nx_wpt_runner.nx

nx_wpt_runner.nx -- BR1 v3: THE SOVEREIGN WPT HARNESS with TIME-SPINE TELEMETRY (2026-08-19). v1 ran real web-platform-tests through the browser's own JS+DOM lane. v2 fixed this harness's own overclaim (assert_throws any-throw) and made async_test synchronous with arithmetic incompleteness. v3 integrates the August-2026 test-intelligence field, measured against it honestly: - CONTENT-ADDRESSED RUNS (the Bazel/TIA class, but EXACT): a run is keyed by (engine_sha = sha256 of this very binary, corpus_sha = sha256 over the SORTED name+bytes of every .html in the directory). Same key => the result CANNOT differ (the engine is bit-deterministic), so re-running is refused with the prior row as the answer (--force overrides). Develocity-class predictive selection is PROBABILISTIC ("99% of failures"); ours is an invariant, not a bet -- and it is only available because determinism was proven first (same bytes twice => bit-identical render). - THE TIME SPINE: every counted run appends ONE TSV row to the history file -- results documented against time, so two rows say "and it is improving/regressing" where one row can only say a level. Schema mapped onto the OpenTelemetry CI/CD semantic conventions (v1.27+ cicd.* / test.* attribute families) so an OTLP exporter is a future formatter, not a redesign: epoch_s(cicd.pipeline.run start) dir(test.suite name) engine_sha16(service.version) corpus_sha16(vcs.ref analog) counters(test.case results) dur_ms(pipeline duration) verdict. - REGRESSION DETECTION: against the newest prior row with the SAME corpus (engine free to differ -- that comparison is the point), pass-down or fail-up flags REGRESSION and exits 4 so a beat can alarm. Improvement is reported and exits 0. - FLAKINESS REFUTED, NOT MANAGED (--twice): the field marks a test FLAKY when it both fails and passes across runs and then works around it; our engine is deterministic, so the whole directory runs twice and ANY divergence in any per-file counter is a REAL BUG (exit 5 DETERMINISM-VIOLATION), never noise to be tolerated. - CONTROLS FIRST, ALWAYS (unchanged law): runtime-synthesized neg+pos control; a harness that cannot see a failure counts nothing. All buffers derived from inputs, no fixed caps. usage: nx_wpt_runner <dir-of-.html-tests> [historyfile] [--force] [--twice] ENTRY: wpt_run(dir, histpath, force, twice) IS the harness (the /compare/browser BR1 contract symbol); main only parses argv and delegates, so the contract names the code that does the work, never a stub wearing its name (re-keyed 2026-08-23: the board had watched `wpt_run` while the harness lived in main + wpt_run_page, so a LANDED rung ranked #1 OPEN for five days). exit: 0 counted (incl improvement) | 2 harness unproven | 3 usage | 4 REGRESSION | 5 NONDETERMINISM license_tier: ORIGINAL expect_exit: 0