refactor: use dbterd 1.28 built-in json ERD target

[il-dat]

What & why

Migrate ERD extraction from our custom @register_target("json") adapter to dbterd 1.28's built-in, schema-validated json target. dbterd now owns the {nodes, edges, metadata} contract, so the locally-registered target (and its module) are no longer needed. build_erd(target="json") forces the official target and build_erd_data maps its payload into the SPA's {nodes, edges} shape, patching two known dbterd quirks: short-name edge ids under entity_name_format (resolved via a name_to_id map) and missing is_foreign_key flags on some algos like model_contract (back-filled from each edge's from_columns). The dbterd floor is bumped to >=1.28 (the built-in json target floor), and dbdocs/extract/erd_json.py plus its unit test are deleted.

Type of change

• 🐞 Bug fix (non-breaking change that fixes an issue)
• ✨ Feature (non-breaking change that adds capability)
• 💥 Breaking change (existing behavior changes)
• 🧹 Refactor / internal (no user-facing change)
• 📖 Docs only

Area

• CLI (generate / serve / deploy)
• extract/ (nodes / erd / graph / column_lineage / health)
• site/ (data dict / builder / 3-tier SPA / deploy)
• frontend/ (React Flow graph bundle)
• Config (dbdocs.yml) / packaging / CI / docs

How to test

task test           # 331 tests, 100% coverage — covers build_erd_data, _resolve_edge_id, _backfill_fk_flags
task frontend:e2e   # verifies the ERD renders against a real demo build (the dbterd-quirk regression gate)
task generate && task serve   # open the site, click a model, check the ERD tab renders nodes + FK edges

Checklist

• task lint passes (ruff format --check + ruff check). — clean, 60 files formatted, all checks passed.
• task test passes at 100% coverage (the gate is enforced in CI). — 331 passed, TOTAL 100.00%.
• I followed the load-bearing patterns in .claude/design_patterns.md (extended an existing seam, didn't fork one). The "ERD from dbterd's built-in json target" pattern entry + TOC are updated in this PR to match the new implementation.
• Data-dict / SPA changes keep producer ↔ consumer in sync — the {nodes, edges} shape the builder emits is unchanged; the graph bundle consumes the same keys. Payload is not re-inlined.
• Graph-UI changes — dbdocs/site/bundle/assets/graph/index.js carries a small aligned tweak to erd_algo / is_foreign_key handling; frontend/src/** is unchanged so no full bundle rebuild was needed.
• New files under dbdocs/site/bundle/** covered by the pyproject.toml artifacts glob. — n/a, no new bundle files (a module was deleted).
• Docs / dbdocs.yml.example updated for any new config knob or CLI flag. — n/a, no new config knob or CLI flag; .claude/skills/spa-site/SKILL.md and the design-pattern doc were updated to reflect the dbterd 1.28 floor.

Screenshots / notes

• dbterd dependency floor moves >=1.27 → >=1.28; uv.lock updated to match.
• The two patched dbterd quirks (short-name edge ids, missing FK flags) are the regression-prone seams — re-run task frontend:e2e on future dbterd bumps to confirm they still hold.