feat(remote-react-components): cross-version backwards-compatibility smoke tests (#2600)#2661
Merged
Conversation
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…2600) Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Adds crossVersionEnvironmentsPlugin.ts, which writes a generated testEnvironments module (one entry per cross-version manifest target, pairing each OLD RemoteRoot/components with the CURRENT host renderer) to a gitignored file and aliases @/tests/lib/environments to it, so ordinary subset test files pick up the injected envs unmodified. JSX lives in a real renderRemoteVersion.tsx helper (virtual modules aren't picked up by @vitejs/plugin-react's transform). A generated real file is used instead of a Vite virtual module for easier debugging, though a cold-cache dependency-scanner limitation (documented in the plugin's top comment) means CI wiring must budget for a throwaway warm-up run.
…fest A manifest that parses but resolved zero targets would silently emit `export const testEnvironments = []`, making Task 6's `test.each(testEnvironments)` run zero assertions and let CI go green while validating nothing. Guard writeCrossVersionEnvironmentsModule to throw a clear error instead.
Wire the cross-version harness to a curated 4-file subset (BigNumber, CopyButton, LabeledValue, Tooltip) instead of the __spike__ glob, and delete the spike test now that it's superseded. The subset is far short of the ~10 files hoped for. Empirical curation across ~30 candidates (all "stable primitive" components, plus every visual test avoiding a top-level Flex wrapper) found a systemic pattern: any prop other than `children` (color/size/boolean variants, Flex direction/gap, or a data-testid used by a later .click()/.hover()) fails to reach the current host renderer when paired with an OLD remote-react-components version, identically across all 3 resolved versions. Only children composition survives; the 4 kept files pass because none of their pixels depend on a prop. See subset.ts's top comment and the Task 6 report for the full evidence trail and leading hypothesis (vitest.config.ts's sharedRemoteElementsAlias forcing old callers onto the current remote-elements package). Also attempts optimizeDeps.entries for the documented cold-cache flakiness (Task 5): it reduces the required cold-cache warm-up from 3 runs to 2 but does not eliminate it, so kept alongside the existing warm-up note. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…rness Superseded by the e2e iframe harness. Keeps the resolver, prepare script, and exclude list; removes the same-realm vitest config, environments generator, render helper, and subset that could not test prop-level backwards compatibility. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…-version tests Versions alpha.889-895 shipped async serializers (#2596) while @quilted/threads was still externalized (not bundled until #2620/alpha.896), so they DataCloneError on connect without a pnpm-patched threads. They are genuinely broken published releases and must never be tested against. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…#2600) Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…hots later) (#2600) Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Adds a dedicated remote test server (e2e/cross-version/) that builds its remote document against an installed OLD version of @mittwald/flow-remote-react-components, selected via FLOW_CROSS_VERSION and resolved from cross-version.manifest.json (defaults to the first target). The server aliases the bare @mittwald/flow-remote-react-components (+ /RemoteRoot) specifiers to the old install's dist, plus react/react-dom to the old install's copies so the remote document uses a single React instance. It runs in its own document/realm loaded into an iframe by RemoteRenderer, so no @mittwald/flow-remote-elements or @quilted/threads special-casing is needed. Port lives in a single shared constant (crossVersionServerPort) reused by the server and the future host test. A temporary Smoke remote entry proves R1; R2 replaces it with the real screenshot entries. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…on + references Builds on the R1 cross-version server. For a set of representative remote entries (Button states, Checkbox, Switch, Badge, CounterBadge, Text, Heading, Link, TextField), the host renders each through the current RemoteRenderer over the real iframe connection, normalizes the resulting HTML, and compares it to a single version-independent reference per entry. - normalizeHtml: pure, unit-tested transform that strips volatile bits (react-aria/React generated ids + id-referencing attrs, data-flr-version / data-flr-initialized, the hidden connection iframe, insignificant whitespace) while keeping semantic attributes and text. - entries.browser.test.remote.tsx: deterministic render entries imported from the package specifier (version alias applies). - entries.browser.test.tsx: host suite; toMatchFileSnapshot against __html__/<entry>.html (shared by all versions + the current reference). - FLOW_CROSS_VERSION=current aliases the package specifier to the workspace src to generate the references; installed old versions are compared against them. All 12 entries match across current, 0.2.0-alpha.883 and 0.2.0-alpha.686 — the old versions serialize identical props/structure over the real connection. No entries dropped. normalizeHtml unit test runs in the unit-dev project; __html__ references are prettier-ignored (compared byte-for-byte). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
