cosyte · NSchatz · Aug 11, 2026 · Aug 11, 2026
@@ -0,0 +1,22 @@
+---
+"@cosyte/mllp": patch
+---
+
+The `@cosyte/hl7` dev/test peer now installs from the npm registry instead of a
+vendored packed tarball, and `vendor/` is deleted.
+
+The tarball held `@cosyte/hl7` at `0.0.0` and never moved while that package
+shipped ten releases. One of them made the MSA-2 control-id echo byte-verbatim
+across the full escape alphabet. Because the pinned copy never picked it up,
+this package's own test suite went on asserting the old behaviour as a
+guarantee: an inbound `MSH-10` of `ID\X` echoed back as `ID\E\X`, which is a
+different control id on the wire, so an ACK the sender cannot correlate, so a
+resend, so a duplicate clinical message. That case is now pinned inverted, as a
+round-trip through the parser path that must produce the same bytes as the
+byte-copy path and must not warn.
+
+No runtime behaviour of this package changes: `@cosyte/hl7` remains an optional
+peer dependency referenced only from the `@cosyte/mllp/ack-from-hl7` subpath,
+and the emitted framing, ACK and warning-code surfaces are untouched. What
+changes is which version of the peer the tests run against, and therefore what
+they can prove.
@@ -17,13 +17,15 @@ name: Em-dash gate
 # the script it calls refuses to print OK from a scan that did not read everything it
 # was handed, and why every one of those refusals was checked red before this landed.
 #
-# mllp tracks ONE binary, `vendor/cosyte-hl7-0.0.0.tgz` (the packed `@cosyte/hl7`
+# mllp used to track ONE binary, `vendor/cosyte-hl7-0.0.0.tgz` (the packed `@cosyte/hl7`
 # devDependency, 472,097 bytes, 1,774 NUL bytes). It is why this repo runs website's
 # NUL-exclusion variant of the script rather than the text-only one hl7, fhir,
 # knowledgebase, pathways, x12, ncpdp and dicom run. The reasoning, the cost, and the
 # measurements behind that choice are in the script header; the short version is that
 # a compressed stream can contain the em dash bytes by coincidence, and that red would
-# have no fix.
+# have no fix. The tarball left on 2026-08-10 (`@cosyte/hl7` installs from npm) and the
+# VARIANT STAYS: the argument is about tracked binaries as a class, and the script header
+# says why the specimen's departure does not retire it.
 #
 # WHY A SEPARATE WORKFLOW, and not a job in ci.yml:
 #

@@ -26,10 +26,10 @@ deleting the pointer or the heading.** Why, and every disclosed miss:
 - **Phase 9 of 11.** Client / server / framing / connection / transport shipped; Phases 6, 7, 8
   and 9 are done. Next: `operations/roadmaps/mllp.md`. What each phase actually shipped:
   `documentation/agent-notes.md#shipped-phases-and-the-vendored-hl7-peer-tarball`
-- **For dev/test the `@cosyte/hl7` peer is a vendored packed tarball**
-  (`vendor/cosyte-hl7-0.0.0.tgz`, a devDependency), interim until the cross-repo consumption
-  decision lands. **Re-vendoring strips the `peerDependencies` entry: restore it afterwards.** The
-  exact command sequence: `documentation/agent-notes.md#shipped-phases-and-the-vendored-hl7-peer-tarball`
+- **The `@cosyte/hl7` peer installs from npm**, like every other `@cosyte/*` dep. The vendored
+  tarball is GONE: it pinned dev/test to a `0.0.0` snapshot and hid a FIXED MSA-2 correlation bug
+  for ten releases. **Never re-vendor one:**
+  `documentation/agent-notes.md#shipped-phases-and-the-vendored-hl7-peer-tarball`
 - **This package is on the npm registry, and this file names no version, deliberately** (derive it:
   `npm view @cosyte/mllp version`). Never quote a version here, never move a published version
   backwards, and never infer repo visibility from publish state or the reverse: they are
@@ -48,12 +48,12 @@ Full rationale, measurements and residuals: `documentation/agent-notes.md#the-em
   parentheses.
