gsd-test found three real regressions in the migrated STATE.md checks:
- W002 didn't exempt phase refs whose only directory lives in an
archived milestone (#3652) — now consults planning-snapshot.cts's
archivedPhaseTokens field (added alongside this fix, shared with
W006's identical need).
- W011 (STATE/ROADMAP cross-validation) never fired: currentPhaseLabel
only read the current template's bare "Phase:" field, silently
missing legacy STATE.md fixtures that use the older bold
"**Current Phase:**" field (mirrors state.cts's own resolveStatePhase
fallback ladder, which the migration didn't carry over).
- W026 (STATE milestone-complete vs. unstarted ROADMAP phases) had two
independent defects: roadmapDeclaredPhases's milestone attribution
can't see <details>/<summary>-shaped ROADMAP sections, and current-
milestone resolution could go null — both silently emptied the
"unstarted" set every time. Fixed by scoping ROADMAP.md to the
current milestone via the same <details>-tolerant extractCurrentMilestone
every other milestone-aware consumer uses, in a new dedicated
planning-snapshot.cts field (currentMilestoneRoadmapPhaseIds) rather
than reusing roadmapDeclaredPhases, which exists for a narrower,
<details>-blind derivation (W021's own original logic) and would
have regressed it if repurposed.
Also fixes an unrelated drift-guard violation this same rule file
introduced: its own phase-token regex was independently re-derived
instead of built from the canonical PHASE_NUMBER_TOKEN_SOURCE.
a0c82f2bd8