feat(core): allow explicitly authored empty risk lists

[rainypilgrimage]

What and why

Closes #164.

Core required every Capability to author risk, but Registry also rejected an explicit empty list. That forced adapters to publish a known-inaccurate label when the author had reviewed an operation and none of the current closed-set RiskLabels applied.

This change allows an explicitly authored risk: [] while keeping the field required. Missing or malformed metadata remains invalid, every non-empty member remains closed-set validated, and load() preserves the authored empty list unchanged.

The authoring guidance ships with the contract: contributors should raise a focused Core vocabulary issue when evidence identifies a missing reusable danger semantic, use [] only after review finds no applicable current category, and request maintainer review when uncertain. Receipt evidence remains authoritative and can refute the declaration.

Type of change

• Protocol / Capability / Query
• Core / simulator / MCP server
• Bug fix
• Documentation / example
• Tooling / dependency

Framework and package impact

Registry now accepts an explicitly authored empty risk array while preserving required-field, array-shape, sparse-array, and unknown-label validation. The public RiskLabel[] shape does not change, and load() returns [] exactly as authored.

No new RiskLabel value is added, and no adapter, Simulator, MCP transport, or Monad Runtime behavior is changed. Downstream adoption remains with #158, #187, and the separately coordinated aPriori follow-up.

Verification

• pnpm build
• pnpm typecheck
• pnpm lint
• pnpm test
• User-facing package changes include a changeset
• Docs and examples match the implemented API

pnpm test currently reaches one unrelated Aave mainnet failure: PT_AUSD_8OCT2026 cleared the address-book age guard; vendor it and remove this quarantine. This branch does not modify Aave. The full offline-equivalent workspace run (MOSS_SKIP_E2E=1 pnpm test --reporter=dot) passes.

Protocol changes

N/A — this is a focused Core contract change and does not modify a Protocol adapter.

• Parameters separate reusable Zod value types from field-purpose descriptions
• Every Capability owns one direct TransactionNode and one typed Receipt
• Receipt tests preserve every original Change object in exact length and order
• Positive and @ts-expect-error fixtures cover exported type behavior
• Fixed addresses and ABIs include sources and verification
• A live Monad happy path returns zero Warnings

Evidence

• Core: 8 test files and 73 tests pass.
• Compile-time fixtures accept risk: [], reject a missing risk field, and preserve the closed RiskLabel set.
• Runtime coverage accepts [] and rejects missing, non-array, sparse-array, and unknown-label metadata.
• Registry load() projects the authored empty list unchanged.
• Receipt coverage retains the original inbound and outbound nativeTransfer Changes, verifies that the inbound case has no account outflow, and demonstrates that outbound evidence refutes the no-outbound claim.
• CONTEXT.md, MCP documentation, the protocol-onboarding guide and issue template, and an Agent-facing Core patch changeset carry the approved semantics and authoring rule.

[pillowtalk-Qy]

Hi Box and @rainypilgrimage,

I independently reviewed a2d25655a0ee99a32d6f28780291257390f302df against main@19e602e and the scope accepted in #164. No blocking finding in this change.

The implementation distinguishes an explicit empty array from missing/non-array metadata; the per-member validation also rejects sparse entries and unknown labels. The public field remains required, the type fixtures retain the closed vocabulary, and Registry load() preserves []. The glossary, MCP documentation, onboarding guide/template and Core changeset consistently describe an authored classification, not a guarantee of safety.

The Receipt regression retains the original inbound/outbound Changes and demonstrates that contrary evidence refutes the declaration. It does not introduce a new runtime risk detector or change downstream adapters, which matches the agreed scope.

Independent verification on this exact head:

• frozen install; lint; full workspace build and typecheck; git diff --check: pass;
• pnpm test:offline: 685 passed / 42 skipped;
• Core: 73/73; MCP: 22/22;
• pnpm -r --no-bail test: completed, with 18 workspace commands passing and 2 failing. Aave is 35/36 due to the expired quarantine in #185; Kuru is 85/86 due to the native-swap ROUTE_QUOTE_UNAVAILABLE reported in #194. Both also reproduced on unchanged main. Pendle is 155/155.

This approval is for the reviewed Core change; it is not a claim that the full live gate is green or a replacement for maintainer merge/CI decisions. The existing Aave/Kuru failures remain separate follow-ups. No branch changes were made.

[nishuzumi]

@rainypilgrimage @pillowtalk-Qy — I audited exact head a2d25655 against main@19e602eb and the #164 decision. The Core change is correct and complete: Registry.use() accepts an authored risk: [], still rejects missing, non-array and unknown-label metadata, and now also rejects sparse arrays the old .some() check let through; load() projects [] unchanged and the MCP load shape is untouched. Compile-time fixtures are load-bearing in both directions, and all three constraints from #164 are met in substance.

Rather than another review round, I pushed one follow-up commit 0e9eb5b onto this branch (docs only, no code):

• ADR 0003 gains a dated section "An authored empty risk list (2026-09-02)" recording the decision, the rejected alternatives (fundIn, a sentinel, an optional field), the authoring rule, and the Receipt-refutation requirement. The decision now lives with the rest of the closed-set taxonomy rather than only in the glossary.
• AGENTS.md review rule: reject knowingly inaccurate placeholder labels, and reject risk: [] without Receipt test coverage asserting that no Change moves assets out of the account.
• Issue template: the paragraph becomes an HTML comment under the "Risk labels" column, matching every other guidance line; as body text it would have been copied verbatim into every filed onboarding issue.
• Changeset: @themoss/core bumped as minor, consistent with the prior feat(core) vocabulary changes (debt, perps); this is an Agent-facing contract change and belongs under "Minor Changes".

Independent verification on a2d25655: frozen install, lint, build, typecheck and the full offline suite pass (685 passed / 42 skipped; Core 73/73, MCP 22/22). Mutation checks confirm the new tests fail when the registry change is reverted, when for...of is swapped back to .some(), or when the Array.isArray guard is removed, and that the @ts-expect-error fixture fails both without the directive and with risk: [] supplied. Full live run with --no-bail: 18/20 workspace commands pass; Aave 35/36 (#185 expired quarantine, a wall-clock assertion) and Kuru 85/86 (#194) fail identically on unchanged main and are unrelated to this diff. The red ci check is that Aave failure stopping the recursive run early.

For the downstream owners: packages/protocols/apriori/README.md:78-80 and src/adapter.ts:190-192 still teach the superseded "Registry rejects an empty risk list" rule and must be fixed in the aPriori adoption follow-up. Adapters adopting [] should note that the fixture helper in this PR only inspects nativeTransfer; a no-outbound Receipt assertion for an ERC-20-paying claim (#187) must also cover event Transfer Changes.

Merging once the offline checks report on 0e9eb5b.