chore: roadmap, readme and general docs updates

[haoshan98]

Summary

Several user-facing docs had fallen behind main or contradicted each other and the code. This PR brings them in line with what has shipped. It changes documentation only.

Corrected claims

• README listed the Messages API as "🚧 Planned" while another README section documents /v1/messages. It is now listed as shipped; only stateful Messages stays planned.
• README and docs/index.md advertised "Background execution" and gateway-executed file search. Neither exists yet: background mode is Support asynchronous/background Responses #357, and file_search declarations parse but are dropped before inference (tool/normalize.rs). Both claims are removed.
• docs/api/index.md quoted fixed WebSocket and executor limits (256 KiB SSE lines, 1 MiB events and response budget). These now come from the configurable [responses] settings, which default to 16 MiB and 8 MiB.
• docs/api/index.md listed only some of the conditions that route a Responses request through the executor. The list now also includes non-function tools, multi_agent.enabled, and tool-search state, matching handler/http/responses.rs.
• CHANGELOG listed Tavily under 0.8.0, but it merged on Sep 28, after the Sep 19 release. The two Tavily entries move to [Unreleased].

Newly documented features

• README: a multi-agent Responses section with an example and the gateway's limits (store: true, HTTP/SSE only, default of 3 concurrent subagents, no max_tool_calls or reasoning summaries). Also new rows in the API table for /v1/responses/compact, the full Conversations routes, /v1/messages with /v1/messages/count_tokens, and Swagger/OpenAPI (ENABLE_OPENAPI_DOCS). The features list, diagram, and tool ownership table now include MCP, the code interpreter, Claude Code, compaction, and OpenTelemetry.
• docs/api/index.md: new sections for GET /v1/responses/{response_id}, multi-agent, Conversations, and Messages.
• CHANGELOG [Unreleased] was empty. It now has entries for fix: reject incomplete and malformed Messages streams #345, feat: preserve prompt_cache_key in Responses execution #347, feat(telemetry): add execution traces and upstream context propagation #349, feat: add OpenAI-compatible Conversations API #351/fix: address Conversations API review comments and add cassette tests #370, fix: support retrieval of stored Responses payloads #354, feat: support code interpreter as gateway execution tool #368, Forward Chat Completions to the upstream. #369, feat: [1/2] Model multi-agent Responses items and event lifecycles #372/feat: [2/2] Execute multi-agent HTTP Responses and verify recorded contracts #373, fix: disable Nagle's algorithm on accepted gateway connections (#376) #380, and fix: do not store a store=false continuation of a stored response #387.

Roadmap

• ROADMAP.md marks multi-agent HTTP execution as shipped, with WebSocket injection (match OpenAI multi-agent WebSocket injection and record transport conformance #300) and stateless requests (support stateless multi-agent Responses requests #393) still open. It adds a status note for Messages, a code interpreter entry, and a file search item linking feat: add Files and Vector Stores APIs with in-tree file search #34 and Support a Files API and file_id hydration for Responses requests #356.
• README "Roadmap at a Glance" checks off Messages and Conversations and adds unchecked items for multi-agent, file search, and stateful Messages.

Smaller fixes

• The repository layout in README and AGENTS.md now lists agentic-llm-d, agentic-cli-docs, and python/agentic_api/.
• ARCHITECTURE.md's list of multi-agent rejections now includes max_tool_calls, matching executor/rehydrate.rs.
• mkdocs.yaml adds the existing but unlinked design/shell-tool.md to the nav.

Not changed: the README's "Rust 1.85+" badge is wrong (1.85 cannot build the lockfile, see #377), but #379 already updates it to 1.88 along with rust-version and a CI check, so this PR leaves it alone to avoid a conflict.