README: rewrite for accuracy, honesty, and clarity (per 3-lens review)

Full rewrite addressing a fact-check + clarity + positioning review. The headline
numbers were all verified correct; the fixes are everything around them.

Accuracy / credibility:
- Tests section now documents the real prerequisites (clone TypeScript to /tmp/ts-repo;
  coverage needs VS Code's bundled grammar) and points at conformance-matrix.ts as THE
  bidirectional metric — not run-conformance.ts, which reports a raw 94.2% accept-rate
  the README used to contradict. Notes these suites are excluded from CI.
- Reframe the over-acceptance gap honestly: it's context-sensitive parse-phase
  constraints a CFG can't express (reserved-word placement, modifier soup, super
  type-args) — NOT 'semantic errors TS parses cleanly then rejects' (the matrix measures
  parseDiagnostics, i.e. the parse phase).
- 'Embedded languages' no longer sells the joint seam-test as built — the embed
  annotation emits standard TextMate contentName today; the verified-seam is the design goal.
- 'fall out for free' qualified: lexer/parser/TextMate/vscode-config/CST-types are
  functional; tree-sitter/Lezer/Monarch are scaffolds with marked incomplete sections.
- 'no manual scope assignment' qualified: structure is inferred, the ~45-line scopes
  vocabulary is declared. Adding-a-language notes the harness is TS-specific.

Clarity / honesty:
- State TS-only status + a quick start up top; define CST/bidirectional/over-acceptance/
  Pratt/TextMate-scopes on first use; cut the core claim's repetition.
- Source comparison is now precise (5x fewer LINES than the 3331-line official YAML);
  'Known differences' says 4 token instances across 3 patterns (was an apparent miscount).

Author: johnsoncodehk