Analytics window, and the demo data that made it possible to build

[AllTerrainDeveloper]

Two things, and the order matters.

Why the demo data comes first

Analytics are impossible to build against an empty database and misleading to build against a handful of test rows. Five submissions make every chart look fine — no long tail, no skew, no segment small enough to be noisy, no date range wide enough to need a scale. The first honest look at a report is the moment somebody points it at a real form with four hundred responses, which is much too late to discover the bar chart cannot cope with a label longer than twelve characters.

The survey

Seven questions, chosen for what they make possible downstream rather than for what a pulse survey would ideally ask: two categorical dimensions to group by, a 0–10 recommendation score, a 1–5 rating, a lumpy small integer, a multi-select whose percentages deliberately sum past 100, and an optional free-text box answered by well under half.

These are real submissions. Every one goes through atf_process_submission() — the same function a stranger's POST reaches. Sanitising, validation, spam screening, storage, the counters and atf_entry_created all run exactly as in production, and the request carries a properly signed timestamp rather than being waved past the time trap. A seeder that wrote rows directly would produce data the pipeline could never have produced, and the first thing that would hide is a bug in the pipeline.

The population is not random, because random answers produce flat charts — every question at 1/n, every segment identical, every correlation zero, which looks like a working report and proves nothing. Each respondent gets a latent morale depending on their team and tenure, and their answers follow from it:

Team	n	Office days	Recommend	NPS	Satisfaction
Design	44	2.18	8.61	+52	3.86
Data	51	2.04	8.43	+39	3.80
Engineering	136	1.84	7.76	+14	3.51
Sales	77	2.82	7.19	−14	3.27
Support	63	2.40	5.78	−51	2.56

It comes back out. The form and every entry are tagged; removal deletes exactly what carries the tag, never "everything on the demo form" — that form has a working shortcode, so somebody will eventually answer it for real, and there is a test asserting their answer survives.

Developer mode

Gated on OpenStation's own per-user switch, read rather than duplicated. It writes hundreds of entries into whatever database this happens to be.

The preference is not the permission. It answers "show me these"; atf_edit_forms answers "you may use them"; both are checked on every route. Verified with dev mode off: GET and POST /demo both 404, the panel is gone, the dock row is gone, and the report itself still renders.

The report

Headline counts, submissions per day, a tally or distribution per question, NPS for any 0–10 question, and a cross-tab of every numeric answer by any categorical one — the difference between "the average is 7.5" and "Support is at 5.8 and everyone else is above 7".

Three things that look fine when wrong, so each has a test:

• A timeline includes its empty days. Keep only the days that had a submission and the gaps close up, so one spike and three weeks of silence renders as a steady trickle.
• NPS is not an average. Six sevens and eights average to a healthy-looking 7.5 and are an NPS of exactly zero.
• A mean needs its distribution. Everybody answering 3, and half answering 1 with half answering 5, are opposite findings with the same mean.

No charting library. Every chart is ordinary elements sized in percentages — a bar chart is a list of labelled quantities, and as a list it can be read aloud, selected and searched.

Bugs found while building it

• Fifty submissions in one request exhausted a 128MB limit, failing as a bare 500 with the entries already written. WordPress adds every post it touches to the object cache and nothing evicts it; under WP_DEBUG, $wpdb keeps every query with a backtrace. Both freed per chunk, chunk size lowered, and a failed chunk now retries instead of abandoning the batch.
• The window's token block listed three roots, not the new fourth, so every var( --atfb-* ) in the report was invalid — which does not fail loudly: the declaration is dropped and a chart of coloured bars renders as a blank rectangle.
• The survey put notifications under settings, where nothing reads it. Harmless by luck, but the test asserting "no notifications" was passing against a key nothing uses and would have gone on passing after somebody added one in the right place.

Verified

516 PHPUnit (36 new), 202 Vitest (8 new), PHPCS clean. The full 500-submission run was driven from the browser end to end; the dev-mode gate was checked from both sides; the demo-data row's gate is mutation-tested.

🤖 Generated with Claude Code

https://claude.ai/code/session_0119U4sRRWGcTQdwYreTwpAp