Repository navigation
finding(plugin-charts): scatter is NOT MEASURED for the degenerate-magnitude class — its all-positive CONTROL drew zero marks, so the sweep's zeros say nothing #7171
Description
Filed by the domain:ui PM seat from a measurement in objectui#7147 (PR #7169). The implementer reported it rather than filing, per the report-don't-duplicate fence on that surface — but unlike objectui#7148's finding, this one has no existing card to attach to, so it gets its own.
What is NOT known
objectui#7147 swept 8 chart families × 9 datasets = 74 tiles in real Chromium, MD5'd and pixel-diffed against a literally empty div, to find where a degenerate magnitude (zero, negative, null, unparseable) produces a chart that misrepresents its data.
It reached verdicts for seven families:
| family | verdict |
|---|---|
| pie / donut / funnel / treemap | defective — fixed in PR objectui#7169 |
| bar / radar | measured and clean — deliberately untouched |
| scatter | ⛔ NOT MEASURED |
Scatter is not clean. It is unmeasured, and the distinction is the whole point of this card.
Why the sweep could not read it
Every scatter tile drew zero marks — including the all-positive CONTROL (markN 0).
A control that returns zero is no control. With the control dead, scatter's zeros on the degenerate datasets carry no information: they are indistinguishable from "the fixture never rendered a scatter at all." Reporting scatter as clean on that evidence would have been a false green across nine datasets.
The cause is identified: scatter takes two measures (x and y), and objectui#7147's fixture supplies a single measure. The fixture does not bind scatter, so scatter never had a chance to draw.
What a correct measurement needs
- A two-measure fixture for the scatter arm, so its all-positive control actually draws marks. ⛔ Until that control returns non-zero, no scatter reading means anything.
- Then the same nine datasets, with the degenerate value applied to each measure independently and to both — a scatter can plausibly degenerate on x, on y, or on both, and those may not behave alike.
- Per-tile screenshot hashing against per-family all-positive and one-row controls, the instrument that produced verdicts for the other seven families. A hash collision between a degenerate tile and a legitimately-sparse one is the finding; that is how pie's full-circle case and treemap's four-datasets-to-one-image case were caught.
Blocked, and on what
pm:blocked on PR #7169 landing — it is live in packages/plugin-charts/src/AdvancedChartImpl.tsx, which is where any scatter arm work would sit. Unblocks automatically.
Why this is worth tracking rather than dropping
Related
objectui#7147 / PR objectui#7169 (the sweep and the fix) · objectui#7148 (silent row drop) · objectui#7140 (the empty-frame early return) · objectui#7146.
Activity
os-warren commentedon Sep 1, 2026
Claim —
domain:uiexecution seat, R19. Unblocked: PR #7169 landed (5eddeeb68).session_012wwHa4aaFybxXrfmfHioDMclaude/issue-7171-scatter-degenerate-sweepui· Worktree: dedicated, offorigin/mainplugin-chartshas no other open PR. The three landed answers inAdvancedChartImpl.tsxare all in the tree and verified (no-positive-flow3,no-positive-magnitude1,unsized-rows1,omitted-rows1).Clause ② — not engaged
A sweep is a measurement. If it finds a defect, the repair shape is the one #7146 / #7147 / #7148 already landed at normal tier — a refusal or a footnote inside the renderer, no declared surface moved.
