fix(design-drift): the member parser ate the comma, so every per-file check was blind

`pnpm design:drift` reported 18 fail on the merged tree. The real number is 98.
The difference was not progress — it was the checker failing to look.

parseMemberList matched each frontmatter field with `(\S+)`, which is greedy
across the comma separating them. `path: shell, prefix: …` therefore yielded the
path "shell,", and resolve(REPO_ROOT, "shell,", "src") does not exist. So
findMemberFiles returned an empty array for all nineteen members, and every
check that walks member files — F8 hardcoded hex and box-shadow, F4 raw z-index,
F1a Tier-1 consumption, leaked selectors — found nothing to inspect and passed.

The failure was well disguised. The one check that still fired was F6, whose
own path resolution broke the same way, so it failed for every member — and "no
DESIGN.md at member root" reads as Phase 8 work not yet done, not as an
instrument that cannot see the tree. A near-clean report with a plausible
explanation for its only failures.

This is the exact failure mode the Phase 1 handoff warned about in its own
notes: a checker reporting success because it failed to look. It warned about
two instances (an argv off-by-one and a CRLF read) and this is a third.

Each field now stops at the comma, and root_class is stripped of the quotes it
carries. With the parser fixed: 98 fail · 0 warn — 59 F8, 21 F4, 17 F6, 1 P2.

Not reconciled: the Phase 1 handoff reports 480 fail · 212 warn and 108/108
contrast pairs, against 98 · 0 and 30/30 here. The parser fix accounts for the
18 → 98 jump but not the rest, and the contrast pair count is a separate
question. Flagged on the PR rather than guessed at — those figures were measured
on another machine and their author should confirm what they counted.

Files changed:
- scripts/design-drift.mjs

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019a8tSPbFdvF1pKtADnWyDg

Author: mpstaton

scripts/design-drift.mjs

@@ -91,9 +91,27 @@ function parseMemberList(fm) {
   const members = [];
   for (const entry of list) {
     if (typeof entry === 'string') {
-      const parts = entry.match(/name:\s*(\S+).*?path:\s*(\S+).*?prefix:\s*(\S+).*?root_class:\s*(\S+)/);
+      // `(\S+)` was greedy across the comma that separates frontmatter fields,
+      // so `path: shell, prefix: …` yielded the path "shell," — and
+      // resolve(REPO_ROOT, "shell,", "src") does not exist. findMemberFiles()
+      // then returned [] for EVERY member, so every per-file check (F4 z-index,
+      // F8 hardcoded hex / box-shadow, F1a Tier-1 consumption, leaked
+      // selectors) silently found nothing and the run reported near-clean.
+      // The only surviving symptom was F6 failing for all 19 members, which
+      // reads as "per-member DESIGN.md files are Phase 8 work" rather than
+      // "the checker cannot see the tree".
+      //
+      // This is the same failure this script's own notes warn about: a checker
+      // reporting success because it failed to look. Stop each field at the
+      // comma, and strip the quotes root_class carries.
+      const parts = entry.match(/name:\s*([^,\s]+).*?path:\s*([^,\s]+).*?prefix:\s*([^,\s]+).*?root_class:\s*([^,\s]+)/);
       if (parts) {
-        members.push({ name: parts[1], path: parts[2], prefix: parts[3], rootClass: parts[4] });
+        members.push({
+          name: parts[1],
+          path: parts[2],
+          prefix: parts[3],
+          rootClass: parts[4].replace(/^["']|["']$/g, ''),
+        });
       }
     }
   }