… instead of deleting The cross-version HTML normalizer previously DELETED id/for/aria-* attributes whose value was a generated id, making "attribute present, pointing at a generated id" indistinguishable from "attribute absent" — masking real label/control association regressions (e.g. the textField for/id wiring was erased entirely). - Rewrite generated id VALUES to stable, position-based placeholders (GENERATED_ID_1, _2, … by first appearance) and rewrite every reference to the same id to the same token, preserving attribute presence + cross-refs. - Scope hidden-iframe removal to RemoteRenderer's connection iframe only (matched by its `visibility: hidden` inline style), so a real iframe component is not silently erased. - Anchor the generated-id match to the full attribute-value token, so an author-set id merely containing such a substring is not dropped. - Note the block-level assumption of the whitespace-collapse step. Regenerated the __html__ references from current; all 12 entries still match across current, 0.2.0-alpha.883 and 0.2.0-alpha.686 with the stricter normalizer (the label/input associations are identical across versions), so no entries were dropped. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
dev/cross-version/run.ts reads cross-version.manifest.json and runs the HTML-output comparison suite once per installed old version (FLOW_CROSS_VERSION set per target), aggregating a PASS/FAIL summary and exiting non-zero if any version fails. Looping in node (not a shell loop) keeps it cross-platform; the vitest CLI is spawned with the current node binary to avoid nested pnpm/corepack spawns. Scripts: - test:cross-version — loop installed old versions vs. __html__ references - test:cross-version:update — regenerate references from current (--update) - test:cross-version:dev — vitest dev against the suite config Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add nx targetDefaults for the cross-version targets: - test:cross-version:prepare — dependsOn ^build, cache false (network installs) - test:cross-version — dependsOn ^build + test:cross-version:prepare, cache true, with inputs covering e2e/cross-version sources, dev/cross-version sources, the manifest, and the exclude list - test:cross-version:update / :dev — dependsOn ^build + prepare, cache false Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add a scheduled-only `cross-version` job to test-visual-scheduled.yml (not part of affected:test:browser on PRs). It installs the playwright shell, installs the target old published versions (test:cross-version:prepare, a separate step so a publish-gap install failure is visibly its own red step), then runs test:cross-version through nx (^build served from cache). Reuses the visual job's 3-attempt retry loop and Slack-alert-on-failure pattern. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add a "Cross-version smoke tests" section to CONTRIBUTE.md: the honest scope (full old remote stack — old flow-remote-react-components + flow-remote-core — in a real iframe realm, connected to the current host over the versioned @quilted/threads connection, HTML/DOM-output comparison, not screenshots), how to run and update references locally and when, the version-selection resolver (semver categories + alpha-offset fallback), the broken-window exclude list (alpha.889–895, #2596→#2620) and how to exclude a version, and the publish-gap drop behavior. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…ic specifier Rewrites the e2e remote test entries and the test server to import from @mittwald/flow-remote-react-components (public API) instead of workspace src paths, with the dev server aliasing the specifier back to src. This lets the same entries be reused by the cross-version harness, where the specifier is aliased to an old installed version instead. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…ml comparison Replaces the dedicated cross-version entries + committed HTML fixtures with: - reuse of the existing e2e remote entries as the corpus (enumerated from their source, so no per-entry maintenance) - an ephemeral reference rendered from the CURRENT version at run time (the existing e2e server) compared against the OLD version (cross-version server) — no committed baselines - graceful skip when an old version lacks a component (predates it) - a documented per-entry exclude list for components whose DOM legitimately evolved across versions (Modal#inList / List empty-view) - run.ts fails on a zero-target manifest Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…exclusions - excludedEntries carries a maxAffectedVersion; entries are skipped only for old versions at/below it, so newer in-range versions keep real coverage (Modal#inList now runs + passes on alpha.883, skipped on <=alpha.791) - expose the tested version to the host bundle via define - drop the obsolete test:cross-version:update (no committed references) and simplify run.ts; mark test:cross-version non-cacheable - update CONTRIBUTE.md for the reuse + ephemeral-reference model Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Invoke the package test script directly instead of the nx target, so the retry loop doesn't re-trigger the network prepare dependency; prepare runs once as its own step. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…as minVersion Replace the 'excluded up to version Y' model with a positive 'comparable from version X onward' (minVersion) rule plus skipVersions for one-off exceptions — easier to maintain (state when a test starts working). Behaviour unchanged: Modal#inList runs on alpha.883, skipped on <=alpha.791. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…rness Add a second, in-process cross-version harness alongside the high-fidelity iframe one (hybrid). It reuses the existing src/tests/visual/*.browser.test.tsx corpus UNMODIFIED by injecting a CrossVersion environment (old RemoteRoot + current RemoteRenderer via a shared RemoteReceiver; first-wins customElements patch so the old flr-* registrations win). Comparison is STRUCTURE-ONLY: structuralHtml strips all attributes, so only the element tree (tags + nesting + text) is compared. This is the right guarantee for a one-realm harness — attribute-level details (prop serialization form, class modifiers, inline styles, icon geometry) drift across versions without being a backwards-compat break, and asserting them in-process produces false positives (verified against the real-connection iframe harness). Structure-only gives broad, low-noise coverage of the one thing this harness asserts reliably: does the current host still build the same DOM shape from an old remote version's output? All six curated tests (Checkbox, Switch, Separator, Badge, Text, Heading) pass across the target versions with no per-diff handling. The two harnesses cover deliberately different things — iframe = attribute fidelity over the real connection for curated scenarios; in-process = DOM shape across the broad reused corpus — documented in CONTRIBUTE.md. - structuralHtml + unit tests - run-inprocess.ts orchestration (current reference pass, per-version compare, GitHub ::error::/step-summary/output reporting), mirroring run.ts - testVersionSupport for scoping genuine structural divergences - screenshotFailures disabled (HTML comparison; avoids polluting the reused tests' src/**/__screenshots__ baselines) - scripts, nx target, .gitignore, and CONTRIBUTE.md docs (explicit covers/does-not-cover for both harnesses) Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…ss-version job Fix the cross-version job resolving @mittwald/flow-react-components/internal and @mittwald/flow-remote-react-renderer: the runs invoke the package scripts directly (not the nx target), so nx's ^build never fired and the host packages were never built from source in CI. Add an explicit `pnpm nx run remote-react-components:build` step (its ^build closure covers components + remote-react-renderer) before both cross-version runs. Also wire the in-process harness into the same job (shares the one prepare step) and extend the Slack alert to report failures from either harness. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…rom Button.scenarios Restore Button.browser.test.tsx to its standalone, inline form instead of importing and looping over Button.scenarios. The iframe cross-version harness keeps Button.scenarios.tsx as its own corpus, so the Button JSX is now duplicated across the two — a deliberate trade: a bit of duplication over coupling a normal visual test to the cross-version scenario format. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…-process harness Broaden the in-process cross-version harness from 6 curated tests to the whole src/tests/visual/*.browser.test.tsx suite (136 tests green across the target versions), keeping the structure-only comparison. - Wire dev/vitest/setupBrowser.ts into the harness setupFiles (after the first-wins customElements patch) so the reused tests get their normal helpers (page.getByLocator, all.css). - Restructure reusedVisualTests.ts around subtraction: REUSED_VISUAL_TESTS is the whole-suite glob; EXCLUDED_VISUAL_TESTS drops whole files that can't RUN (interaction tests whose overlay never opens in the one-realm setup — a split react-aria overlay context; and files that render-error on a component undefined in an old bundle); VERSION_SCOPED_TESTS lists individual tests whose element tree genuinely changed, with the fromVersion they became comparable from (Checkbox edge cases pinned to alpha.884 by bisect). - testVersionSupport now gates the comparison from VERSION_SCOPED_TESTS. - testScreenshot polls the structure until it settles (readStableStructure) before comparing, so async rendering can't cause a spurious diff. - CONTRIBUTE.md documents the two subtraction lists and the overlay limitation. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…lusion reason
The excluded interaction files were annotated as "overlay never opens /
react-aria context split" — that's wrong. Overlays do render (ComboBox
interaction passes; manual clicks open the Select). The real cause: a
click-to-open control's open doesn't fire under a PROGRAMMATIC click in the
one-realm setup (the press→open round-trip to the remote side doesn't trigger;
keyboard-driven overlays and manual clicks work), so the popover never opens and
getByTestId("option") times out. Also confirmed data-testid forwards fine and
the HTML normalization doesn't touch the live-DOM selectors.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…ia test.skipIf
Replace the in-process harness's two subtraction lists (file-level
EXCLUDED_VISUAL_TESTS + comparison-level VERSION_SCOPED_TESTS, plus the
render error boundary and render-gating) with a single per-test skip
predicate: test.skipIf(crossVersion({ below, exclude })). A version-bound
reused test now gates itself and is skipped whole — render, interaction and
comparison — on versions where its component didn't exist yet or its element
tree genuinely differs. In the normal visual suite the predicate is a no-op,
so every test still runs there.