This card exists because scatter's result in #7147's sweep was NOT MEASURED, and NOT MEASURED is neither red nor green. Its all-positive control drew zero marks (
markN: 0on every scatter tile, the control included), so a control that returns zero is no control, and none of its zeros carry information.⭐ And the predicate that fixed the other four families almost certainly does not transfer. Pie, funnel and treemap size by magnitude, so
Number.isFinite(v) && v > 0is the right test there. Scatter plots position — a negative or zero coordinate is perfectly legitimate data. Mechanically reusingno-positive-magnitudeon scatter would refuse correct charts. The dispatch order carries that as a fence: the degenerate class for a two-measure positional chart has to be derived, not copied.Generated by Claude Code
os-warren commentedon Sep 1, 2026
os-dev-report
{ "issue": 7171, "status": "done", "branch": "claude/issue-7171-scatter-degenerate-sweep", "pr": "https://github.com/objectstack-ai/objectui/pull/7197", "premise_still_valid": true, "summary": "THE SCATTER CONTROL DREW — 3 of 3 marks, 245px of mark area, on a run whose live console positive control also fired; every reading below is therefore a measurement, not a zero. A2.1 confirmed: scatter's XAxis and YAxis are both type=number bound to xAxisKey and series[0].dataKey, and #7147's single-measure fixture put a category column on the numeric X axis — reproduced as its own tile, 0 marks, byte-identical to five other broken shapes. The sweep (33 tiles, real Chromium, one page load per tile, 520x360 above the 280 floor) found three fix-grade defects: rows unplaceable on BOTH coordinates rendered byte-identically (0.000%) to a scatter handed no rows at all; six distinct authoring failures shared one image; and one placeable row among three was 99.75% pixel-identical to a genuinely one-row scatter. Shipped a positional predicate — no-plottable-points refusal and unplotted-points footnote, scatter only — DERIVED not copied: #7147's magnitude predicate would have blanked the all-negative (3 of 3 marks), all-zero (2 of 2) and mixed-sign (3 of 3) tiles, which are correct charts. Zero variance, constant-x and constant-y are measured-and-DECLINED (A2.3's axis-domain-collapse hypothesis did not reproduce as collapse; the real failure is an axis that renders no scale at all). 14 tiles moved, 19 byte-identical, all three landed answers still firing under their own codes.", "tests": "Union run on the final commit 9953dd5fc (git rev-parse --short HEAD), tree clean. `pnpm exec vitest run packages/plugin-charts/src/AdvancedChartImpl.degeneratePosition.test.tsx` -> `Test Files 1 passed (1)` / `Tests 40 passed (40)`; `pnpm exec vitest run packages/plugin-charts/` -> `Test Files 40 passed (40)` / `Tests 371 passed (371)`. The FIRST test asserts the control draws a non-zero mark count, as an assertion rather than an observation. ABLATION, direction and counts predicted in writing before running (leg A 17 red, leg B 7 red, disjoint): leg A neutered the refusal guard (points.plottable === 0 -> === -1) and gave 17 failed / 354 passed, split 15 refusal block + 1 seam + 1 unprojectedSeriesDimension; leg B inverted the note guard (unplotted <= 0 -> >= 0) and gave 7 failed / 364 passed, all in the note block. Both legs proved mutation on disk by anchor count AND blob hash (aceb44fc -> d740c74f in A, -> 53138e45 in B; anchor 1->0, injected 0->1), and proved restore by STATE (git diff HEAD empty, git diff --cached empty, blob back to aceb44fc), under an absolute-path trap using `git checkout HEAD -- path` rather than a bare checkout. The browser instrument was rebuilt (`vite build`) on each tree and the marker strings verified present in the bundle the browser actually reads before every reading. GATES on 9953dd5fc: type-check exit 0 with tsc --listFiles confirming all four edited files in the compiled set (1 hit each); package lint `299 problems (0 errors, 299 warnings)` exit 0; check-changeset-presence/no-major/overwrite/fixed all exit 0; check-control-bytes `OK (scanned 5956 tracked text file(s))`; check-vi-mock-specifiers/inherit `OK (4106 tracked source file(s) ... 526 carry a mock)` — counts moved from #7169's 5942/525, which proves the gates saw the new test rather than skipping it as untracked; check-i18n-keys/drift, check-self-import, check-phantom-deps, check-esm-specifiers, check-side-effects-array, check-shell-escape-residue all exit 0; check-lint-coverage `46/46 packages linted, 0 with outstanding errors`; check-type-check-coverage `41/41 packages compile their tests`. NOT MEASURED, reported as such and not as failures: check-eager-closure exited 2 with `not applicable to a local run — this check compares a pull_request merge ref` (CI owns it); the probe's first vite build failed to resolve @object-ui/core because the dependency closure was unbuilt, built with `pnpm --workspace-concurrency=2 --filter '@object-ui/plugin-charts^...' build` and re-run to a real verdict. DECLARED NARROWING: repo-wide `pnpm lint` not run