Signal comparison qualification

Date: 2026-10-05. Local branch: codex/foldkit-0-166-0. Approved feature baseline: 7b5db30. No push, publication, version bump or deployment is included.

Delivered contracts

Pure errorBarGeometry returns readonly numeric stems and caps for either axis, preserves datum identity/order and rejects invalid or overflowing coordinates. Root and exact chart/errorBars exports were exercised by packed Bun, Node and strict TypeScript consumers without optional framework peers. Bounds remain caller-supplied metadata; no statistical inference is performed.

The native Foldkit model owns a separate captured baseline with copied readings, nested bounds and source snapshot. Capture/clear first roll back provisional gestures; late pointer facts cannot commit that obsolete gesture. Inspection, viewport and selection remain independent. Dataset replacement always clears the baseline, including reused revisions and IDs.

Comparison preserves exact current-minus-baseline arithmetic, ms/percentage-point units, quality, support zero, methods, bounds and captured provenance. Missing/Invalid and absent inspection remain explicitly unavailable. Selected error bars, baseline time/value references and HTML disclosures compose with existing clips, bands, domains and quality lanes. The route discloses all 13 actual application source files.

Task commits: 2915d08 geometry; 273a871 capture state; 7438bd0 derivation. Composition and this evidence are committed together after qualification.

Automated evidence

Each task observed RED before implementation and GREEN afterwards. Geometry: 6 tests; packed consumers: 5 tests. Baseline state plus existing input/state regressions: 20 tests. Derivation plus geometry/state/input: 24 tests.

Initial final Signal checks: 53 passing tests, 456 assertions across 10 files. Root bun run check and bun typecheck exited zero. Sequential bun run test: 520 pass, 1 skip, 0 fail before the review fix (Astro 74, Viz 211, promo 117, web 118). Pages-base production build generated 10 routes; Pages regression passed with 290 assertions. These prove host contracts and production output, not native device or assistive-technology acceptance.

Session logs: /tmp/comparison-all-signal.log, /tmp/comparison-task4-check.log, /tmp/comparison-task4-types.log, /tmp/comparison-task4-tests.log, /tmp/comparison-pages-build.log, /tmp/comparison-pages-test.log. Logs are local session evidence, not committed dependencies. Review server restarted from apps/promo at http://127.0.0.1:4321 after production qualification.

Direct native observations

Firefox Developer Edition on macOS, desktop 1909×1050, dark theme: hydrated Capture/Clear disabled with no inspection; keyboard inspected signal035; pin/capture retained inspection and captured 100 ms, bounds 92–108 and support 0. Resume and keyboard inspection of signal040 showed current 0 ms/0%, retained baseline 100 ms/0.2%, and exact differences −100 ms/−0.2 percentage points.

Fresh→Stale changed current provenance while the captured source stayed Fresh at capture. Missing signal010 latency and Invalid signal022 errors displayed unavailability and retained diagnostics. Replace explicitly captured signal022 with Stale provenance; Clear removed baseline marks/references without changing inspection. Recapture signal035 restored Fresh provenance. Range selection 20–60 retained it, temporary absent current inspection was explicit, and keyboard inspection of signal040 restored comparison. Desktop cards, sources, bounds and deltas remained readable without overlap in the observed viewport.

Open native gates at initial implementation

This run did not complete off-screen pan, 390/1280 layout and light-theme checks, table/source inspection or a new contextual screenshot. The browser foreground changed to another active task during the native pass; further interaction stopped to preserve that activity. Earlier M2 layout observations are historical and do not qualify the added comparison panel.

Held-pointer capture/clear, native Escape/lost capture, hardware touch/outside-plot scrolling, actual screen-reader announcements and representative-user comprehension remain unperformed. The native API exposes completed drags, not independently held pointer phases; connected DOM browser inventory was empty. Host rollback/render tests do not close these gates. No critical-system suitability or complete P2/P3/visx parity claim is made.

Implementation rulings

Named pure-helper lint scope extends only to baseline.ts: synchronous validated copying, not Effect handling. Cost if wrong: exceptions in that named helper.
Named pure-helper scope extends only to comparison.ts: synchronous arithmetic/projection. Cost if wrong: exceptions in that named helper.
Comparison promises reading content rather than object identity: Foldkit schema constructors copy nested records; only pure geometry promises datum identity. Cost: comparison readings may have different references, with exact data retained.
Named pure-render scope extends only to comparison-view.ts. Cost if wrong: existing renderer exceptions in that named file.
Fresh review uses the whole approved feature range from 7b5db30 rather than earlier, already reviewed branch work. Cost if wrong: unrelated historical changes are outside this review.
Native acceptance stays partial where the surface was unavailable or changed to other user activity. Cost: those native paths remain unqualified despite passing host regressions.

Independent review

One fresh read-only reviewer inspected 7b5db30..e30413e. Critical: none. Important: selected singleton/equal intervals were rendered by both the legacy fallback and the new comparison layer, obscuring dash cues. Minor: none.

Three full-composition regressions failed first: singleton unequal, singleton equal and wholly equal-band rendering. Selected intervals now belong to the comparison layer; unselected fallback marks and bands remain, and coincident fallback caps draw once. All three then passed; the complete comparison view file passed 10 tests/89 assertions. After the fix: root check/typecheck pass; sequential workspace tests 523 pass, 1 skip, 0 fail (promo now 120). All Signal tests: 56 pass, 482 assertions. Pages-base build still generates 10 routes; its regression passes 290 assertions. Logs: /tmp/comparison-review-{check,types,tests,signals,pages-build,pages-test}.log. Native observations above predate this composition fix; no native verification of the fix is claimed.

Reviewer boundaries and executor rulings:

Earlier branch work stands outside the approved range because it was reviewed previously. Cost if wrong: unrelated historical changes are not requalified.
Off-screen pan, responsive/light layouts, table/source and screenshot acceptance remain open. Cost: those actual presentation paths are unqualified.
Held-pointer/Escape/lost capture, hardware touch/scrolling, assistive technology and comprehension remain open. Cost: host contracts cannot establish actual operability or announcements.
Late MovedChartPointer may continue Following inspection after cancellation: gesture rollback and rejection of late commits are required; ordinary hover inspection remains independent of captured evidence. Cost if wrong: a late move may change the current inspected record, while baseline and committed view remain intact.
Missing/Invalid captured latency has no overview time marker because that marker is gated on numeric baseline reference. The design permits the marker; exact captured time/text remains available. Cost: no overview locator for nonnumeric captured latency.
Publication/deployment and critical-system suitability are outside this authorised local slice. Cost: local qualification establishes neither a deployed release nor operational certification.

No deferred minors. The three named pure-file lint exceptions and reading-content identity ruling above stand.

Follow-up browser qualification

The post-review browser follow-up now records390/1280 light/dark layouts, table/source agreement, an off-screen baseline and Capture/Clear/Escape during held pans and brushes. These paths were observed directly in the in-app browser on68f9f72. Independent capture loss, hardware touch/scrolling, actual screen-reader announcements and comprehension remain open. Earlier open-gate descriptions above retain their historical meaning.

Remaining scope

P2 error-bar geometry and P3 captured comparison are delivered locally. Wider axis/unit/time edge qualification, interaction acceptance and the other precision goals remain partial. Existing fixture decimal-authoring polish remains deferred; arbitrary caller values are never rounded. The available comparison-specific browser qualification is complete; remaining device/assistive-technology gates are explicit. Next recommended product work: design caller-owned event annotations and a shared event lane.