Bring all adapters to token-safety parity via packaged scripts

[uncoooloj]

Depends on #28 (packaged script paths).

• codex.md: replaces four long inline jq blocks with one-line packaged-script invocations (session-candidates.py / session-events.py / session-digest.py incl. sidecar cache); keeps exactly one short jq example (the session_meta cwd check), the Candidate Ranking section, and the event-shape table. Net size down.
• claude-code.md: adds the raw-grep trap (grepping ~/.claude/projects/ for a topic/skill name mass-matches the available-skills boilerplate embedded in every transcript's system-reminder blocks — project user messages first, then rg the projection); documents the real per-session sidecar layout (<sessionId>/subagents/agent-*.jsonl, <sessionId>/tool-results/*.txt) so evidence reads come from cheap targeted files; replaces the longest jq blocks with script calls.
• cursor.md / antigravity.md / opencode.md: each gains a compact Safe Reading section (size check before read, metadata peek before bodies, strongest-signal candidate ranking, bounded rg/sed slices; packaged scripts for JSONL).

Documented script flags verified against the actual fixtures (e.g. claude-code events are tagged text/user, docs corrected to match).

🤖 Generated with Claude Code