Count only rendered steps, and keep a data page's paginator in frame

[woksin]

Summary

Two layout defects that only appear once there is enough content to expose them.

Fixed

• A stepper counts only the steps it actually renders. A conditionally rendered panel still reaches it as a child, so a hidden step was counted: Submit never appeared on the real last step, a dead Next took its place, and clicking it landed on a step with nothing in it. Each hidden step cost another click, and the per-step validation gate read past the end of its own array and stopped blocking (StepperCommandDialog miscounts conditional steps #120)
• A step that disappears after the user has already passed it no longer strands the wizard. The index is kept within the steps that still exist, so Submit stays reachable, Previous moves rather than absorbing a click per vanished step, and hiding every step no longer leaves a dead Next behind (StepperCommandDialog miscounts conditional steps #120)
• A data page keeps its paginator inside the height it was given. The table grew to its content instead of filling the pane, so a full page of rows pushed the paging controls past the bottom of the frame and clipped them, leaving no way to reach the next page. A short page fit and looked correct (DataPage clips its paginator in a bounded-height container when actions are present #126)
• A data page with a details component keeps its paginator in frame too. The split view sizes its panes from a stylesheet the package never imported, so the pane grew to content and clipped the paginator exactly as before, and the two panes stacked instead of siding. The stylesheet now ships with the component, so there is nothing for a consumer to import (DataPage clips its paginator in a bounded-height container when actions are present #126)

Changed

• A data page without a details component no longer renders a split view, and a data page now requires an ancestor with a real height — documented, with a minimum height so a page that does not provide one stays usable rather than collapsing (DataPage clips its paginator in a bounded-height container when actions are present #126)

[woksin]

Reviewer context, kept out of the description because it is not release notes.

Verification. npx tsc -b no errors; yarn lint clean; yarn test 375 passed (baseline 333, so +42 across 6 new spec files); yarn build and yarn build-storybook both clean. Each item verified by mutation — reverting the fix must turn a named spec red — and restored by hand before the next step.

DataPage, DataTableForQuery and CommandStepper had no specs at all before this. That absence is why both defects shipped: every existing spec imports source and every DataPage story renders a single short page, so neither defect could appear.

Environment note that matters for anyone reproducing this. yarn.lock is gitignored and no lockfile is tracked, so installs are not reproducible and a branch switch silently carries dependency state across. This tree had drifted to PrimeReact 11 with jsdom missing, producing 125 phantom tsc errors that have nothing to do with either change. yarn install restored what main pins (PrimeReact 10.9.8, jsdom 30.0.1) and the baseline above is measured against that. Worth tracking separately — it is plausibly why #118's ESM defect reached publish.

One deliberate deviation. The single-pane branch keys on props.detailsComponent alone rather than detailsComponent && selectedItem. With the latter, the element flips between a plain div and the split view the moment a row is clicked, so React unmounts the subtree and the table loses its filter state and refetches on every select. Branching on the stable prop keeps the subtree mounted and still removes the split view from the single-pane case.

Deliberately out of scope.

• Fragment support in the stepper. A fragment still counts as one step and pools its panels' fields, exactly as before. Supporting it properly is a new input shape and a separate change.
• allotment/dist/style.css is still imported nowhere, so a data page with a details component stacks its panes instead of splitting. That is a second, larger defect than DataPage clips its paginator in a bounded-height container when actions are present #126 and fixing it is a visible behavior change — it needs its own change and a visual review.
• DataTableForObservableQuery sizes itself with a ResizeObserver and a pixel height rather than the flex pattern its sibling uses. It inherits the bounded parent so DataPage clips its paginator in a bounded-height container when actions are present #126 is fixed for it too, but converging the two is a behavior change beyond this defect.

Sequencing. #121 (footer Cancel) touches the same footer block as #120 and should be rebased onto this rather than developed in parallel. Open PR #107 (PrimeReact 11) rewrites the stepper wrappers; if it lands first this needs a rebase onto it.

[woksin]

Reviewer context for 2c91e93 and 643ac2b, kept out of the description.

Where these came from. Independent code and security review. The security review found no concerns in the diff — and noted the stepper change actually closes a small disclosure, since React.Children.count counted false/null slots while toArray dropped them, so {isAdmin && <StepperPanel/>} gave a non-admin an inflated count, a live Next onto an empty panel, and a stepErrors[phantom] ?? false that read as valid. The code review required changes, and both of its merge-blocking findings are fixed here.

