Three-stage scan pipeline + schema simplification

.changeset/add-bucket-schema-and-merge.md

@@ -0,0 +1,27 @@
+---
+"ghost-expression": minor
+---
+
+Land the three-stage scan pipeline: topology (`map.md`) → objective (`bucket.json`) → subjective (`expression.md`). All three stages are now owned by `ghost-expression`; the previously separate `ghost-map` package is folded in.
+
+**New artifact: `ghost.bucket/v1`** — catalogues every concrete design value with structured specs, occurrence counts, and deterministic content-hashed IDs.
+
+**New verbs:**
+- `ghost-expression inventory [path]` — emit deterministic raw repo signals as JSON (manifests, language histogram, registry, top-level tree, git remote). Feeds the topology recipe. (Migrated from `ghost-map inventory`.)
+- `ghost-expression bucket <op>` — `merge` (concat with id-based dedup, idempotent — useful for modular rollups and fleet cohort views), `fix-ids` (recompute every row's `id` from content, so surveyor agents can author rows with empty `id` fields and finalize in one pass).
+- `ghost-expression scan-status [dir]` — report which scan stages have produced artifacts (`map.md`, `bucket.json`, `expression.md`) and which stage to run next. The build-system glue orchestrators call between stages.
+
+**Updated verbs:**
+- `ghost-expression lint` now auto-detects file kind by extension/content and dispatches to the right validator (`expression.md`, `map.md`, or `bucket.json`).
+
+**New skill recipes:**
+- `map.md` — author `map.md` from a target (the topology stage). Migrated from the standalone `ghost-map` package.
+- `survey.md` — author `bucket.json` from a target (the objective stage). Walks the agent through LLM-driven extraction with dialect-specific grep strategies, exhaustiveness discipline, and saturation predicate.
+- `scan.md` — meta-recipe that orchestrates topology → survey → profile end-to-end via `scan-status` checkpoints. Use when the user wants a full scan rather than a specific stage.
+
+**Refactored skill recipe:**
+- `profile.md` — now strictly the subjective interpretation stage. Reads `bucket.json` as ground truth; cannot fabricate values not in the bucket; cites bucket rows as evidence. Pre-requires `map.md` + `bucket.json`. Hard split from the previous one-pass extract+interpret recipe.
+
+**Removed:** the `ghost-map` package is deleted. `ghost.map/v1` schema and types now live in `@ghost/core`; `inventory` and `lint` (for `map.md`) move to `ghost-expression`. Consumers that imported from `ghost-map` should switch to `@ghost/core` (schemas/types) or `ghost-expression` (CLI verbs / library functions).
+
+Bucket schema, deterministic-id generation, lint, merge, and fix-ids primitives live in `@ghost/core`.

.changeset/canonical-decision-vocabulary.md

@@ -0,0 +1,5 @@
+---
+"ghost-expression": minor
+---
+
+Add a controlled vocabulary of 12 canonical decision dimensions (`color-strategy`, `surface-hierarchy`, `shape-language`, `typography-voice`, `spatial-system`, `density`, `motion`, `elevation`, `theming-architecture`, `interactive-patterns`, `token-architecture`, `font-sourcing`) so fleet-aggregation primitives can group decisions across members. Profile recipe nudges authors toward canonical slugs; novel project-flavored slugs may pair with an optional `dimension_kind` that maps to a canonical bucket. New soft `non-canonical-dimension` lint warning suggests the closest canonical match. The schema accepts the optional `dimension_kind` field on `decisions[]`; existing expressions remain valid.

.changeset/config.json

@@ -7,11 +7,5 @@
   "access": "public",
   "baseBranch": "main",
   "updateInternalDependencies": "patch",
-  "ignore": [
-    "ghost-ui",
-    "ghost-docs",
-    "@ghost/core",
-    "ghost-fleet",
-    "ghost-map"
-  ]
+  "ignore": ["ghost-ui", "ghost-docs", "@ghost/core", "ghost-fleet"]
 }

.changeset/drop-roles.md

@@ -0,0 +1,6 @@
+---
+"ghost-expression": major
+"ghost-drift": patch
+---
+
+Drop `roles[]` from the `expression.md` schema. Slot → token bindings either fall out of decisions[] (pattern consequences) or live in bucket.json components[] (exhaustive catalog). The hybrid `roles[]` slot was filling neither role cleanly and didn't scale to systems with many components. Existing files that carry `roles:` will fail strict lint — drop the section to migrate. Drift skill recipes that referenced `roles[]` as part of the expression frontmatter have been updated.

