viz: scatter, histogram, stacked bar + analyse: filter_rows

[moonlight-lupin]

Summary

Three chart types and a filtering primitive — the gaps that forced agents to hand-roll the same code every run — plus the docs-accuracy fix that reviewing #24 surfaced.

Charts (viz.py, HTML only — the xlsx path refuses them with the existing message)

scatter_chart(x, y, …, trend_line=False)	Correlation / outlier spotting. Both axes float via _nice_ticks (matching line_chart). The OLS overlay spans only the observed x-range and carries slope / Pearson r / n in its tooltip.
histogram(values, bins=10, …)	Bin count (equal-width) or explicit edges like [0,30,60,90,365]. Y-axis forced to 0; bars touch.
stacked_bar(data, …)	Takes a pivot() result directly, or {segment: [(cat, value)]}, or {categories, series}. Legend with swatches.

Three judgment calls worth reviewing:

• histogram edges are half-open [lo, hi) except the last, which includes its upper bound — otherwise the maximum observation silently vanishes.
• scatter_chart draws no trend line when x has zero variance. There is no defensible slope for a vertical cloud, and drawing one would invent a relationship the data doesn't contain.
• stacked_bar stacks negatives below the zero line. Folding a credit note into the positive stack would inflate the very bar it reduces.

All three reuse the engine's parse_number when reachable ('15%' → 0.15, '1.2m' → 1200000, '(500)' → -500) rather than inventing a second numeric dialect — otherwise the chart and the working paper could show different numbers for the same cell. Non-numeric inputs are skipped and reported, never plotted at the origin where they'd read as a real zero.

filter_rows(header, rows, filters) (analyse.py)

12 operators (== != > >= < <= in not_in between not_between contains is_empty not_empty), ANDed, returning (rows, report) with in/out totals and per-filter removed counts.

Two comparison rules exist because the alternative is a plausible wrong answer, not an error — both found by testing, not by reading:

1. Dates compare as dates. parse_number('15/02/2026') returns Decimal('15022026') — it strips the separators. A number-first coercion therefore sorts 15 Feb after 1 Mar. Values that look like dates are now tested as dates first.
2. A type mismatch is incomparable, not a string compare. The string fallback made 'n/a' > 1000 true. Such rows are excluded and counted in report["incomparable"].

between is inclusive of both ends (finance quotes ranges that way). Unknown columns and operators raise — a typo must not look like "no results".

Docs accuracy

The README's "no network calls / no cloud OCR / no credentials" and DATA-HANDLING.md's "no external APIs, no third-party uploads" predated the opt-in vision image extract shipped in #19, which calls a user-configured vision endpoint and needs a key. All three claims now carve out that one exception explicitly. SKILL.md and COMPATIBILITY.md already disclosed it correctly — it was the top-level pitch that overstated, which is the wrong way round.

Verification

Wired through _viz_block and schemas/dashboard-spec.schema.json (new block types, their properties, and per-type required-property rules; additionalProperties:false still holds).

Five regression tests, all mutation-checked — each was confirmed to fail when its guard is narrowed. One of them initially escaped: the stacked-negatives test only asserted relative ordering, which holds even when both segments stack upward. It now measures against the zero line.

Gates: data-lint OK · 101 pytest · 68 standalone · 19 runtime · 12 schemas · 6/6 golden plans · quickstart smoke.

🤖 Generated with Claude Code