fix(security): repoint custom-call-classifier completeness caveat off closed #741 (#764)

[graciangabriel8]

Closes #764

What changed and why

src/security/custom-call-classifier.ts carried two in-code pointers marking the send-family recipient gate's COMPLETENESS as an open question:

• The JSDoc above SEND_FAMILY_RECIPIENT_GATE (~:55): "...other third-party-source→recipient selectors may exist and are pending a qualified security review (issue security: ERC-777 operatorSend untracked in custom-call classifier — exfil sibling of #711 (SEC/ARCH weigh in on send-family set) #741)."
• The trailing note on assertSendFamilyRecipientIsWallet's non-bypassable refusal message (~:150): "...recipient-verified send path, which does not exist yet — tracked as issue feature: recipient-verified send path (ERC-721/1155 transfer, ERC-4626 receiver) — restores #753-retired capability behind durable recipient binding #756. (issue security: ERC-777 operatorSend untracked in custom-call classifier — exfil sibling of #711 (SEC/ARCH weigh in on send-family set) #741)"

#741 was closed by PR #753 on its own acceptance criterion (the nine gated selectors are correct — no wrong recipient index, no member that doesn't belong). SEC's certification on that PR explicitly left completeness unsettled; the caveat rode along with the closure anyway. A reader following (issue #741) lands on a resolved, closed issue and reasonably (and wrongly) concludes completeness was settled.

This PR repoints both references from #741 to #764 — the open issue that is itself the "Required end state (1): Immediate, scoped (DEV)" target this commit satisfies, and the precise open home for the completeness gap.

Scope is intentionally narrow, per ARCH's disposition on #776: "Ship NOW (trivial, DEV-ready — do not wait on the substrate design): the in-code caveat in custom-call-classifier.ts (~:55 and ~:150) cites CLOSED #741 as its completeness home — a dangling pointer. Repoint it to the OPEN substrate home (#776), or to #757... Mechanical, no design dependency."

What was deliberately left out

The substantive completeness fix — #757's structural recipient/beneficiary-parameter decoding on the signing path, needed to cover the 7 confirmed-unclassified selectors (Permit2 approve/transferFrom/batch transferFrom, unwrapWETH9, ERC-777 send, ERC-1363 transferAndCall, Yearn-style withdraw) — is not attempted here. Per both #764's own body and ARCH's #776 comment, hand-enumeration does not converge (unwrapWETH9 sits at the exact index the current model already supports and was still missed) and extending the selector list is explicitly the wrong fix. That work is folded into #776's design, gated on SEC/ARCH sign-off, and is not a DEV task.

The #764 body's item (3) — settling the U1 premise (selector signature-strings vs. published EIP standards) against the actual EIPs — is also out of scope here; it's listed as low-priority and separate from the repoint.

Test — falsifier

Added test/764-completeness-caveat-repoint.test.ts, following the same source-text-scan pattern as the existing test/presign-annotation-guard-735.test.ts (raw readFileSync + fileURLToPath(import.meta.url), no source import needed):

1. Asserts the classifier source no longer contains the literal (issue #741) substring anywhere — RED on unfixed main (contains it twice), GREEN on this branch.
2. Slices to the JSDoc block preceding SendFamilyGateEntry and asserts its qualified security review (issue #NNN) caveat does not resolve to 741.
3. Slices to the assertSendFamilyRecipientIsWallet function body and asserts its trailing (issue #NNN) caveat does not resolve to 741.

Tests 2 and 3 assert "not #741" rather than "must be #764" specifically, so a future re-scope (e.g. onto #757 once the seam ships) can repoint the source and stay green without touching this test — matching #764's own falsifiable acceptance criterion: "every in-code pointer to a completeness/security caveat resolves to an open issue."

No existing test asserts the literal (issue #741) trailing string (the existing send-family-gate tests in test/custom-call.test.ts match the refusal message via /CUSTOM_CALL_REFUSED[\s\S]*NOT[\s\S]*bypassable/, not the trailing issue pointer), so nothing needed updating.

Tests were NOT run locally — Node/npm are not installed on the authoring machine. CI (Build & Test on Node 20 + 22) is the authoritative check for this PR.

Blast radius

Two string literals changed inside comments/an error-message string in one file (src/security/custom-call-classifier.ts); one new, self-contained test file. No behavioral change — assertSendFamilyRecipientIsWallet's refusal logic, SEND_FAMILY_RECIPIENT_GATE contents, and every other classifier code path are untouched. The refusal message's content changes by two characters (issue number), which is cosmetic — no existing test asserts that trailing substring (see above).

Residual concerns

• Line 33's section header ("Issue #741 — send-family recipient gate...") was deliberately left unchanged — it documents the design issue security: ERC-777 operatorSend untracked in custom-call classifier — exfil sibling of #711 (SEC/ARCH weigh in on send-family set) #741 shipped under (historical provenance), not a completeness caveat, and is outside both security: exfil-surface completeness has no open home — in-code caveat cites CLOSED #741; 7 confirmed unclassified selectors; hand-enumeration does not converge #764's stated scope (:55 and :150 only) and ARCH's arch(substrate): hand-maintained fail-open enumerations are this codebase's default guard habit — 4 instances (#757/#764/#772/#771) #776 comment.
• This PR does not touch arch(substrate): hand-maintained fail-open enumerations are this codebase's default guard habit — 4 instances (#757/#764/#772/#771) #776 itself or INCIDENT: pre-sign block 5 checks selector but never arguments — recipient-bearing fns on RECOGNIZED destinations are signable drain paths (Aave/Morpho/Uniswap/Lido) #757; once the structural seam design lands, the completeness caveat here will likely need a further repoint (e.g. to INCIDENT: pre-sign block 5 checks selector but never arguments — recipient-bearing fns on RECOGNIZED destinations are signable drain paths (Aave/Morpho/Uniswap/Lido) #757) — the falsifier test is written to tolerate that without modification, but the actual source text will need a follow-up edit at that time.