Fix a self-contradiction, two stale numbers, and dead citations in the docs

[digitalmasterykit-rgb]

Summary

Full adversarial fact-check of every markdown file in the repo root against the actual
source code (not against the docs' own cross-references), ahead of sharing this repo cold
with prospective clients. Fixes:

• README.md said the architecture is "identical" to production while a later section in
  the same file says the five-source generalization is this repo's own — a direct
  self-contradiction. Reworded to be consistent with the correct account.
• The "thirteen of fifteen gates never ask a model" claim (README/ARCHITECTURE/LIMITATIONS)
  was stale — two gates were added by an earlier PR and this was never updated. Verified
  the real count by grepping every gate string: 22 pure-code, 2 model-verdict, 24 total.
  The grep command is now inline so the number is self-verifying instead of another fact
  that can silently drift.
• PROVIDERS.md's comparison table had two scenario counts that no longer matched a live
  demo run (checked by actually running both providers).
• ADAPTERS.md described the ClickUp status-fallback as checking a type field the code
  never reads (it matches on status name only), and separately contradicted itself about
  which Drive comment carried a reply.
• README.md and AGENTS.md cited "PRD §N" as if it were a followable reference; that
  document never ships in this repo, so it's a dead end for any outside reader. Removed
  the citation where the sentence didn't need it, and made AGENTS.md's quote self-contained
  where it did.
• The "~40ms" demo timing (repeated 3x in README) was stale against a live measurement.

Verification

• npx tsc --noEmit clean
• npm run lint clean
• npm test — 1000/1000 passing
• npm run demo and npm run demo -- --provider anthropic both run correctly
• Every fixed number/claim re-derived directly from live command output or grep, not
  assumed from the doc's prior text