fix(geometry): carry the no-void probe reading through the census rec… #1371
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| # This Source Code Form is subject to the terms of the Mozilla Public | |
| # License, v. 2.0. If a copy of the MPL was not distributed with this | |
| # file, You can obtain one at https://mozilla.org/MPL/2.0/. | |
| # Content-matching validation fixture (issue #1891). | |
| # | |
| # The matcher's unit tests are hand-built fingerprints: they check that the | |
| # tiers do what their author intended. This lane checks something else -- that | |
| # the shipped matcher, over the shipped fingerprints, recovers a KNOWN | |
| # correspondence on real models. It is the only automated place where the | |
| # geometry hash, the data fingerprint and the matching tiers are measured | |
| # together against an answer key produced by construction rather than by the | |
| # program itself. | |
| # | |
| # G1 the fixture's own guards re-GUID trap, key disjointness, population | |
| # G2 the scored run precision/recall by tier, by kind, by class | |
| # G3 the calibration stratum re-sampled curves must NOT match at tier 1 | |
| # G4 the negative controls deleted/inserted elements must not be paired | |
| # G5 the harness mutation check always-match, always-abstain AND over-eager | |
| # must all be rejected | |
| # | |
| # See scripts/xmatch/SPEC.md for what each stratum means and where the numbers | |
| # come from. Two numbers per stratum, and the difference matters when this lane | |
| # goes red: | |
| # | |
| # * the GATING FLOOR is a ratchet against REGRESSION - the measured baseline | |
| # minus a 0.02 margin. Red means something got WORSE than it was, which is | |
| # a real finding and is fixed by fixing the regression, never by lowering | |
| # the floor to meet it. | |
| # * the PRE-REGISTERED TARGET is what was written down before the first | |
| # measurement existed. Two strata sit below theirs today (issue: the | |
| # geometry-less residue recalls 0.468 against a 0.5 target), and the run | |
| # prints those as BELOW PRE-REGISTERED TARGET lines every time. A floor | |
| # being green is not a claim that the number is good. | |
| # | |
| # A floor is NOT set so far below its baseline that a regression still passes: | |
| # at these populations 0.02 is one to two elements, so "one element worse" is | |
| # red. That was the point of not simply making this lane non-blocking. | |
| # | |
| # COST. Deliberately cheap and non-blocking; it MUST NOT become a required | |
| # check: | |
| # - weekly `schedule` + on-demand `workflow_dispatch`, the same profile as | |
| # determinism.yml / wide-arithmetic.yml / moonshot.yml; | |
| # - `push`/`pull_request` ONLY on the paths that can invalidate the result: | |
| # the diff engine, the geometry kernel and the wasm bundle it is measured | |
| # through, the CLI adapter that builds the fingerprints, and the harness | |
| # itself. A docs-only or viewer-only PR never sees this job. | |
| # - free `ubuntu-latest` runners, never Depot. | |
| # The measured work is ~15 s for three models; the wall clock is the wasm build | |
| # (Swatinem cache) and ~9 MB of fixtures. | |
| name: Content-matching fixture | |
| on: | |
| schedule: | |
| # Weekly, Tuesday 05:23 UTC. A day and minute no other weekly lane uses | |
| # (determinism Mon 03:17, wide-arithmetic Wed 04:29, moonshot Sun 06:41). | |
| - cron: '23 5 * * 2' | |
| workflow_dispatch: | |
| push: | |
| branches: [main] | |
| paths: | |
| # The matcher itself. | |
| - 'packages/diff/**' | |
| # The adapter that turns a store into the fingerprints it consumes. | |
| - 'packages/cli/src/commands/diff-engine.ts' | |
| - 'packages/cli/src/commands/diff-scope.ts' | |
| # The parse the whole measurement rests on: `IfcParser` builds the store, | |
| # `getInheritanceChainAcrossSchemas` decides which entities are keyed at | |
| # all, and the extraction APIs feed every data hash. A change here can | |
| # move the measured POPULATION without touching the matcher, which is the | |
| # same class of blind spot this lane exists to close. | |
| - 'packages/parser/**' | |
| # `RelationshipType` and the store types the adapter reads through. | |
| - 'packages/data/**' | |
| # `generateIfcGuid` / `isValidIfcGuid`: the answer key's re-GUID and the | |
| # guard that checks it. A change here changes the head revision itself. | |
| - 'packages/encoding/**' | |
| # The world geometry hash and AABB, and the bundle they cross into JS | |
| # through. The first run of this fixture found a real ordering | |
| # sensitivity here, so a kernel change is exactly what must re-run it. | |
| - 'rust/geometry/**' | |
| - 'rust/processing/**' | |
| - 'packages/wasm/**' | |
| # The harness, the answer key generator, and the pre-registered floors. | |
| - 'scripts/xmatch/**' | |
| - '.github/workflows/xmatch-fixture.yml' | |
| # | |
| # DELIBERATELY ABSENT, having walked the import closure rather than | |
