Contract — R93 B1 outputSchema SDK Passthrough Envelope Mismatch

Round: R93 debug-sweep Branch: feature/r93-b1-output-schema-envelope Audit: docs/audits/r93-b1-output-schema-envelope-audit.md β task: 6b2da36a-8f1d-4a90-95c1-ac549b3fd60e

§1. Behavioural invariants (post-fix)

ID	Invariant	Verifier
I-1	A tool registered with outputSchema returns a {ok: true, data: <handler-result>} envelope through the MCP SDK without triggering -32602.	Regression test: register with_output_schema, call via Client, assert response.structuredContent.ok === true.
I-2	The eight live tools (router_score, router_fallback, router_stats, consensus_propose, consensus_vote, consensus_finality, consensus_gossip, vrf_eval) are reachable through the MCP wire and return their documented payloads.	Smoke probes against a freshly-bootstrapped server (manual verification + unit-level test asserting no -32602).
I-3	ColibriToolConfig.outputSchema remains an accepted field; passing it does not throw, and the existing accepts an optional description and outputSchema test continues to pass unchanged.	Existing test at src/__tests__/server.test.ts:441.
I-4	Handler return shapes are unchanged. The 8 affected handlers in src/domains/router/tools.ts + src/domains/consensus/tools.ts are not modified by this slice.	Source-level diff inspection.
I-5	The Zod outputSchema exports (RouterScoreOutputSchema, RouterFallbackOutputSchema, RouterStatsOutputSchema, ConsensusProposeOutputSchema, ConsensusVoteOutputSchema, ConsensusFinalityOutputSchema, ConsensusGossipOutputSchema, VrfEvalOutputSchema) remain in their source files and continue to flow into TypeScript types via z.infer<typeof X>.	rg "Output(Schema|Type)" src/domains/{router,consensus} returns unchanged sets.
I-6	The α envelope contract on failure is unchanged: handler throws → {ok:false, error:{code:"HANDLER_ERROR",…}}; Zod input failure → {ok:false, error:{code:"INVALID_PARAMS",…,details:{issues}}}.	Existing middleware tests.
I-7	The middleware contract documented in docs/2-plugin/middleware.md Stage 1-5 remains accurate. No new stage; no stage removed.	Manual cross-check; doc untouched.
I-8	The 3492 currently-green tests continue to pass after the patch.	npm test exit 0; suite count preserved.
I-9	npm run build && npm run lint && npm test is the gating triple.	CI + local verify run.

§2. Non-invariants (explicitly out of scope)

• N-1. Output runtime validation. Post-fix, outputSchema is a TypeScript-only artifact for the listed tools. Handlers may return any value; the wire layer does not enforce shape. A future slice (call it B1-extended) may layer Zod-driven output validation inside the middleware with a typed OUTPUT_VALIDATION_FAILED error code; that work is not in this slice.
• N-2. Per-tool output documentation updates. The slice doc + tool description strings already describe the response shape; no doc rewrite is required.
• N-3. Existing handler logic. Any latent bug in a handler’s return shape is preserved by this slice; the regression test only verifies the SDK doesn’t reject a well-shaped handler return — it does not verify handler correctness.

§3. Required test coverage

Test	File	Asserts
outputSchema-declared tool round-trips through the SDK without -32602	src/__tests__/server.test.ts (new it inside describe('5-stage middleware (InMemoryTransport end-to-end)'))	Registers a tool with both inputSchema and outputSchema, invokes it via Client.callTool, asserts response.structuredContent matches {ok:true, data:{<expected handler payload>}} and that the call does not throw.

This is the load-bearing regression test. It must FAIL on the pre-fix tree and PASS on the post-fix tree. The existing acceptance test at server.test.ts:441 is preserved unchanged.

§4. Acceptance criteria

• AC-1. git diff origin/main..HEAD -- src/server.ts shows a single-region change in registerColibriTool removing the conditional outputSchema spread on the SDK registerTool config object (audit §2.2).
• AC-2. git diff origin/main..HEAD -- src/__tests__/server.test.ts shows one new it inside the InMemoryTransport e2e block (audit §5).
• AC-3. npm run build exits 0 with no new warnings.
• AC-4. npm run lint exits 0 with no new warnings.
• AC-5. npm test exits 0; suite count = 79 (no removed/skipped suites); test count ≥ 3493 (the new regression test).
• AC-6. The PR body documents the bug + reproduction + fix + the live tool surface unblocked.
• AC-7. Writeback per CLAUDE.md §7: thought_record(reflection) followed by task_update(status="DONE").

§5. Risks

• R-1. A test elsewhere in the suite might depend on the SDK enforcing outputSchema and would silently pass when validation is gone. Mitigation: grep outputSchema in the test corpus and audit each hit.
• R-2. The SDK could in future enforce outputSchema purely on the typed return of registerTool’s handler arg without inspecting structuredContent. That’s a wider-API question and not part of this slice.
• R-3. A downstream consumer of the *.OutputSchema Zod exports might be using them for runtime validation (e.g. in a sub-agent dispatcher). Grep OutputSchema across src/** to confirm only TypeScript-type use, not parse() / safeParse() consumption.

Proceeding to packet.