refactor(responses): migrate the single-pass example chains to one parse

[cdoern]

Progresses #741. Independent of #1374 — they touch disjoint files and neither needs the other.

Change

Swap openai_responses_format for openai_responses_request in the nine example chains that classify only Responses traffic, so each create body is parsed once rather than twice.

Migrated: response-store, response-store-postgres-mtls, responses-routing, vllm-agentic-api, doc-extract, file-resolve, model-rewrite, http-passthrough, mcp-outbound-chain.

One chain needed more than a rename

responses-routing branches on the classifier's published results. A branch condition must name the filter it is attached to, since every other filter's results are cleared before it is evaluated, so renaming the filter alone left a condition that could never match.

Configuration validation rejects this at load with an explicit message rather than failing silently at runtime, which is how it surfaced:

branch 'stateful': on_result.filter 'openai_responses_format' must name the filter
the branch is attached to ('openai_responses_request')

Worth knowing for the chains still to migrate: a rename is not always sufficient, and the failure is loud rather than silent.

Not migrated, and why

Nineteen example chains still run the previous pair. They are not mechanical swaps:

• Three use the classifier for non-Responses traffic — unified-gateway, format-routing, and time-to-first-token classify Anthropic and Chat Completions bodies. openai_responses_request recognizes Responses operations only, so these need the Chat Completions registry work in feat(chat-completions): register operations for request-head classification #1048 rather than a swap.
• The translation chains rewrite the path away from /v1/responses, so head-based matching no longer recognizes the operation when the chain re-enters. That is a design problem, not a configuration one.
• The agentic-loop chains follow the translation work.

Verification

Full integration suite: 754 passed. The 26 failures are all examples::openai_agentic_loop, which fail identically on this machine without these changes and pass in CI; none are attributable to this change.

Unit tests 4513 pass, 0 fail. make lint passes, including the lean build and generated-doc sync checks.

Checklist

• Unit tests
• Integration tests
• Example config in examples/configs/
• Functional integration test for the example config
• Generated filter documentation regenerated