Add rugcheck.xyz holder-quality screening filters

[agushudono]

What

Wires four previously-inert user-config.json keys to real screening behavior, backed by a new rugcheck.xyz enrichment:

Key	Effect
maxSniperPct	Drop candidates whose sniper insider-network concentration exceeds the threshold
maxSuspiciousPct	Drop candidates whose flagged-insider holder concentration exceeds the threshold
maxBundlePct	Drop candidates whose bundler insider-network concentration exceeds the threshold
convictionScoreMin	Drop candidates below a rugcheck-derived conviction score (100 − normalized risk)

All four default to null (disabled), so behavior is unchanged unless an operator opts in.

How

• tools/rugcheck.js — new internal screening helper getRugcheckSignals({mint}). Fetches the rugcheck report once per mint (15-min cache, 10s timeout), derives conviction_score, suspicious_pct, sniper_pct, bundle_pct, top10_pct. Returns null for any metric rugcheck doesn't supply.
• index.js — the screening recon loop fetches rugcheck only when at least one threshold is set, applies the filters alongside the existing maxBotHoldersPct check, and surfaces the signals in the candidate block shown to the LLM. Each filter fails open on missing data (same pattern as maxBotHoldersPct).
• Keys added to config.js, reloadScreeningThresholds, update_config's CONFIG_MAP, and user-config.example.json; documented in CLAUDE.md.

Notes / caveats

• sniper_pct and bundle_pct depend on rugcheck's insiderNetworks[] carrying typed entries with token amounts, which is often absent/partial — in that case those two filters fail open (no-op) for that token. convictionScoreMin and maxSuspiciousPct are the most reliably populated.
• Not an LLM tool — purely an internal pre-LLM enrichment, so no changes to tools/definitions.js / agent.js tool sets.

Testing

• node --check on all touched JS.
• Unit-verified the derivation math against a sample rugcheck payload: conviction 100−18=82, suspicious 8+5=13%, sniper 5%, bundle 3%, top10 41.5%; 404/error path returns all-null (fail-open).

🤖 Generated with Claude Code