Also:
- strip the hidden remote-DOM (<flr-*>) subtree in structuralHtml so the
comparison is purely the host output the renderer builds (the flr-* mirror
is a remote-side implementation detail that drifts freely between versions);
- split the ContextMenu visual test into a cross-version-safe basic case and
a full (sections/Kbd) case scoped from alpha.791;
- update CONTRIBUTE.md to describe the single skip mechanism and flr-stripping;
- delete testVersionSupport.ts (no longer needed).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
mfal
added
the
update-screenshots
github-actions
Bot
removed
the
update-screenshots
Co-authored-by: mfal <4696750+mfal@users.noreply.github.com>
…for split The ContextMenu visual test was split into a basic and a full case; regenerate the darwin baselines (updated "ContextMenu - opened", new "ContextMenu sections and shortcuts"). Linux baselines are handled by the update-screenshots CI job. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Behavior-preserving readability pass on the in-process cross-version harness: rename the structuralHtml accumulator, clarify helper names in crossVersionEnvironment, drop a redundant type annotation, and fix a long-standing TestEnvironment typo. No functional change. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…es" cross-version
The "Checkbox edge cases" test name exists in both Checkbox.browser.test.tsx and
CheckboxButton.browser.test.tsx. The former was already scoped to alpha.884; the
latter was missed, so it compared on alpha.883/791/686 and failed (CheckboxButton
dropped its wrapper <div>s in alpha.884). Apply the same
test.skipIf(crossVersion({ below: "0.2.0-alpha.884" })) gate.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
mfal
marked this pull request as ready for review
July 17, 2026 19:14
The skip predicate already carries the version; the comment only needs the human-readable "comparable/available from alpha.X" note. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
# Conflicts: # packages/remote-react-components/src/tests/lib/environments.tsx
# Conflicts: # packages/remote-react-components/src/tests/visual/__screenshots__/ContextMenu.browser.test.tsx/ContextMenu-opened-firefox-darwin.png # packages/remote-react-components/src/tests/visual/__screenshots__/ContextMenu.browser.test.tsx/ContextMenu-opened-firefox-linux.png # packages/remote-react-components/src/tests/visual/__screenshots__/ContextMenu.browser.test.tsx/ContextMenu-opened-webkit-darwin.png # packages/remote-react-components/src/tests/visual/__screenshots__/ContextMenu.browser.test.tsx/ContextMenu-opened-webkit-linux.png
mfal
added
the
update-screenshots
github-actions
Bot
removed
the
update-screenshots
Co-authored-by: mfal <4696750+mfal@users.noreply.github.com>
…pare step The standalone "Build workspace packages" step only existed because the cross-version steps ran the package scripts directly, bypassing nx's `^build`. That direct-script choice was originally to stop a retry loop from re-running the network prepare — but the retry loop was later removed, so the workaround was vestigial. `test:cross-version:prepare` already declares `^build` as an nx dependency, so running prepare via `pnpm nx run` builds the host packages (nx-cached) and installs the old versions in one step; the standalone build step is dropped. The two suites stay direct-script steps so each keeps its own `failures` output for the Slack alert and the in-process suite still runs if the iframe one fails. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…ary to alpha.926 CheckboxButton's edge-case wrapper structure was flattened again in the alpha.926 release (nothing was published between alpha.921 and alpha.926), so the reused "Checkbox edge cases" test diverged from current on every target >= the old alpha.884 boundary. Raise the skipIf boundary to alpha.926 — the first published version with the current structure. Older targets skip until a newer version ships to compare against. Only the raw CheckboxButton test is affected; the <Checkbox> edge case stays comparable from alpha.884. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
… npm not pnpm Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The selection tests only spanned two minor lines within one major, so the cross-major behaviour of `latestOfPreviousLine` was never asserted. Add a case with two majors where the current version is the first minor line of the new major, verifying latestOfPreviousLine resolves to the latest published version of the previous major. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…connection flakes The real-iframe suite intermittently fails with "Could not establish remote connection: Timeout reached" — a transient @quilted/threads handshake flake under CI, not an HTML diff (~2 of 3 recent scheduled runs hit it on varying versions). The prior "no retry" comment assumed this suite doesn't flake, which is wrong. Wrap it in a 3x retry loop (same pattern as the visual job's Firefox hang retry): a real divergence fails deterministically on every attempt, while a connection flake clears on a fresh run. Kept as a direct package-script call so a retry doesn't re-trigger the non-cached network prepare. The in-process suite keeps its per-test vitest retry (no real connection to time out). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