| # guessed at it: `packages/ifcx` and `packages/pointcloud`. The parser | |
| # imports ifcx (and ifcx imports pointcloud), so their code is LOADED, | |
| # but the corpus is STEP-only and neither is exercised by a single | |
| # measurement here. Listing them would fire this lane on PRs that cannot | |
| # move a number in it. | |
| pull_request: | |
| branches: [main] | |
| # Same list as `push`. Path-filtered, non-blocking, and NOT to be added to | |
| # branch protection. | |
| paths: | |
| - 'packages/diff/**' | |
| - 'packages/cli/src/commands/diff-engine.ts' | |
| - 'packages/cli/src/commands/diff-scope.ts' | |
| - 'packages/parser/**' | |
| - 'packages/data/**' | |
| - 'packages/encoding/**' | |
| - 'rust/geometry/**' | |
| - 'rust/processing/**' | |
| - 'packages/wasm/**' | |
| - 'scripts/xmatch/**' | |
| - '.github/workflows/xmatch-fixture.yml' | |
| concurrency: | |
| group: ${{ github.workflow }}-${{ github.ref }} | |
| cancel-in-progress: ${{ github.event_name == 'pull_request' }} | |
| permissions: | |
| contents: read | |
| env: | |
| FORCE_JAVASCRIPT_ACTIONS_TO_NODE24: true | |
| jobs: | |
| content-matching-fixture: | |
| name: Content matching vs a known correspondence | |
| runs-on: ubuntu-latest | |
| timeout-minutes: 30 | |
| steps: | |
| - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1 | |
| with: | |
| lfs: false | |
| persist-credentials: false | |
| - name: Setup pnpm | |
| uses: pnpm/action-setup@0977fd99725f1db4007ccb2928dbb4e90d06cc86 # v6.0.10 | |
| - name: Setup Node.js | |
| uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0 | |
| with: | |
| node-version: 22.14.0 | |
| cache: pnpm | |
| - name: Setup WASM build toolchain | |
| uses: ./.github/actions/setup-wasm-build | |
| with: | |
| cache-prefix: ci-xmatch | |
| - name: Install dependencies | |
| run: pnpm install --frozen-lockfile | |
| # From source, never a published bundle: the geometry hash this fixture | |
| # measures lives in the Rust kernel, and a prebuilt bundle would hide | |
| # exactly the change the lane exists to catch. | |
| - name: Build WASM bindings from source | |
| run: bash scripts/build-wasm.sh | |
| # The harness imports built `dist/` by relative path: the diff engine and | |
| # the CLI's fingerprint adapter (which pulls parser/data transitively). | |
| - name: Build the packages the harness measures | |
| run: pnpm turbo build --filter=@ifc-lite/diff --filter=@ifc-lite/cli | |
| - name: Cache fixtures | |
| uses: actions/cache@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0 | |
| with: | |
| path: tests/models | |
| key: ci-xmatch-fixtures-${{ runner.os }}-${{ hashFiles('tests/models/manifest.json') }} | |
| # Unconditional, cache hit or not: the fetcher hashes what is on disk and | |
| # re-downloads only what is missing or wrong, so it is the fetch and the | |
| # integrity check at once. | |
| - name: Fetch the corpus | |
| run: | | |
| node scripts/fixtures/fetch-fixtures.mjs \ | |
| ara3d/duplex.ifc \ | |
| ara3d/AC20-FZK-Haus.ifc \ | |
| various/rvt01.ifc | |
| # G5 first, and in its own step: if the harness cannot fail, nothing the | |
| # next step reports means anything. `--self-test` asserts that BOTH an | |
| # always-match and an always-abstain matcher are rejected. | |
| # Three mutants, and the generator's own refusals. `--self-test` asserts | |
| # that always-match, always-abstain AND over-eager are each rejected on | |
| # per-pair clauses, and that the three places which can detect a corrupt | |
| # answer key still throw. Its exit code deliberately does NOT depend on | |
| # whether the real matcher clears its floors: "can this harness reject a | |
| # wrong matcher" has to be answerable when the scored run is red, which | |
| # is when it matters most. | |
| - name: Mutation-check the harness (always-match, always-abstain, over-eager) | |
| run: node scripts/xmatch/run.mjs --self-test | |
| - name: Score the matcher against the answer key | |
| run: node scripts/xmatch/run.mjs | |
| # The committed scorecard is the standing artifact, and the run is | |
| # deterministic (seeded mutations, wall clock excluded from the file), so | |
| # a drift is a real change in the measurements. | |
| # | |
| # `continue-on-error` because the committed copy was produced on macOS | |
| # and this runs on Linux: the geometry is wasm (bit-deterministic by | |
| # spec) and the hashing is integer arithmetic, so the two SHOULD agree, | |
| # but until one scheduled run has demonstrated that, a platform | |
| # difference here would be a measurement artifact rather than a | |
| # regression, and it must not be the thing that reddens the lane. The | |
| # scored step above is the verdict. | |
| - name: Committed scorecard is current (informational) | |
| if: always() | |
| continue-on-error: true | |
| run: | | |
| node scripts/xmatch/run.mjs --write || true | |
| git --no-pager diff --exit-code -- scripts/xmatch/scorecard.json |