docs(labels): refresh stale observed-value statistics in survey_documentation

[elleggert]

While working through reviewer comments on the paper I went to verify the cholesterol figures in data/labels/survey_documentation/ and found the numbers there don't match the released dataset. That docs snapshot predates several label-preparation steps, so 18 pages publish statistics that no longer describe the data we ship.

This PR regenerates those blocks. Nothing else.

What changed

Regenerated from the released full-cohort label files — last_labels.json (md5 521c157a…) and context_labels.json (md5 220f5946…), the same bytes served by the public bundle. Per page, only these are rewritten:

• the **Total observations** / type-enforced / unique / raw types line
• the **Type-enforcement rejections** line
• the | stat | value | table
• the **Top N most frequent values** table
• the _Generated …_ provenance footer

Prose, ## Source, ## Question, ## Answer options, ## Git history and ## Notes are byte-identical. The diff contains no line outside those five kinds.

Representative corrections

Variable	Published	Released data
Watch_BasalEnergyBurned	min −4666 kcal	500.3
Watch_RespiratoryRate	min −1, max 106	5, 39
Watch_HeartRateVariabilitySDNN	min 0.002341 ms	5.18
WeightKilograms	min/max 30 / 300 (slider clamps)	31.3 / 223.6
BMI_values	max 125	58.04
age	median 32	39 (de-identified birthdates)
Watch_VO2Max	max 99	59.89
TotalCholesterol	median 9.49	4.55 mmol/L
Hdl / Ldl	median 2.77 / 5.61	1.32 / 2.61 mmol/L

Full list of touched pages: BMI_values, WeightKilograms, Hdl, Ldl, SystolicBloodPressure, TotalCholesterol, framingham_risk, age, Watch_BasalEnergyBurned, Watch_HeartRateVariabilitySDNN, Watch_RespiratoryRate, Watch_RestingHeartRate, Watch_StandTime, Watch_VO2Max, Watch_WalkingHeartRateAverage, vigorous_act, sleep_time, happiness.

One additional one-word change, called out explicitly: cardiometabolic_labs/summary.md described TotalCholesterol as (mg/dL), which the refreshed statistics in the same PR directly contradict. Changed to (mmol/L).

Scope / non-goals

• No data, code or API change. The released dataset is unaffected — it was already correct; only these markdown pages were stale.
• Nothing outside survey_documentation/. The bundle-side docs (data/labels/README.md, RELEASE_NOTES.md, which live in Google Drive and the Dataverse deposits) carry no per-variable statistics, so no dataset re-version is needed.
• Pages with categorical/ordinal value tables were checked and are unaffected.

[elleggert]

Short version of what this does and how it was checked.

What we did. The Observed values block on these pages was generated on 2026-04-28 from a pre-release snapshot of the label files and never regenerated, so 18 pages published statistics that don't describe the data we ship. Regenerated those blocks from the released label files (last_labels.json md5 521c157a…, context_labels.json md5 220f5946…). Per page only the counts line, rejections line, stat table, top-N table and provenance footer are rewritten — prose, source references, git history and notes are byte-identical.

How we checked it. Every number was re-derived a second time through an independent code path (numpy percentiles, ddof=1) and diffed against what's committed here: 18/18 pages reproduce exactly — every stat cell, count, unique and Top N header, with all 18 footers citing the md5s of the files actually read.

Cross-checked against labels_wide.parquet, a separate release artifact written by a separate pipeline step: 10 of 18 match it exactly (same n, same sorted values to 1e-9) — BMI_values, WeightKilograms, Hdl, Ldl, SystolicBloodPressure, TotalCholesterol, framingham_risk, age, vigorous_act, sleep_time. The remaining 8 (the seven Watch_* metrics and happiness) cannot match by construction: labels_wide.parquet holds one row per user, while these variables store a per-user series and the doc block summarizes every observation — the same convention the original generator used.

Also confirmed no label-name collisions between last_labels.json and context_labels.json for these 18, so there is no ambiguity about which file each page reads.

On the large count drops (TotalCholesterol 9,937 → 3,558, SystolicBloodPressure 9,867 → 3,691, age 57,939 → 11,893): these reflect the release deduplicating clinical targets to one record per user, not a counting change here — which is why labels_wide.parquet matches those pages at exactly the user count.

Scope: docs only. 19 markdown files, all under data/labels/survey_documentation/; no code, no data, no API change. The released dataset was already correct — only these pages were stale.