Codify per-protocol vs. prepare_custom_call cutoff rule in CLAUDE.md

[szhygulin]

Adopted from #638 / #645. The research recommendation was (c) hybrid — keep per-protocol prepare_* for flows that encode prepare-time invariants, route the long tail through prepare_custom_call. The cutoff rule needs a written home so future protocol-add decisions don't re-litigate the same question.

Cutoff rule (verbatim from #645's research note)

Keep a per-protocol prepare_* when the tool encodes any of:

1. Slippage / MEV math (swap-class)
2. Protocol-pause / cap / threshold preconditions (lending-class)
3. Approve+action bundling with burn-address + unlimited-approval gates
4. Durable-binding to a verified candidate (Inv Add TRON Phase 1: read-only balances + portfolio integration #15 — validator, market, bank, comet, ATA)
5. Non-standard token semantics (rebasing, fee-on-transfer)

Otherwise route the agent to prepare_custom_call. The cutoff is structural, not popularity-based: a Uniswap V3 collect that doesn't take slippage args could legitimately go generic; a brand-new Layer-N farm that takes a min-out arg should not.

Why this is the right home

Project CLAUDE.md (/home/szhygulin/dev/recon-mcp/CLAUDE.md) already carries the project-specific gate rules (Pre-Sign Gate Surface Sweeps, Typed-Data Signing Discipline, Cross-Repo Scope Splits). The cutoff rule is the same shape: a structural decision that future contributors / agents need to apply when scoping new protocol coverage. ROADMAP.md tracks individual protocol adds; it isn't where structural meta-rules live.

Threat-model rationale (for the rule body)

assertTransactionSafe blocks 4 (catch-all unknown destination) and 5 (per-destination ABI-selector check) are bypassed for ack-stamped prepare_custom_call txs; only blocks 2 (approve spender allowlist) and 3 (transfer on unknown token) still fire. Per-protocol tools concentrate prepare-time invariants the generic path cannot enforce — picking prepare_custom_call for a new protocol that meets criteria 1–5 silently drops those invariants behind a single ack.

Scope

• Add a "Per-protocol vs. generic-call cutoff" section to project CLAUDE.md with the 5-criterion rule and the threat-model one-liner.
• No code change.

Related

• Research: generic tx-handling tool vs. per-protocol prepare_* tools for DeFi coverage #638 (parent research issue)
• docs(research): generic tx-handling tool vs per-protocol prepare_* (#638) #645 (research deliverable, closed unmerged — content moved here)
• feat(custom-call): add prepare_custom_call escape-hatch tool #494 / feat(security): add get_contract_abi — standalone Etherscan-V2 ABI fetch (closes #495) #497 / fix(custom-call): honor acknowledgeNonProtocolTarget at preview/send time #498 (prepare_custom_call and ack gate)
• feat(scope): conditional tool-surface gating via VAULTPILOT_CHAIN_FAMILIES + VAULTPILOT_PROTOCOLS #492 (VAULTPILOT_PROTOCOLS — the catalog-growth lever, complementary)