Normalize bounded Codex Responses requests into the Exchange IR

[tiammomo]

Summary

Implement the second policy-sized child of #39 after #40 merges: normalize the typed Codex request into ModelPort Exchange while preserving Responses-native, metadata-minimized evidence and keeping all runtime routes disabled.

Actual behavior

The parser slice can validate the pinned Codex request syntax, but ModelPort cannot yet represent that request in its provider-neutral Exchange IR without losing protocol/path identity or tool-call causality.

Expected behavior

A Responses-native ClientRequest normalizes the bounded text/function subset into Exchange with stable semantic evidence and strict causal validation, while provider rendering and /v1/responses remain fail closed.

Acceptance criteria

• Starts from main after Add the pinned Codex Responses parser and content-free fixture #40 merges; normalizes ordered text/function history and parallel calls; rejects duplicate, missing, or out-of-order call evidence; preserves openai-responses and /v1/responses identity; excludes client_metadata from fingerprints, estimates, parameters, forwarding, and ledger evidence; provider rendering and the HTTP route remain disabled; focused tests and boundary documentation pass.

Additional context

Context

This is the Exchange-normalization child of #39 and follows the parser/fixture slice in #40. It must start from main only after #40 merges; it is not a stacked pull request.

The parser establishes the bounded request syntax pinned to codex-cli 0.148.0-alpha.21. This issue establishes the semantic boundary needed before any experimental HTTP handler: ModelPort must retain Responses-native client evidence even if a later provider adapter renders the normalized request through another protocol.

Scope

• Add a ClientRequest variant for the typed bounded Codex Responses request.
• Normalize instructions, text messages, function calls, function-call outputs, function tools, tool choice, parallel-tool intent, reasoning effort, and typed response controls into the existing Exchange IR.
• Preserve item order. Group adjacent parallel function calls into one assistant turn and require every pending call to receive exactly one matching output before a new message or call batch begins.
• Reject duplicate/empty item IDs, duplicate/empty call IDs, missing outputs, out-of-order outputs, duplicate tools, and role/content mismatches.
• Preserve openai-responses, /v1/responses, route name, and a stable semantic fingerprint as client evidence.
• Parse and bound Codex client_metadata but exclude it from the semantic fingerprint, input estimate, normalized parameters, provider forwarding, and ledger evidence.
• Keep provider rendering fail closed for the new client variant until the response adapter slice is implemented.
• Add focused unit tests for ordering, parallel causality, evidence identity, metadata minimization, and render rejection.
• Document the internal Exchange behavior while stating that no Responses HTTP route or SSE renderer exists.

Non-goals

• Registering POST /v1/responses or changing Dashboard compatibility status.
• Rendering provider requests or translating provider output into Responses events.
• Supporting multimodal input, reasoning-history items, structured output, hosted tools, WebSocket, Realtime, background responses, or general Responses compatibility.
• Depending on an unmerged branch: implementation starts only after Add the pinned Codex Responses parser and content-free fixture #40 is on main.

Dependencies

• Blocked by Add the pinned Codex Responses parser and content-free fixture #40 / PR feat(protocol): add bounded Codex Responses parser #41.
• Parent contract issue: Define and test the bounded Codex Responses request contract #39.
• Parent capability issue: Add a Codex-bounded OpenAI Responses ingress behind an experimental gate #38.