v2.0.0 — MIE v3 format redesign; retire the discovery trio (MAJOR)

The MIE-format redesign (v3). A ground-up rewrite of every MIE around verified, executable
worked examples — the load-bearing query-construction content the 2026-07 ablations isolated —
collapsing the v2 format's 4× restatement (schema list + ShEx shape + worked query + sample triple)
into one atomic unit. All 36 served MIEs were rebuilt; the corpus is 50.9% smaller
(1.57 MB → 770 KB). Validated as an equivalence gate over the full 100-question benchmark (×3, API
answering + judging): judge score flat within the declared ±0.5 margin (paired Δ +0.29/20, 95%
CI [−0.09, +0.68]), factoid correctness up (+1.0 on factoid questions), and measured runtime
−15% input tokens / −15% cost / −6% latency. Durable record:
benchmark/redesign/release/FINDINGS.md.

MAJOR because the discovery trio is removed from the tool surface.

Removed

• find_databases, list_databases, list_categories — the discovery trio. The full
  database= roster is already always on the run_sparql / get_MIE_file schema, and the
  descriptions/keywords/categories they served moved to a static, generated Database Catalog
  section of the Usage Guide. Removing them retires a whole class of "which tool tells me what
  exists" round-trips (agents can pick a database by reading the guide).

Added

• Database Catalog in the Usage Guide — a build-time section listing all 36 databases (title,
  one-line description, keywords) grouped by category, generated from the MIE discovery: blocks by
  scripts/generate_usage_guide_catalog.py (--check / --list-categories modes), with a
  test_catalog_in_sync drift guard and a CI job (catalog.yml).

Changed

• Served MIE corpus flipped to the v3 format (togo_mcp/data/mie/*) — same databases, new
  structure (discovery / header / examples / schema_delta / id_join_map), 29–65% smaller per
  file. No server-code change: the reader already handled both formats (discovery-or-schema_info,
  global_gotchas-or-critical_warnings).
• Workflow STEP 0 is now a no-tool catalog scan (was find_databases(...)) — updated in the
  TogoMCP_Usage_Guide docstring and guide parts 01/02/03.
• Empirical budgets + tool tiers refreshed from the v3 100-question run (optimal total tool
  calls 6–15 → 4–10; consecutive-SPARQL penalty 1.26 → ~1.1; find_databases removed from the tiers).

---

Full validation record: benchmark/redesign/release/FINDINGS.md · PR: #167 · Compare: v1.7.1...v2.0.0