feat(api): add bounded Responses compatibility (#153)

[tcballard]

Summary

Adds the bounded authenticated Responses API compatibility surface for #153.

Scope

Included

• /v1/responses and /responses on local and authenticated managed data-plane routers
• strict text and bounded multi-turn input normalization over the existing Chat Completions path
• explicit support for model, input, instructions, stream, max_output_tokens, and temperature
• normalized buffered response shape and input/output/total usage fields
• bounded Responses SSE events with one terminal response.completed or response.failed event
• explicit unsupported-field and unsupported-content errors rather than silent field loss
• input history, output, response-body, and upstream event bounds
• decision-only/dry-run preservation, route/auth/privacy/budget/accounting reuse, tests, ADR, and docs

Excluded

• changes to the deterministic scorer or provider credentials
• tools, background jobs, stored conversations, previous-response references, or multimodal parts
• a second provider execution path or a Chat Completions contract change
• ChatGPT consumer-account access through OpenAI Platform credentials

Product / Architecture Decisions

• Responses is an adapter boundary over the authoritative authenticated Chat Completions execution path, so routing, eligibility, privacy, budgets, failover, and accounting remain one implementation.
• The accepted field set is intentionally narrow and deny-unknown; unsupported features fail closed with wayfinder_router_unsupported_request.
• Text and message history are bounded before delivery; provider response bytes and streaming event/output growth are bounded during relay.
• Streaming emits response.created, ordered text deltas, and exactly one terminal event. Duplicate [DONE] markers and late upstream events cannot create a second terminal or second accounting operation.
• Provider usage is normalized only when present; missing usage is omitted rather than fabricated. Existing Chat Completions behavior remains unchanged.

User-Facing Contract

CLI / Human Output

No CLI command changes. OpenAI-shaped clients may POST to /v1/responses (or /responses) using the documented bounded text/multi-turn contract.

JSON Output

Buffered success returns object: response, status: completed, a message output item, and normalized usage when the provider supplies it. Unsupported fields return wayfinder_router_unsupported_request; existing route headers and decision-only responses remain available.

Exit Codes

No CLI exit-code changes.

Verification

Ran

• cargo fmt --all -- --check
• cargo clippy --workspace --all-targets --all-features --locked -- -D warnings
• cargo test --workspace --all-targets --locked (242 gateway tests plus all workspace crates and live Redis tests)
• cargo +1.85.0 check --manifest-path rust/Cargo.toml --locked -p wayfinder-gateway
• cargo +1.85.0 test --manifest-path rust/Cargo.toml --locked -p wayfinder-gateway responses_ -- --nocapture

Covered

• buffered aliases and normalized provider usage
• fragmented SSE, ordered text deltas, duplicate terminal markers, and one terminal event
• bounded multi-turn translation and oversized history rejection
• strict unknown/unsupported field behavior
• dry-run/decision-only preservation for buffered and streaming requests
• existing Chat Completions, Redis, parity, concurrency, and provider behavior through the full suite

Review Path

Stacked on #166 (bounded canary tripwires), targeting codex/enterprise-canary-tripwires. Review after #166.

Notes For Reviewer

This is compatibility, not a claim of full OpenAI Responses feature parity. Tools, background execution, stored conversations, previous-response references, and multimodal content are intentionally rejected until each has an explicit bounded contract. The existing Chat Completions route remains the sole execution and accounting path.