 - **Never add `grep -I`.** Measured on GNU grep 3.8: it skips a text file whose bad byte shares a
   line with the em dash, in total silence, and the gate prints OK.
-- **Never switch this copy to the text-only shape the other parsers run.** The vendored tarball is a
-  compressed stream that can carry `E2 80 94` by coincidence, and a red with no possible fix is a
-  gate someone disables.
+- **Never switch this copy to the text-only shape the other parsers run.** The tarball that forced
+  it is gone; the shape stays. A binary carrying `E2 80 94` by coincidence is a red with no possible
+  fix, which is a gate someone disables.
 - **The NUL exclusion is a disclosed miss, not a pass**, and the tell is the excluded count on the OK
-  line: **it reads 1 today**. If a NUL-bearing text fixture ever lands, revisit the partition, never
-  the ban. The at-risk fixture class already exists, so do not round this off to hypothetical.
+  line: **it reads 0 today**, 1 until the tarball left. If a NUL-bearing text fixture ever lands,
+  revisit the partition, never the ban. The at-risk class exists, so do not call this hypothetical.
 - **Never count over markdown alone.** A markdown-only count is what wrongly cleared `dicom`.
 - **The script is composed from three sibling copies; understand the composition before editing it**,
   and fix shared limits in the script header, not here.

@@ -20,14 +20,15 @@ thing a rule guards, or when you are tempted to relax one. Every paragraph here
   Mirth/NextGen (`test/differential/`, `MLLP_DIFF_ADAPTER`-gated live tier), the §3 quirk corpus
   (`test/conformance/`), and a PHI/observability audit that closed the `MLLP_FRAME_TOO_LARGE`
   `snippet` payload-slice leak) done. Next: see `operations/roadmaps/mllp.md` for what follows Phase 9.
-  For dev/test the
-  `@cosyte/hl7` peer is consumed as a **vendored packed
-  tarball** (`vendor/cosyte-hl7-0.0.0.tgz`, a devDependency): an interim mechanism until the
-  cross-repo consumption decision (umbrella `PW-5` gate) lands; refresh it by re-running
-  `pnpm -C ../hl7 build && pnpm -C ../hl7 pack --out ../mllp/vendor/cosyte-hl7-0.0.0.tgz`
-  (`--out` resolves relative to the `-C` directory) then `pnpm remove @cosyte/hl7 &&
-  pnpm add -D @cosyte/hl7@file:vendor/cosyte-hl7-0.0.0.tgz`. Note `pnpm remove` also
-  strips the `peerDependencies` entry; restore it (`"@cosyte/hl7": ">=0.0.0"`) after.
+- **The vendored tarball is GONE. The `@cosyte/hl7` peer installs from npm** at `^0.0.N` (ADR 0032,
+  which superseded the interim ADR 0008 five weeks after its own `PUB-FLIP` trigger fired).
+  **Why you never re-vendor one.** The tarball pinned dev/test to `0.0.0` while `@cosyte/hl7`
+  shipped ten releases. One made the MSA-2 echo byte-verbatim across the escape alphabet, naming
+  THIS package's correlation as the reason. The fix never arrived, so
+  `test/ack-from-hl7/control-id-verbatim.test.ts` asserted the DEFECT as a guarantee: `ID\X` echoed
+  back `ID\E\X`, a different control id on the wire, so an unmatched ACK, a resend, a duplicate
+  clinical message. **A pinned blob does not read as stale, it reads as the contract** -- which is
+  what makes it worse than an out-of-date range, since a range announces its age.
 
 ## The em-dash brand gate
 
@@ -53,11 +54,13 @@ thing a rule guards, or when you are tempted to relax one. Every paragraph here
   applies the `./` prefix in the list-building loop instead of through `sed -z`, so the scan is a
   single command with the stderr capture bound to all of it, which closes a residual the `ncpdp`
   and `dicom` copies still carry.