locally, CI owns it — the per-package run is a measurement: eslint's own config selected 53 files (--format json count, up from #7169's 52, the delta being the new test file), all four edited files in that set with 0 errors, and eslint.config.js extends tseslint.configs.recommended with zero matches for parserOptions.project or projectService, so no rule is type-aware and this diff cannot move a verdict in a file it does not touch.", "mcp_calls": "10", "open_questions": [], "out_of_scope_findings": [ "filed as #7194: a multi-series scatter draws every series at the FIRST series' y values — measured, two series over `ym` and `zm` produced 4 symbols at only 2 distinct positions, each drawn twice, and `zm`'s values (5 and 90) appear nowhere on the plot; good data, confidently wrong picture, and it needs a shape decision rather than a one-liner", "filed as #7195: an all-boolean scatter coordinate draws nothing and says nothing — deliberately left unguarded by this PR, because a boolean beside a genuinely numeric row draws every mark (3 of 3), so rejecting booleans would put a false 'not drawn' count over visible points; the first draft DID reject them and the pin went red, which is how the mixed reading was found" ] }Notes on the two report fields PM reads first
premise_still_valid: true— the card's diagnosis was correct in full. Scatter takes two measures; finding(plugin-charts): funnel and pie handed all-zero rows render a pixel-identical blank tile — same silence as #7140, a different trigger #7147's fixture supplied one; the control was dead for exactly that reason. Verified by reproducing finding(plugin-charts): funnel and pie handed all-zero rows render a pixel-identical blank tile — same silence as #7140, a different trigger #7147's own fixture as a tile in the new instrument (scatter--fixture7147, 0 marks) beside a two-measure control that drew.Generated by Claude Code
os-warren commentedon Sep 1, 2026
✅ LANDED — PR #7197 squash-merged to
main(f08cbd996)Verified by content with live controls. The strongest available shape here is differential: every previously-landed seam answer is still in the same file, so they discriminate a real landing from a broken probe.
no-plottable-points(the new refusal code)unplotted-points(the new note)AdvancedChartImpl.degeneratePosition.test.tsxno-positive-flow(#7146)no-positive-magnitude(#7147)unsized-rows(#7148)⭐ All five seam codes now coexist in
AdvancedChartImpl.tsx, which is the mutual-exclusivity assertion holding in the tree rather than only in the suite.The card's premise was NOT MEASURED, and this closed it the right way
This card existed because #7147's scatter result was neither red nor green: its all-positive control drew zero marks, so none of its zeros carried information.
The control drew — 3 of 3 marks, 245px of mark area, on a run whose live console positive control also fired. And the first test asserts that non-zero mark count rather than observing it, so the sweep cannot silently degrade back into blank tiles.
⭐ Better still, the old zero is now explained rather than merely superseded: scatter's
XAxisandYAxisare bothtype=numberbound toxAxisKeyandseries[0].dataKey, and #7147's single-measure fixture put a category column on the numeric X axis — reproduced as its own tile, 0 marks, byte-identical to five other broken shapes. A known artefact beats an unexplained one.⭐ The predicate was DERIVED, and the derivation is proved by what copying would have broken
The fence I set was: ⛔ do not reuse #7147's magnitude predicate. It was not merely respected, it was measured:
⇒ Mechanical reuse would have refused three working charts, which is worse than the silence being fixed. Scatter plots position; negative and zero coordinates are legitimate data. That is why a separate scatter-only positional predicate earns its existence.
What the sweep found
14 tiles moved, 19 byte-identical, and all three landed answers keep firing under their own codes.
⭐ #7195 — their own pin caught their overreach
An all-boolean scatter coordinate draws nothing and says nothing, and it was deliberately left unguarded: a boolean beside a genuinely numeric row draws every mark (3 of 3), so rejecting booleans would put a false "not drawn" count over visible points. And the discovery route is the part worth keeping:
That is the ablation discipline running forward, as a design instrument rather than an after-the-fact proof, and it prevented shipping a refusal that lies about what is on screen.
Filed, not ridden
#7194 — graded p1⚠️ Unlike everything this card fixed, that needs no degenerate input — valid data, confidently wrong picture, and invisible to inspection because 4 symbols at 2 positions look exactly like legitimate coincident points. It needs a shape ruling and is escalated on #5560.
bugthis round. A multi-series scatter draws every series at the first series' y values: two series produced 4 symbols at 2 distinct positions, each drawn twice, with one series' values appearing nowhere.Clearing
pm:dispatched.Generated by Claude Code