[NSchatz]

What

Integration tests in every starter that chains packages, through the libraries' public APIs only,
on synthetic data from @cosyte/synth (the DICOM starter's own hand-written header, since synth has
no DICOM generator). Each starter's tests live in test/integration/, tagged by file name:

• *.fast.test.js, the fast slice: npm test now runs them, so starters.yml runs them on every
  pull request and its existing All starters check stands for the slice.
• *.full.test.js, the full suite: seed sweeps and many-process runs. npm run test:full runs every
  test and fails unless the helpers in test/integration/support.js have 100% line, branch and
  function coverage.

The pipelines, each tagged fast and full:

Starter	Pipeline	Fast	Full
hl7-mllp-listener	7 HL7 v2 kinds from synth over a real MLLP link on 127.0.0.1 and back through @cosyte/hl7, each ACK correlated; the starter's listener on each ADT trigger	2 seeds per kind	150 seeds per kind on one link; 4 senders at once
hl7-to-fhir	hl7, transform, fhir validation, terminology (UCUM, code-system resolution), deid over the FHIR Bundle	2 seeds per kind	250 seeds per kind
deid-hl7	deid over 7 HL7 v2 kinds, the output parsed back	2 seeds per kind	250 seeds per kind
x12-remittance	835, 837P, 837I, 837D round trip; deid over X12; the de-identified 835 reads as the same payment	2 seeds each	250 seeds each
x12-eligibility	271 and the 270 built for it, both round-tripped and read	3 seeds	300 seeds
ccda-summary	CCD and Referral Note round trip; terminology on every codeSystem OID; deid over C-CDA	2 seeds each	150 seeds each
astm-lab-results	framed upload decoded and re-framed byte for byte; orders round trip	3 seeds	100 seeds, 1 to 8 results
ncpdp-claim	B1, B2, B3 and the starter's B1 round trip; deid over NCPDP Telecom	2 seeds each	250 seeds each
dicom-metadata	Part 10 round trip; deid, the de-identified bytes round-tripped, UID remapping	1 header	20 maps by 10 passes
cli-tour	cosyte as a child process: convert piped into validate - and redact -; validate and fmt over 8 formats; redact refusals; map-codes through a ConceptMap	seed 1	12 seeds per HL7 kind and per format

Workflows:

• integration.yml: the full suite, every starter on Node 22 and 24, one Integration (full)
  check. Nightly on main, on a pull request that changes an integration test, on dispatch, and
  through workflow_call with package and artifact: each starter whose lockfile carries the
  package installs the artifact's tarball as the dependency and as an npm override, checks that every
  copy in the tree is the tarball's version resolved from the file, then runs the suite.
• integration-self-test.yml: packs the locked @cosyte/hl7 under a version the registry does not
  have (<locked>-selftest.<run id>.<attempt>, checked absent with npm view), uploads it, and
  calls integration.yml with it. The gate's install step fails unless that marked version is what
  every starter installed.

Every lockfile was regenerated from scratch, so each starter is on the current @cosyte/* 0.1
releases. Five starters gain test-only devDependencies (@cosyte/deid, @cosyte/terminology, and
for cli-tour @cosyte/ccda, @cosyte/fhir, @cosyte/x12). Each starter README gains an Integration
tests section that says what its tests assert; AGENTS.md, CLAUDE.md and the root README describe the
tags and the workflows. pathways is not covered: the published @cosyte/pathways 0.0.1 needs Node 24
and says it is not ready to run.

Why

The libraries are each tested on their own; this suite tests them together, the way an integration
uses them: a parser's output fed to the transform, the transform's to the validator and the
de-identifier, the CLI's output piped into the CLI. The fast slice runs on every pull request, the
full suite nightly, and a library's release can run it against the tarball it is about to publish.

Proof

• Gate: the four steps in AGENTS.md passed, 7 of 7 commands (no-emdash, claude-card, npm ci and
  npm test in every starter on Node 24 and on Node 22, actionlint).
• npm run test:full in every starter on Node 22 and on Node 26: all pass, two todo, helpers at
  100% line, branch and function coverage.
• CI on this pull request, all green. The fast slice is Starters (check All starters): 30
  seconds from start to finish, 20 jobs, the longest 14 seconds. The full suite (Integration,
  check Integration (full)): 62 seconds, the longest job cli-tour at 45 seconds.
• The self-test (Integration self-test) packed @cosyte/hl7 0.1.8-selftest.42873672033.1 and the
  gate's log shows each starter installed it from the tarball, for example cli-tour on Node 22:
  node_modules/@cosyte/hl7 0.1.8-selftest.42873672033.1 file:../../../_temp/override/...tgz.

Red first. The tests were written against the base and run red until the scripts, helpers and
expectations were right. The red runs that shaped them: hl7-to-fhir's ORU check failed when the
vital-signs starter profile was applied to lab Observations; the X12 de-identification check failed
on a rendering provider who shares the patient's synthetic name; dicom-metadata's UID check failed on
the assumption that a new map gives new UIDs (the remap derives from the original UID); the
x12-remittance full run failed its coverage threshold at 97.37% branch coverage until a dead branch
went. Two tests stay red as todo, one per library defect below.

Library defects found, each a todo test that shows it, none worked around:

1. @cosyte/deid 0.1.5, DICOM: deidentifyDicom writes De-identification Method (0012,0063) as
   one 113-character Value, over the 64 characters PS3.5 allows an LO, and @cosyte/dicom reports
   DICOM_DEIDENT_METHOD_VALUE_OVER_LENGTH. Repro, in dicom-metadata:
   deidentifyDicom(parseDicom(syntheticCtHeader())).warnings.
2. @cosyte/cli 0.1.8: NCPDP is read as SCRIPT only. A Telecom D.0 B1 (generateB1({ seed: 1 }))
   is undetected (exit 65, CLI_FORMAT_UNDETECTED), and with --format ncpdp fails with
   NCPDP_SCRIPT_NOT_XML, although --help lists ncpdp for validate, inspect and fmt.

Seams the tests pin down, not defects: synth's ORU^R01 leaves OBR-25 empty, so the transform
withholds the DiagnosticReport; synth's OBX-6 units carry no UCUM coding system, so they never reach
Quantity.code; @cosyte/terminology resolves neither three HL7 v3 code systems the transform writes
nor four HL7 v3 OIDs and the NCI Thesaurus OID in the synthetic C-CDA; the CLI does not detect NCPDP
SCRIPT without --format.

Choices

• The fast slice runs inside npm test, so All starters is its one check rather than a new one.
  Undo: drop test/integration/*.fast.test.js from each test script and add a fast job to
  integration.yml.
• Tags are file names, not test names, so a copier's npm test stays one command. Undo: rename.
• The release override sets the tarball as the dependency and as an npm override, so a copy deeper in
  the tree (the CLI's own @cosyte/hl7) is replaced too. Undo: drop the .overrides line in the
  install step.
• Library defects are todo tests, which report without failing the run, so the defect stays
  visible and the suite stays green. Undo: remove the todo option.

Kind: code