feat: Talk to agents over A2A (agent card / send / task)

[a1594834522-coder]

What

Binds the A2A protocol endpoints in the open-api to the CLI, under the existing agent tree:

memorylake agent card <agent>
memorylake agent send <agent> --text … [--context] [--task] [--stream|--no-wait] [--raw]
                                 [--actor] [--project] [--read-only-project]… [--skip-memory] [--metadata-json]
memorylake agent task list|get|cancel|feedback <agent> …

All take --workspace with the same default as agent bind.

Design decisions

• A2A v1.0 HTTP+JSON only. The unversioned /a2a/... paths are v1.0; the /a2a/v1/ paths are protocol 0.3 (confirmed by the agent card's supportedInterfaces). JSON-RPC and the 0.3 REST surface are not bound.
• Responses are serde_json::Value. The API documents them without a schema; success bodies are bare A2A documents, not envelopes. Errors still arrive as envelopes, and describe_a2a_error lifts the A2A message + reason out of the escaped document so task cancel on a finished task says Task cannot be canceled - current state: 3 (TASK_NOT_CANCELABLE) rather than a blob of \".
• send blocks by default (measured: production returns TASK_STATE_COMPLETED with artifacts in a few seconds). Reply text → stdout, task … context … state … → stderr, so scripts can pipe the reply and still continue the thread. TASK_STATE_INPUT_REQUIRED prints how to answer.
• --stream reads the SSE body with a small parser in core::sse. Production streams the reply token-by-token in statusUpdate events and then once more in full as an artifactUpdate; the artifact is printed only when nothing was streamed.
• A2A-Extensions: extensions://task-feedback/v1 is sent on task get, task list and task feedback. Measured: feedback records without the header, but task get only shows metadata."task-feedback/v1" with it.
• --actor is not defaulted to actor me; production accepts an actor that is not bound to the workspace, but the server's own default is the right fallback.

Not in this PR

agent task subscribe (reattaching to a --no-wait task's stream), and flattened flags for overrides — those go through --metadata-json.

Tests

• core: SSE parser, request serialization (camelCase, absent-when-empty), path constants, error rewrite, header/stream behaviour against the loopback server.
• CLI wire: method/path/query/body/header for every subcommand; stream stub writes frames one at a time with pauses.
• CLI offline: help, workspace required, content flags exclusive, --stream/--no-wait conflict, feedback validation, --page-size range.
• CLI live: a2a_round_trip_against_the_default_agent — card → send (--skip-memory) → task get → list → feedback → cancel refused → stream prints once. Passes locally; full cargo test --workspace green apart from the known transport flakiness when live tests run in parallel.

🤖 Generated with Claude Code