.changeset/drop-signature-section.md

@@ -0,0 +1,9 @@
+---
+"ghost-expression": major
+---
+
+Drop the `# Signature` body section and `observation.distinctiveTraits` field from the `expression.md` schema. Every claim Signature carried — distinctive traits, load-bearing absences — was already covered better elsewhere: `rules[]` with `presence_floor` codifies absences as enforceable patterns, decisions cite evidence directly, and `observation.personality` carries the lightweight summarizing role. Signature had collapsed into restating frontmatter as bullets.
+
+The parser silently ignores legacy `# Signature` blocks in older `expression.md` files, so existing artifacts continue to load without erroring. The writer no longer emits the section, and lint no longer expects it. `INVARIANTS.md` §3 has been amended to remove "Signature" from the body partition list.
+
+Removed: `DesignObservation.distinctiveTraits` (TypeScript), `# Signature` body section, `observation.distinctiveTraits` field. Context-bundle emit now uses `observation.personality` for the trait-phrase hooks in README and skill description.

.changeset/fix-oklch-backfill-and-tighten-survey.md

@@ -0,0 +1,14 @@
+---
+"ghost-expression": patch
+"ghost-drift": patch
+---
+
+Fix self-distance bug + tighten the survey recipe's exhaustiveness rule.
+
+**Bug fix.** `loadExpression` now backfills `oklch` on palette colors that arrive hex-only (frontmatter without an explicit `oklch` tuple). Without this, `comparePalette` treated hex-only colors as fully unmatched and contributed distance `1.0` per color — even when comparing an expression to itself. Self-distance was reported as 17.5% on a freshly authored expression. Backfill is deterministic (same hex → same oklch), so re-parsing the same file always yields the same in-memory shape.
+
+**Defensive fallback.** `comparePalette` also now resolves oklch on-the-fly when missing, and falls back to hex-string equality when even on-the-fly compute can't parse the color (CSS variables, opaque external refs). This covers third-party producers that don't backfill on write.
+
+**Recipe tightening.** `survey.md` now states the exhaustiveness rule up front and applies it per section. The rule is repo-agnostic — the recipe doesn't name specific signal sources (no "use registry.json"); the agent identifies the canonical signal in *this* repo, enumerates exhaustively, and cross-checks counts. New `Never sample` rule and explicit guidance against placeholder/glob names. Triggered by a dogfood scan that produced ~10% recall on `components[]` (6 rows for a 97-component package).
+
+**Schema cut: `bucket.libraries[]` removed.** External libraries (icon sets, primitive collections, motion libs, charting) no longer have a top-level bucket section. Whether a system uses Radix or hand-rolls primitives doesn't change what its design language *is*. When a library is load-bearing for the design language (icon family, font sourcing), it surfaces as prose evidence in the interpreter stage under the relevant decision dimension. Bucket sections are now `values`, `tokens`, `components`. `LibraryRow` / `LibraryRowSchema` / `libraryRowId` removed from `@ghost/core` exports. Existing `bucket.json` files with a `libraries` field will fail lint (use a no-op migration: drop the field).

.changeset/v0-rules-and-perceptual-prior.md

@@ -0,0 +1,5 @@
+---
+"ghost-expression": minor
+---
+
+Add `rules[]` as the v0 authoring surface for drift review and a perceptual-weight prior that calibrates emitted severity. An `expression.md` can now declare grep-friendly rules in frontmatter (id, canonical, kind, pattern, optional severity / match / tolerance / presence_floor / support); the emitter renders them as a Rams-shaped slash command grouped by Critical / Serious / Nit, with rationale + match shape + support cited per rule and a calibration footer that explains why severities landed where they did. Severity is computed from a fixed perceptual tier (color and font-family are loud, shape and elevation structural, spacing and motion-detail rhythmic) plus presence-floor escalation against the bucket — adding to a silent dimension is louder than tweaking a populated one. Existing `decisions[]`-driven expressions continue to emit the previous structured-section layout; the new rules-driven path activates only when `rules[]` is present. Ships a one-liner curl installer at `install/install.sh` so the skill bundle can be installed without npm or a build step; recipes carry prose fallbacks for the most-called CLI verbs (`scan-status`, `inventory`, `lint`) so the no-CLI path is real, not degraded.

.gitignore

@@ -16,3 +16,6 @@ packages/ghost-ui/.ghost/
 
 # Emitted skill bundle — re-generate with `ghost emit skill`
 .claude/skills/
+
+# Scratch dir for dogfood scans and ad-hoc experiments
+.scratch/