-  **Why not the text-only shape the other parsers run, since `vendor/cosyte-hl7-0.0.0.tgz` has no
-  em dash in it today:** the reason is durability, not a present-day red. That tarball is a
-  compressed stream, it is re-vendored by hand (see the Phase 9 note above), and a compressed
-  stream can contain `E2 80 94` by coincidence. Measured both ways against the real file: `dicom`'s
-  copy is green on today's bytes and **red, unremediably, on a copy seeded with those three bytes**.
+  **Why not the text-only shape the other parsers run:** durability, not a present-day red, and the
+  reason OUTLIVED the file that prompted it. The argument was about `vendor/cosyte-hl7-0.0.0.tgz`, a
+  hand-refreshed compressed stream that can contain `E2 80 94` by coincidence. Measured both ways
+  against the real file at the time: `dicom`'s copy was green on those bytes and **red, unremediably,
+  on a copy seeded with those three bytes**. **The tarball is gone and the shape stays**, because the
+  property belongs to binaries in general, not to that one file. Read the live excluded count off the
+  gate's OK line.
   You cannot rewrite a compressed byte stream with a period, and a red with no fix is a gate
   someone disables.
   **The disclosed cost, said plainly: a tracked TEXT file holding a NUL byte is silently exempt,

@@ -91,7 +91,7 @@
     "@arethetypeswrong/cli": "0.18.4",
     "@changesets/cli": "2.31.0",
     "@cosyte/eslint-config": "^0.0.3",
-    "@cosyte/hl7": "file:vendor/cosyte-hl7-0.0.0.tgz",
+    "@cosyte/hl7": "^0.0.10",
     "@cosyte/prettier-config": "^0.0.2",
     "@cosyte/test-utils": "^0.0.2",
     "@cosyte/tsconfig": "^0.0.2",

@@ -116,10 +116,11 @@
  *  (v)   [PINNED] A NUL-BEARING FILE IS SKIPPED WHOLE, SO ITS POINTERS ARE NEVER READ. This is
  *        the one that can print `all resolving` over a dangling pointer, and it is a disclosed
  *        miss rather than a pass. THE TELL IS THE SKIPPED COUNT ON THE OK LINE, exactly as
- *        `check-no-emdash.sh`'s NUL exclusion works and for the same reason: the one file it
- *        excludes today is the vendored `@cosyte/hl7` tarball, a compressed stream that cannot
- *        be read as markdown and cannot be edited to fix a red. If a NUL-bearing TEXT file ever
- *        lands here, revisit the partition rather than the rule. `CLAUDE.md` requires that
+ *        `check-no-emdash.sh`'s NUL exclusion works and for the same reason. It excludes NO file
+ *        today: the one it used to exclude was the vendored `@cosyte/hl7` tarball, a compressed
+ *        stream that could not be read as markdown and could not be edited to fix a red, and that
+ *        tarball is gone (the peer installs from npm). The rule stays. If a NUL-bearing TEXT file
+ *        ever lands here, revisit the partition rather than the rule. `CLAUDE.md` requires that
  *        exclusion be carried as a disclosed miss and says the at-risk class already exists, so
  *        do not round this off to hypothetical.
  *  (vi)  [PINNED] AN ATX HEADING INSIDE A FENCED CODE BLOCK IS NOT AN ANCHOR, and the fence

@@ -25,19 +25,21 @@
 #   * ZERO tracked files carry U+2014, as the literal character or in any encoded form.
 #     So this gate changed no content in this repo. It exists purely to stop a regression,
 #     which is the only reason to add a gate to a clean tree.
-#   * ONE file holds a NUL byte: `vendor/cosyte-hl7-0.0.0.tgz` (472,097 bytes, 1,774 NULs),
-#     a real gzip stream and the only binary this repo tracks. See BINARY EXCLUSION.
