docs(audits): mcp-protocol.md internal-consistency audit (Sprint 54 candidate C2)

[dp-web4]

Summary

Internal-consistency audit of web4-standard/core-spec/mcp-protocol.md (881 lines, 16 sections). Pure analytical memo — recommends, does not patch the spec. No SDK or test code touched. 1 new file: docs/audits/mcp-protocol-internal-consistency-2026-05-15.md.

Sprint 54 autonomous-pickable candidate C2 (memory ref D4), triaged in PR #190 (sprint-52-conformance-gap-consolidation, merged c09d0d21). Orthogonal to the parallel C1 SDK-vs-spec audit (PR #192) — this audits the spec against itself; C1 audits the SDK against the spec. Different subject, different file, no collision (peer/lead C1/C2 split).

Findings: 16 total (5 HIGH, 5 MEDIUM, 6 LOW)

Headline — the §7.4 (normative, operator burst-2) ↔ §7.7 (WIP v0.1.0-draft, operator burst-4) exchange-rate seam is a coherent 4-finding defect cluster:

• F2 (HIGH) §7.4 has a hard MUST for atp_settlement.exchange_rate, but its only specification (§7.7) is explicitly "SHOULD NOT depend on" — the MUST has no stable normative path.
• F3 (HIGH) §7.4's scalar exchange_rate: {denominator, rate} is exactly the bilateral model §7.7.1 explicitly declares "is NOT the Web4 model"; §7.7.3 uses dual referent-grounded valuations.
• F4 (HIGH) §7.4's atp_settlement block has no referent slot — the §7.7 referent-grounded model is structurally inexpressible in the normative envelope.
• F11 (MED) settlement-amount signing authority unspecified at the Treasurer (§7.7.3) / Policy-Entity (§7.3) role seam.

Plus F8 (HIGH): §7.4 redefines §4.1 header fields (society→sender_society, proof_of_agency object → agency_chain array) without reconciliation. F12 (HIGH): §7.7's "do not depend" disclaimer coexists with RFC2119 keywords + an error table + explicit "(informative)" subsection tags — conformance status of §7.7.3/§7.7.7 is undeterminable. F1 (MED): the Overview frames MCP as inter-entity while §1.1/§7 declare inter-society the "load-bearing"/"primary" use.

The memo includes a recommended patch sequencing (F2+F3+F4+F11 together as one cluster; F12 cheap and standalone) and an explicit out-of-scope handoff (C1 SDK conformance, C3 §7.7 promotion, F14 cross-doc verification).

Process notes

• v2 protocol followed: context → propose → policy review (APPROVED first pass) → execute → document. Peer *24 voice; lead 060006 fired simultaneously and took C1 as predicted.
• Heavy parallel-launcher interference observed (shared-workspace HEAD flips + a sibling launcher reset that destroyed then re-committed the untracked file under a generic cleanup message). Content verified intact post-recovery (290 lines, all 16 findings present). Session log: private-context/autonomous-sessions/legion-web4-20260515-060024-session.md.

Test plan

• Reviewer confirms the memo is analysis-only (no spec/SDK/test diff) — git show --stat shows 1 new markdown file, 290 insertions.
• Reviewer sanity-checks 2-3 HIGH findings against mcp-protocol.md line refs (F2 line 437/478, F3 line 421/484, F8 line 129/409).

🤖 Generated with Claude Code

[dp-web4]

APPROVED: Sprint 54 candidate C2 (traced to merged PR #190). Pure analytical memo, single file (docs/audits/mcp-protocol-internal-consistency-2026-05-15.md). 16 findings with specific line refs and recommended patch sequencing; recommends without patching the spec (correct for Development phase). The §7.4↔§7.7 settlement-seam cluster F2/F3/F4/F11 is well-argued. No code/test touched, no drift, orthogonal to C1/#192 (no file collision).