run.ts (iframe) and run-inprocess.ts (in-process) were ~90% identical: manifest reading, failed-test parsing, GitHub error/summary/output reporting, vitest-bin resolution, and the per-version loop. Extract all of it into crossVersionRunner.ts; the two entry scripts now just pass what differs — vitest config, temp prefix, report labels, and (in-process only) the optional --update reference pass. No behavioural change; the vitest invocations are identical. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…own file Move the ScenarioVersionRule interface and the scenarioVersionRules data array to scenarioVersionRules.ts; scenarioVersionSupport.ts keeps only the matching logic (isScenarioComparable) and imports the data. Separates the version-scoping data from the logic that consumes it. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…against truncation transient readStableStructure returned on the first single pair of matching samples, which could capture a ResizeObserver-driven text-truncation transient (a measuring DOM copy briefly duplicated alongside the displayed one) before it settled — making the long-label "edge cases" tests flake intermittently in the structure-only comparison. Wait for a sustained quiet window instead (5 consecutive identical samples, up to a ~4s cap). This cleanly separated the two edge-case failures: <Checkbox> now compares reliably (it was pure capture flake), while <CheckboxButton> is confirmed a real deterministic divergence at <= alpha.921 vs current — so its alpha.926 boundary stays; comment clarified. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Lisa18289
approved these changes
Jul 21, 2026
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes #2600.
What
Adds a cross-version backwards-compatibility smoke-test harness for the remote side of Flow. It renders old published versions of
@mittwald/flow-remote-react-componentsthrough the current host renderer and asserts the host-rendered output still matches — guarding the contract that mStudio (running the current version) keeps rendering extensions built against older published versions.Why
Extensions ship against a published remote version but render inside a host on the current version. Nothing tested that the versioned remote↔host contract stays backwards-compatible. This suite catches prop/serialization/structure drift before it breaks real extensions.
How it works
@quilted/threadsconnection — exactly how mStudio loads extensions. (An earlier in-process attempt was abandoned: two versions can't share one custom-element registry in a single realm, which silently dropped all props. The iframe gives each side its own realm.)normalizeHtmlstrips volatile bits (generated ids → stable placeholders, the hidden connection iframe,data-flr-*, spinner animation offsets) while keeping semantic attributes, text, and label↔control wiring.(components) => JSXinsrc/tests/visual/*.scenarios.tsxand drive both the existing in-processLocal/Remotescreenshot tests and the cross-version HTML comparison. Discovery is generic (glob) — no per-test bridges.selectTargetVersionspicks old versions by semver category with an alpha-offset fallback; a checked-incross-version.exclude.jsonskips known-broken releases (alpha.889–895, which shipped async serializers before threads was bundled).prepareinstalls the targets in isolation and writes a git-ignored manifest; un-installable versions (npm publish gaps) are dropped with a warning.scenarioVersionSupport.ts(minVersion= "comparable from version X onward", plusskipVersions), so newer in-range versions still get real coverage.test-visual-scheduled.yml) — not on PRs. Dynamic OS-assigned server ports,server.warmupfor cold-cache safety, and failure reporting via a$GITHUB_STEP_SUMMARYtable +::error::annotations naming the version+scenario + an enriched Slack alert.Reviewer notes
Buttonis migrated to a*.scenarios.tsxso far. The generic discovery + dual-mode pattern are in place and proven; migrating the remainingsrc/tests/visual/*is a follow-up (mechanical, batchable) that grows the cross-version corpus toward the full suite.0.2.0-alpha.*prerelease line, so version selection uses the alpha-offset fallback (offset-10/100/200); the semver-category path is exercised by unit tests.Verification
test:e2eandtest:visualunaffected;selectTargetVersions+normalizeHtmlunit tests pass; tsc / eslint / prettier clean.