-#   * FOUR files are what GIT calls binary (`git ls-files --eol` reports `i/-text`), which is
+#     ZERO files hold a NUL byte. This read ONE until ADR 0032 deleted
+#     `vendor/cosyte-hl7-0.0.0.tgz`, a real gzip stream and the only binary this repo tracked.
+#     THE BINARY EXCLUSION STAYS; its section explains why the specimen leaving does not
+#     retire the shape. The live count is on the OK line, not here.
+#   * THREE files are what GIT calls binary (`git ls-files --eol` reports `i/-text`), which is
 #     a WIDER set than the NUL rule and is the one worth naming, because it is where the
-#     exclusion could start biting without anyone noticing. The other three are
+#     exclusion could start biting without anyone noticing. They are
 #     `test/differential/fixtures/r1-ack-aa.frame.bin`, `r1-adt-a01.frame.bin` and
 #     `r1-oru-r01.frame.bin`: git classifies them binary on its LONE-CR branch, since an
 #     HL7 v2 segment terminator is CR with no LF. They hold ZERO NUL bytes, so `has_nul`
 #     keeps them IN scope and they are scanned as text. Proved rather than assumed, the way
 #     ncpdp proved its `.xml` fixtures are read: each of the three was seeded with a live em
 #     dash in turn and the gate went RED on each, naming the file.
 #   * ZERO tracked gitlinks (no mode 160000 entry) and ZERO tracked symlinks.
-#   * Every tracked file except that tarball decodes as UTF-8.
+#   * Every tracked file decodes as UTF-8 (the one exception was that tarball).
 #
 # The fix is never to re-encode the character: rewrite the sentence with a
 # period, a colon, a comma, or parentheses.
@@ -62,19 +64,25 @@
 # stderr capture instead of being skipped. That is fail-closed and it is the better shape
 # where it applies.
 #
-# It does not apply here, and the reason is DURABILITY, not a present-day red. Measured:
-# `vendor/cosyte-hl7-0.0.0.tgz` does NOT currently contain the byte sequence `E2 80 94`,
-# so a knowledgebase-shape script would pass over it today, and dicom's copy was run
-# against this exact tree to confirm that rather than to assume it (it printed OK, exit 0).
-# But that tarball is a compressed byte stream, and a compressed stream contains any given
-# three-byte sequence by coincidence with real probability. The tarball is re-vendored by
-# hand whenever `@cosyte/hl7` is rebuilt (see this repo's CLAUDE.md), and the next re-vendor
-# could land a copy that happens to contain those bytes. Measured, not argued: appending
-# those three bytes to the real tarball and re-running dicom's copy makes it exit 1 with
-# `binary file matches` against `vendor/cosyte-hl7-0.0.0.tgz`. That red has NO REMEDIATION
-# AVAILABLE: you cannot rewrite a compressed byte stream with a period, and the tarball's
-# contents are not prose anyone wrote. A gate whose red state has no defined fix is a gate
-# someone disables, which is strictly worse than the hole it closed.
+# It does not apply here, and the reason is DURABILITY, not a present-day red. The argument was
+# made against a specimen that has since left the repo, and it OUTLIVES it; read to the end
+# before concluding the shape can now be relaxed.
+#
+# Measured at the time: `vendor/cosyte-hl7-0.0.0.tgz` did NOT contain the byte sequence
+# `E2 80 94`, so a knowledgebase-shape script would have passed over it, and dicom's copy was
+# run against that exact tree to confirm it rather than assume it (it printed OK, exit 0).
+# But that tarball was a compressed byte stream, and a compressed stream contains any given
+# three-byte sequence by coincidence with real probability. It was re-vendored by hand
+# whenever `@cosyte/hl7` was rebuilt, and the next re-vendor could have landed a copy holding
+# those bytes. Measured, not argued: appending those three bytes to the real tarball and
+# re-running dicom's copy made it exit 1 with `binary file matches` against the tarball. That
+# red has NO REMEDIATION AVAILABLE: you cannot rewrite a compressed byte stream with a period,
+# and the tarball's contents are not prose anyone wrote. A gate whose red state has no defined
+# fix is a gate someone disables, which is strictly worse than the hole it closed.
