Registry-driven releases: one source of truth for tokens, components, and counts
editorial(design-system): fix registry-driven-releases' standing 2026-08-02 flag (Track-B) — confirmed against app-privategit-design/src/tokens_gallery.rs's own source comments: the flag's 'served exclusively from primitive.json+pointsav-brand.json, never tokens.full.json' claim was itself backwards — those files are compiled INPUTS to tokens.full.json, which IS the served/compiled export, matching this article's original mechanism claim; only the leaf count was genuinely stale, now described as current-count-not-fixed rather than re-asserting a specific number that will drift again; de-narrated per register-documentation.yaml; same fix applied to the ES pair; register-clean EN+ES
@@ -95,18 +95,15 @@ no longer maintained. Two details make this example useful. First, neither copy actually fed the site: the tokens design.pointsav.com serves resolve through one pipeline — the vault's source files compiled into a single resolved export (`dtcg-vault/exports/tokens.full.json`, 146 leaf tokens) — so the divergence was invisible in production while remaining a live hazard for any tool that picked the wrong file. (Correction, 2026-08-02: the pipeline claim is backwards — the real commit's own message states design.pointsav.com's served tokens come exclusively from `primitive.json` + `pointsav-brand.json`, never from either `dtcg-bundle.json` copy nor the vault export named here. The 146-token count is also stale by this article's own `last_edited` date — the real file has grown to ~385 leaves since, after two more pillars (`paper`, `writing`) were folded in. The "107 shared paths, 5 new colors" resolution numbers earlier in this section are independently confirmed accurate. Flagged, not resolved.) Second, the resolution was registry-shaped: declare the vault's source files (`primitive.json`, `pointsav-brand.json`, plus the `paper` and `writing` pillars) compiled into a single resolved export, `dtcg-vault/exports/tokens.full.json` — so the divergence was invisible in production while remaining a live hazard for any tool that picked the wrong file. That export has grown well past the two-pillar shape this section first measured; the current published bundle spans four pillars and several hundred leaves total, and the file itself is the current count, not a number fixed to this article's writing date. Second, the resolution was registry-shaped: declare one copy canonical, fold the five genuinely-new values into it, and replace the other with a deprecation stub that points at the canonical file. Not a cleanup of two copies into two better copies — the elimination of the second