Add bundle_gas: Group A Strep across Epic Cosmos, NNDSS, and ABCs

[MadiBal]

Combines Group A Streptococcus surveillance from three sources into six long-format parquets.

Stacked on #186. The base of this PR is ingest-cosmos-gas, so the diff here shows only the abcs_gas source and the bundle. Merge #186 first; this PR's base will retarget to main automatically.

Initialized with dcf::dcf_add_bundle("bundle_gas", source_files = ...), so process.json carries type: "bundle" and scripts: build.R.

Outputs

All six share the bundle conventions: geography holds state names (or "United States"), time is ISO YYYY-mm-dd period-end, value is the plotting column, keyed by a measure identifier column.

Parquet	Source	Rows	Grain
epic_gas.parquet	epic_gas	39,312	state + national, quarterly, by age — n_strep_throat / pct_strep_throat / n_patients, plus a per-measure suppressed flag
nnds_stss.parquet	nnds	24,752	state + national, weekly — stss_cases_weekly / stss_cases_cumulative
abcs_gas.parquet	abcs_gas	674	national, annual, by age/sex/race — case and death rates and counts
abcs_gas_syndromes.parquet	abcs_gas	135	national, annual — 5 clinical syndromes
abcs_gas_resistance.parquet	abcs_gas	180	national, annual — 6 antibiotics + isolate count
abcs_gas_emm.parquet	abcs_gas	421	national, annual — 16 emm types + other

Two things that most need a second pair of eyes

1. NNDSS is published cumulatively, and I de-accumulate it. The raw streptococcal_toxic_shock_syndrome column is a year-to-date running total that resets each MMWR year — national 2024 runs 5 → 647 across weeks 1–52. Plotted as-is it would show a sawtooth, not incidence. build.R differences it within geography-year into stss_cases_weekly and keeps the published cumulative series alongside; the two are not additive. The earlier bundle_gas attempt on the stale gas branch did the same, and the nnds source measure_info documents the column as "Cumulative year-to-date", so this looks right — but it is the one substantive transformation in this PR and worth confirming.

Relatedly, 27 of 12,376 weekly increments come out negative, because NNDSS revises earlier weeks downward. I left them as reported rather than clamping to zero, and build.R logs the count. Clamping is a one-line change if preferred.

2. Which "ABCS" this is. data/abcs on main is invasive pneumococcal disease by serotype — a different pathogen, with no GAS measures. The Group A Strep ABCs data (CDC dataset 9y49-tura) existed only on the stale gas branch and was never merged, so the first commit here brings abcs_gas onto main-line development.

Known gap, pre-existing: that source's own measure_info.json uses generic keys (case_rate, syndrome, pct_resistant, emm_type) that do not match its actual column names, so dcf_check reports 36 undocumented columns on the source. The bundle documents those columns itself via levels inheritance, so the dist files and generated docs are complete — but the source-level check still fails. Happy to fix those 36 entries in a follow-up or here.

Verification

• dcf::dcf_process("bundle_gas") runs clean and is deterministic — re-running produces no diff.
• measure_info.json parses; 0 undocumented columns across all 6 parquets, and all 34 levels source_id references resolve against the contributing sources' measure_info.json.
• _bundle block covers every parquet, its sources equals the union of the per-file sources, and every listed source_file exists.
• scripts/build_docs.R regenerated (14 bundles, 46 sources); bundle_gas renders a short name plus a description or value list for all 29 columns.

Smaller decisions

• Four separate ABCs parquets, per the upstream split, but each melted to long format to match the value-column convention rather than kept wide as the earlier attempt did.
• Epic suppression is mapped per measure: the upstream numerator flag covers n_strep_throat and pct_strep_throat (the percent derives from that same cell), the denominator flag covers n_patients.
• Territories and non-state NNDSS jurisdictions (e.g. New York City) are dropped, following bundle_enteric_diseases.
• abcs_gas is national only (a catchment area, not the whole US), and the three sources' time spans differ (Epic 2017→2025, NNDSS 2022→present, ABCs 1997→2023), so cross-source comparison is limited to the overlap. Both are documented in the bundle README.

🤖 Generated with Claude Code

[MadiBal]

Superseded by #189, which rebuilds this bundle on top of #188. Changes since: Epic GAS is read from epic_resp_infections rather than a separate source, geography_fips is carried alongside the geography name, and ABCs Group B Streptococcus is included (it was missing entirely). The abcs_gas ingest also needed a substantial rewrite — CDC restructured the emm and syndrome topics in their 2026 release, and re-running the version here produced an empty emm file and 100x-too-small resistance values.