+#
+# THE TARBALL IS GONE (ADR 0032) AND THIS SHAPE STAYS. The property that forced it, an
+# unremediable red on bytes nobody authored, belongs to tracked binaries as a class, not to
+# that one file. Revisit the PARTITION if you must, never the ban.
 #
 # Do not read website's stated reason across to here. website regenerates 7 PNGs on every
 # PR, so it quoted a per-PR probability. mllp has exactly one binary, it is static, and it
@@ -105,13 +113,14 @@
 # THE COST, stated rather than left implicit, because it is a real hole in a ban whose own
 # wording has no exceptions: a tracked TEXT file that happens to hold a NUL byte (a UTF-16
 # document, a fixture carrying framing bytes) is excluded here and its em dash would be
-# missed. Say what that means for the tarball without softening it: seed the real
-# `vendor/cosyte-hl7-0.0.0.tgz` with a live em dash and THIS GATE PRINTS OK AND EXITS 0.
-# That is a MISS, not a pass. It is the deliberate price of the durability argument above,
-# and it was checked rather than reasoned, along with the control that proves the NUL rule
-# is what causes it: the same tarball bytes with every NUL replaced, carrying the same em
-# dash, goes RED. mllp has NO tracked text file with a NUL today, so the exclusion currently
-# exempts exactly one file and that file is a genuine binary.
+# missed. It was said for the tarball without softening, and the measurement stands as the
+# demonstration even though the file is gone: seeding the real `vendor/cosyte-hl7-0.0.0.tgz`
+# with a live em dash made THIS GATE PRINT OK AND EXIT 0. That is a MISS, not a pass. It is
+# the deliberate price of the durability argument above, and it was checked rather than
+# reasoned, along with the control that proves the NUL rule is what causes it: the same
+# tarball bytes with every NUL replaced, carrying the same em dash, went RED.
+# mllp has NO tracked file with a NUL today, so the exclusion currently exempts nothing.
+# Read the count off the OK line rather than from here.
 #
 # Do NOT round that off to "the hole is hypothetical here", which an earlier draft of this
 # header did. The at-risk fixture class ALREADY EXISTS: the three `.frame.bin` captures under

@@ -136,8 +136,9 @@
 #                          the audit log for fixture-level PHI-scan bypasses. Internal
 #                          compliance bookkeeping, not consumer documentation, and not in
 #                          package.json `files`.
-#   * CLAUDE.md, .github/, .changeset/, scripts/, test/, vendor/
-#                          internal by definition, or code rather than prose.
+#   * CLAUDE.md, .github/, .changeset/, scripts/, test/
+#                          internal by definition, or code rather than prose. (`vendor/` was
+#                          in this list until it was deleted with the hl7 tarball.)
 #   * src/ DOC COMMENTS    IN SCOPE, as a THIRD PASS at the bottom of this file, with its
 #                          own rule array (SRC_RULE_PATTERN), its own self-tests, and its
 #                          own extractor. `src/` JSDoc IS public: it is compiled into the
@@ -1047,8 +1048,8 @@ fi
 # heuristic calls binary, which includes a genuine TEXT file with a broken encoding, so a
 # violation inside one would be skipped in silence. This gate's surface is markdown, JSON and
 # TypeScript with no binaries (CHECKED ON THIS TREE: no tracked file under any scanned path
-# holds a NUL byte; the repo's one vendored tarball lives under `vendor/`, which this gate
-# does not scan), so losing `-I` makes a future binary a loud red instead of a silent miss.
+# holds a NUL byte, and since the hl7 tarball was deleted no tracked file anywhere in the repo
+# does), so losing `-I` makes a future binary a loud red instead of a silent miss.
 # Fail closed, not open. `-H` is set so every hit carries its filename: grep omits the name
 # when handed exactly one file, which an xargs batch boundary can produce.
 : > "$SCANLIST"