proto: subagent prompt/history as structured AgentStart fields

[christopherwxyz]

Summary

Replace the planner's synthesized envelope string with two typed
AgentStart proto fields.

message AgentStart {
  string agent_id        = 1;
  bytes  agent_config    = 2;
  repeated Message messages = 3;
  string subagent_prompt  = 4;  // NEW: planner's `prompt` arg verbatim
  string subagent_history = 5;  // NEW: planner's `history` arg verbatim
}

Motivation

The Gemini planner currently synthesizes a multi-line string envelope
as Messages[0]:

History Summary:
<stringified conversation history>

Prompt:
<the planner's `prompt` function-call argument verbatim>

Every subagent that wants clean access to the typed prompt arg has
to string-parse this envelope. Any change to the envelope wording
silently breaks every subagent.

Change

• Add subagent_prompt (field 4) and subagent_history (field 5) on
  AgentStart. Planner populates them from the typed Gemini
  FunctionCall.Args directly.
• The planner no longer synthesizes a fake user-role Message. For
  planner-driven dispatch, AgentStart.Messages is empty.
• Messages remains the wire surface for direct (non-planner)
  callers (e.g. grpcurl-driven testing of a subagent).

Breaking change

Subagents that currently read lastUserText(start.Messages) on
planner-driven invocations must migrate to:

prompt := start.GetSubagentPrompt()
if prompt == "" {
    // direct caller path — read from Messages as before
    prompt, _ = lastUserText(start.Messages)
}

The reference python_sandbox_agent example in PR #3 is updated to
this pattern in a sibling commit.

Test coverage

TestHandleSubagentCall_PopulatesStructuredFieldsOnly — asserts:

1. SubagentPrompt + SubagentHistory populated from typed args
2. Messages is empty (regression guard against re-introducing the
   envelope)

go test -race ./... clean.

Related

• PR feat(gemini-planner): support native Gemini tools (google_search, url_context, code_execution, google_maps) #2 (planner native tools) — independent, no line-level conflict
• PR feat(experimental): agent-sandbox Kubernetes backend #3 (agent-sandbox K8s backend + python_sandbox_agent reference impl) — picks up the new contract