness command in the
release checklist to inspect those gates without running them.
React lockfile preparation fetches registry metadata before the separate frozen offline install.
Install the separate React lockfile first as described in the release checklist.
Archive size/hash are written outside the package to avoid a
self-referential archive checksum.
DEV08 history retains the earlier focused-repeat evidence,
including the diagnosed host-sleep timeouts and hosted outer-timeout correction.
DEV10 adds a React portal in development StrictMode, closes/remounts the dialog,
and verifies wrong initial and return focus. Its separate negative process does
not catch the assertion: exit 1, HTML, JUnit and the original incomplete attachment
are checked. Playwright classifies JourneyAssertionError as a JUnit error, so
both error and failure counts are considered; neither is treated as success.
The report guide explains the saved artifacts. Earlier transport revisions have their own four-round development evidence; these
do not certify the current consecutive-read scheduler. Consumer checks remain additional
evidence, not extra sequential rounds.
The beta.2 results are recorded in the release notes.