The specs could not fail on the DataPage defect, and that is now fixed at the root. Reverting the primary production change of c5ee3e1 previously left all 375 tests green: every assertion was on an inline style declaration, and jsdom applies no stylesheets, so it cannot tell an honored height: 100% from an inert one. The runner now processes exactly one stylesheet (css.include scoped to allotment/dist/style.css), so the specs read computed values and an unpositioned pane reads as static / auto. That same mutation now fails.

The containment claims also assert the paginator exists before claiming where it sits — scrollRegion.contains(paginator) was satisfied by the paginator being absent, since Node.contains(null) is false. Forcing it never to render previously left for_DataPage 40/40 green; it now fails four specs.

Route taken for the details path. The stylesheet is imported by the component, which is this package's actual convention — every stylesheet here is imported by the component that needs it, rollup.config.mjs externalises /\.css$/ so the import survives verbatim, package.json declares sideEffects: ["**/*.css"], and allotment is already a hard dependency exporting ./dist/*. Verified in the built output (dist/esm/DataPage/DataPage.js and the cjs twin) and in the Storybook bundle. Drag-to-resize is preserved; replacing Allotment was not needed.

Six mutations, each apply → red → capture → restore → green in one trap … EXIT command, every capture ending identical to pre-mutation. The two that matter most: putting the common path back inside the split view now fails (it was the previously-undetectable one), and commenting out the stylesheet import yields ['static / auto', 'static / auto'] against an expected ['absolute / 100%', …] — precisely the diagnosed failure.

One mutation came back green and was worth more than the five that went red. The first attempt at the dialog's clamp passed, because the clamp added to the shared body module masked it — an assertion guarding an invariant that a different module establishes. The behaviour the dialog's own clamp uniquely owns is that Previous moves the wizard rather than absorbing one click per vanished step; a spec for that is what kills it.

Browser spec: assessed and not built. Issue #126 asked for one. There is no browser-mode capability in this repo — no @vitest/browser, playwright, puppeteer, @storybook/test-runner or webdriverio, and vite.config.ts pins environment: 'node'. Adding it means a new test stack, a second Vitest project (the current config is single-project with fileParallelism: false), a playwright install --with-deps CI step and browser-binary caching, for one assertion. What the computed-style change buys instead is the ability to distinguish honored from inert; what remains genuinely uncovered is layout itself — "the paginator is within the visible frame and clickable" still needs a real browser.

Gate. tsc -b exit 0; eslint 0 errors and 0 warnings; root build and full rollup clean; Storybook builds; markdownlint clean. 394 specs pass, 0 fail (375 + 19). One pre-existing dead eslint-disable comment in DataPage.tsx was removed to reach zero warnings — confirmed against git show HEAD: that it warns there too, so it is not from this PR.

Found, not fixed — both outside this PR.

• useCommandFormContext() identity is load-bearing and unguarded. CommandStepperContent's error effect memoizes on [stepFieldNames, getFieldError]; if the context object is recreated per render, getFieldError's identity changes every render, which fires the effect, which sets state, forever. A spec run hung on exactly this. Every existing stepper spec uses renderToStaticMarkup, which never runs effects — so no spec in this repo can see it. Arc presumably memoises the context, but nothing here asserts it, and a consumer that does not gets an infinite render loop.
• scripts/copy-css.sh copies storybook-static/assets/*.css into dist/ when a local Storybook build exists — its find excludes node_modules, dist and .storybook, but not storybook-static. Thirteen Storybook artifacts landed in dist/{esm,cjs} on a local build. CI never hits it; a local yarn build && npm publish would ship them.

Deliberately left. minHeight: '20rem' is an unconditional floor rather than the documented fallback, with no opt-out — a page inside a bounded 10rem ancestor is forced to 20rem and overflows. The fix wants a minHeight prop defaulting to '20rem', which is a public-API addition and belongs in its own minor PR. Fragments still count as one step (honestly documented and pinned by a spec); #120 should be closed noting that.

Label corrected from minor to patch — both commits are bug fixes, no public surface is added or removed, and .ai/rules/pull-requests.md maps bug fixes to patch.