The doc argued that the first real re-sync is what makes source_commit
truthful, but never said what the restored script actually writes: source is
the repo name (not a path, not a machine-local absolute path) and source_commit
is git rev-parse HEAD in the source checkout.
Naming the second half makes an ordering constraint visible that was implicit:
a re-sync run before this repo's commit reaches the public mirror stamps two
public consumer repos with a hash that exists on one machine. That is the
c1b7bad defect again with a fresher date, so the check to run first is written
out.
The script that produced the existing vendored copies was deleted with the
monorepo layout in 2026-06 and recovered from catalog's history
(e84dffd^:scripts/sync-design-system.mjs). It could not be restored unchanged:
it walks its source directory and copies everything, and this repo's root is no
longer a directory holding only the design system. Run verbatim it vendored 109
files instead of 27, STATE.md and the whole of .git/ among them - and STATE.md
is gitignored precisely because this repo's remote is public.
The boundary is an explicit allowlist (DELIVERED_FILES) rather than a dist/
directory: one definition, read directly by the tests, with no copy step to go
stale. The allowlist's own failure mode - a new design-system file forgotten
from the list ships nothing, and --check stays green because it hashes a
consumer's tree against that consumer's own MANIFEST - is closed by scanning
the source for unlisted root-level *.css, fonts/* and schemas/* and refusing to
run.
tests/sync-source-boundary.test.mjs pins both directions. It was red on the
verbatim script for the right reason (STATE.md present in the target, with
tokens.css as the known-positive control) before the boundary was written.
--target is the consumer's repo root; the script appends
playground/vendor/playground-design-system/ itself. --check verified read-only
against both consumers: MANIFEST OK (27 files) for ms-ai-architect and
llm-security alike. No re-sync was run, so both MANIFEST.json files are
untouched and the stale source_commit is still theirs to inherit at the first
real re-sync.
README and docs/vendoring-and-re-sync.md now publish the command, and warn that
a plain diff/cmp against this repo reports all nine stylesheets as drifted
because the sync injects a generated-header line the source does not have -
17/27 raw, 26/27 with line 1 stripped. Found by ms-ai-architect, re-measured
here.
Vendoring stands: 26/27 vendored files are byte-identical in both consumers 3.5 months on, with no local drift. No re-sync command is published yet - the script is recoverable in full from catalog history and is already polyrepo-ready, but restoring it unchanged would copy this repo's root wholesale, STATE.md included.
Also records that MANIFEST.json's source_commit c1b7bad is not merely unreachable but false: the vendored tree carries components-tier4-project-view.css, which did not exist in that commit. Verified with git cat-file -e plus a known-positive control on components-tier3.css in the same commit.
Verification was find/git cat-file/SHA-256 comparison, not grep - the full log is in docs/vendoring-and-re-sync.md.
