fix: accept string Responses content and prepare v0.3.1

[Zesuy]

Summary

• Accept valid string-form OpenAI Responses input[].content values instead of rejecting the request as malformed.
• Preserve string content unchanged while continuing to inspect rich function_call_output.output[] image blocks.
• Normalize ordered user-context selection across Responses, Chat Completions, and Anthropic Messages.
• Add regression fixtures and tests for string history, rich tool output, empty-turn boundaries, and release contracts.
• Prepare the v0.3.1 version, packaging defaults, changelog, and documentation.

Root cause

The Responses image walker assumed every non-null input[].content value was an array. Standard Responses requests may use a string shorthand, so valid text-only requests were incorrectly terminated with HTTP 400 before reaching the upstream provider.

User impact

String-form Responses input now passes through unchanged. When a string-bearing item also includes a rich function output, image discovery and rewriting still run.

Validation

• go test -count=1 ./...
• go test -race -count=1 ./...
• go vet ./...
• go build ./...
• ./scripts/verify-contracts.sh
• ./scripts/package-smoke.sh
• make verify-version-override
• Docker Compose configuration validation
• Docker BuildKit artifact build
• CLIProxyAPI host E2E with mock providers
• Independent regression and release-readiness review

Release notes

PR CI validates the Linux amd64 package artifact. The full Linux/macOS/Windows amd64/arm64 matrix remains part of the Release workflow. The billable live-provider test was not run.

Fixes #8