feat(db): normalization + shape CHECK constraints migration, seed hardening

.github/workflows/ci.yml

@@ -3,8 +3,9 @@ name: CI
 on:
   push:
     branches: ["main"]
+  # No base-branch filter: stacked sub-MVP PRs target feature branches and
+  # must pass CI independently (CLAUDE.md PR discipline).
   pull_request:
-    branches: ["main"]
 
 permissions:
   contents: read

ARCHITECTURE.md

@@ -225,7 +225,7 @@ interface CsiNode {
 
 interface CsiTree {
   id: string           // spec ID
-  section: string      // CSI section number, e.g. "27 21 00"
+  section: string      // CSI section number, e.g. "27 21 00", "26 00 13.10", "01 32 01.00 10"
   title: string
   parts: CsiNode[]     // root-level Part nodes
 }
@@ -262,7 +262,7 @@ interface ApiResponse<T> {
 -- Specs
 CREATE TABLE specs (
   id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
-  section VARCHAR(20),  -- "27 21 00"
+  section VARCHAR(20),  -- "27 21 00" | "26 00 13.10" | "01 32 01.00 10" (expanded shape, ADR-020)
   title TEXT,
   source VARCHAR(20),   -- 'ufgs' | 'arcat' | 'cpi' | 'unknown'
   created_at TIMESTAMPTZ DEFAULT now(),
@@ -344,7 +344,7 @@ CREATE TABLE spec_references (
   source_spec_id      UUID REFERENCES specs(id)      ON DELETE CASCADE,
   source_paragraph_id UUID REFERENCES paragraphs(id) ON DELETE CASCADE,
   target_type         VARCHAR(20) NOT NULL,  -- 'section' | 'paragraph' | 'standard'
-  target_spec_section VARCHAR(20),           -- "09 91 00" — for section refs
+  target_spec_section VARCHAR(20),           -- "09 91 00" / "26 00 13.10" — for section refs
   target_spec_id      UUID REFERENCES specs(id)      ON DELETE SET NULL,
   target_paragraph_id UUID REFERENCES paragraphs(id) ON DELETE SET NULL,
   standard_code       TEXT,                  -- "ASTM C150" — for standard refs

CLAUDE.md

@@ -249,6 +249,7 @@ docs/adr/
   017-project-manual-publishing.md
   018-document-concurrency-state-model.md
   019-scope-boundaries-content-neutral-platform.md
+  020-section-number-expanded-shape.md
 ```
 
 **ADR format:**

docs/adr/020-section-number-expanded-shape.md

@@ -0,0 +1,39 @@
+# ADR-020: Expanded Section-Number Shape as Opaque Normalized String
+
+## Status: Accepted
+
+## Context
+
+CSI MasterFormat Level 4 (`26 00 13.10`) and UFGS Level 5 agency suffixes
+(`01 32 01.00 10`; 10 = Army Corps, 20 = NAVFAC, 30/40 = NASA/AFCEC) appear in
+36% of the UFGS reference corpus and arrive through every ingest format (.SEC,
+DOCX, plaintext). SpecR previously validated only `NN NN NN`, silently
+truncating suffixes in prose-ref extraction and content inference — collapsing
+distinct sections (e.g. `01 33 23` vs `01 33 23.33`) into one identity.
+
+Two viable designs:
+1. Opaque normalized string, grammar owned by one module.
+2. Structured `SectionNumber` type with decomposed DB columns
+   (division/l2/l3/suffix/agency).
+
+## Decision
+
+Opaque normalized string (`src/lib/section-number.ts` owns the grammar).
+Canonical form: single ASCII spaces, `NN NN NN`, `NN NN NN.NN`, or
+`NN NN NN.NN NN`. Cross-reference linking remains **exact match only** — a ref
+to `26 00 13` never resolves to `26 00 13.10` or vice versa. DB CHECK
+constraints enforce shape on `specs.section` (plus the `'unknown'` inference
+sentinel) and `spec_sections.section_number`;
+`spec_references.target_spec_section` stays unconstrained because it records
+what the source document said.
+
+## Consequences
+
+- One module to change when the grammar grows; consumers embed its fragment.
+- Exact-match keeps broken refs honest (a base ref to a missing base section
+  is genuinely broken) at the cost of no family fallback.
+- Structured queries (e.g. "all agency variants of X") require LIKE prefixes
+  rather than column equality — acceptable; no current feature needs them.
+- Free-prose ambiguity: `Section 26 00 13.10 20 mm` mis-reads `20` as an
+  agency suffix. Documented as KNOWN AMBIGUITY; tagged .SEC refs are immune.
+- Lexicographic ORDER BY remains correct for the fixed-width grammar.