feat(libsy-server): demo libsy + random routing + affinity + builtin proxy

[messiaen]

Summary

Introduces libsy-server, a minimal axum proxy that fronts coding agents
(Claude Code, Codex) and A/B-routes each request between a weak and strong
upstream tier using libsy's RandomAlgo. Also merges in agent-aware routing /
affinity (from agent/subagent-aware-routing) and fixes two Anthropic→OpenAI-
Responses translation bugs that broke tool-using agent turns.

What's included

New crate — libsy-server (crates/libsy-server/)

• HTTP surface for OpenAI Chat, OpenAI Responses, and Anthropic Messages
  (/v1/chat/completions, /v1/messages, /v1/responses, /v1/models, /health).
• Decode → route via RandomAlgo → translate to the chosen tier's format →
  translate back to the inbound format. Buffered and streaming (SSE) supported.
• CLI: --weak/--strong + --weak-format/--strong-format (per-tier upstream wire
  format, credential picked by provider), --affinity [session|subagent],
  --log-routing, --host/--port.
• Populates request metadata from headers so affinity can key on it; recognizes
  Claude Code's X-Claude-Code-Session-Id.
• Caller auth headers are dropped and the real backend credential injected
  (sentinel-key friendly).
• README with quickstart + Claude Code setup; integration tests via
  tower::oneshot + wiremock.

Agent-aware routing / affinity (crates/libsy/, merged)

• affinity module: Affinity trait, SessionAffinity, SubAgentAffinity, and
  metadata_from_headers (Switchyard/Codex/Relay/Dynamo/Claude Code header
  normalization).
• RandomAlgo::with_affinity(...); AgentContext on protocol Metadata.
• demo/libsy-proxy POC.

Translation fixes (crates/switchyard-translation/)

• Anthropic tool_use → Responses function_call now emits arguments as a
  JSON string (was an object → API 400).
• Anthropic thinking → Responses now dropped instead of encoded as a
  content-bearing reasoning input item (was → API 400), matching the OpenAI Chat
  encoder. Reasoning-only messages emit nothing; private reasoning never leaks into
  content.

Testing

• libsy-server: 9 integration tests (buffered/streaming per format, cross-format
  routing, error mapping).
• switchyard-translation: existing suite green + 2 new regression tests for the
  arguments/reasoning fixes.
• libsy: affinity unit/integration tests (incl. Claude Code session header).
• Verified end-to-end with Claude Code: streaming, tool use, and multi-turn work
  across both tiers; concurrent sessions each pin to one tier under
  --affinity session (deterministic: A→weak 12/12, B→strong 12/12).

Notes for reviewers

• Behavior change: with per-tier formats the proxy accepts any inbound endpoint
  and translates to the tier's format (no longer same-format-only), so there's no
  "unsupported inbound format" 400.
• Affinity is all-or-nothing per session — a session pinned to the weak tier
  runs the whole task on it (intended for per-session A/B measurement).
• One change touches the just-merged metadata_from_headers to recognize
  X-Claude-Code-Session-Id.

Checklist

• One class per file; filename = snake_case of the primary class.
• New public symbols exported from switchyard/__init__.py.__all__ if intended for downstream use.
• Unit tests added for new components / bug fixes.
• README / --help updated if customer-facing surface changed.
• Commits signed off (Signed-off-by: Your Name <email>) per the DCO.

Notes for reviewers

Anything reviewers should pay extra attention to — risky paths, follow-up tickets, intentional trade-offs.