docs: make CLAUDE.md an @AGENTS.md import instead of a symlink (#48)

Cursor globs for both CLAUDE.md and AGENTS.md and does not dedupe by
inode, so the symlink caused identical instructions to be injected
TWICE into every agent session.

A one-line `@AGENTS.md` import fixes it, because the tools treat
imports differently:

  - Claude Code expands `@AGENTS.md` and gets the full file
    (it does NOT read AGENTS.md on its own -- verified with a
    marker test plus positive control)
  - Cursor does NOT expand it: it injects 10 bytes plus one copy
    of AGENTS.md
  - Codex keeps reading AGENTS.md directly

Measured on neurobro-terminal-backend: the first request to the model
dropped from 263,425 to 190,461 tokens.



Claude-Session: https://claude.ai/code/session_013mWK5fgUhaMRXo1imDXurn

Co-authored-by: greg-agent[bot] <305014151+greg-agent[bot]@users.noreply.github.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>

Author: 3 people

CLAUDE.md

@@ -0,0 +1 @@
+@AGENTS.md