mcptoon v0.7.15

What's new

• The registry record carries a title and a homepage. server.json had only the
  five keys it started with; the 2025-12-11 schema also defines title (display name)
  and websiteUrl. Both are now set and pinned by a test. Correction worth keeping:
  the schema has no license field — a teardown note that said to add one was
  reading a different registry's schema.
• The landing pages are discoverable by agents. docs/llms.txt is published, so a
  tool that follows the llms.txt convention can read the measured numbers and the
  canonical entry points without scraping HTML.
• Both READMEs carry the mcpservers.org badge. The directory approved the listing
  on 2026-09-15 and asks listed projects to show the badge.
• Both READMEs now lead with the two decisions a reader is making. The hero states
  the two pain points — tool schemas eating the context window, and the same MCP
  servers configured once per agent — before the badges, with the install line directly
  under them. The body below is unchanged in substance.

Fixed

• The demo no longer contradicts itself. One screen promised "Install 1,000 MCP
  tools, 0 token schemas" while the table under it measured 255 tools at 581 tokens.
  The banner carries the measured pair now, each tier keeps its own number, and a guard
  fails if the retired claim comes back.
• The suite total was stale on six surfaces. README.md, README.zh-CN.md,
  DEVELOPERS.md, ROADMAP.md and both landing pages said 925 passed after the
  suite had moved on. All six read the current total now; the landing pages were
  regenerated through the documented pipeline rather than hand-edited.
• The suite-total guards allow only what a skip can explain (one), down from a gap
  of five — the wide gap is what let the stale total above pass unnoticed.
• A version guard no longer breaks Python 3.10. The check added on 2026-09-17 read
  pyproject.toml with tomllib, which is 3.11+, so CI went red on the two 3.10 jobs
  the moment it landed. It parses with a regex now, as test_registry_sync.py already
  did for exactly this reason.

Footprint (re-measured, not remembered)

Full history: CHANGELOG.md