feat: provider dispatch — backend registry (claude/codex/sdk/local)

[justrach]

Parent: #257 — Layer 0 (Plumbing)

What

Provider-agnostic dispatch function that spawns one agent turn on one specified backend. No resolution logic, no preferences — just execute what resolve() decided.

const Backend = enum { claude, codex, sdk, local };

fn dispatch(
    alloc: Allocator,
    backend: Backend,
    model: []const u8,
    prompt: []const u8,
    opts: DispatchOpts,
    out: *ArrayList(u8),
) void {
    switch (backend) {
        .claude => spawnClaude(alloc, model, prompt, opts, out),
        .codex  => spawnCodex(alloc, model, prompt, opts, out),
        .sdk    => querySDK(alloc, model, prompt, opts, out),    // #178
        .local  => spawnLocal(alloc, model, prompt, opts, out),  // future
    }
}

Backends

Backend	How	When available
claude	Spawn claude -p --model X (current tryClaudeAgent logic)	claude on PATH
codex	Spawn codex app-server (current codex_appserver.runTurnPolicy)	codex on PATH
sdk	Claude Agent SDK query()	When #178 lands
local	Spawn local model server (ollama, llama.cpp)	Future

Changes

• Extract tryClaudeAgent() from agent_sdk.zig into spawnClaude() in dispatch module
• Extract codex_appserver.runTurnPolicy() into spawnCodex()
• Each backend handles its own sandbox/permission mapping
• Login shell fallback stays in spawnClaude()

What this does NOT do

• No resolution logic (that's resolve())
• No model/provider selection (that's the grid)
• No prompt injection (that's run_agent)

Depends on

• feat: provider detection — ProviderSet.detect() probes environment #264 (provider detection — to know which backends are available)

Blocks

• run_agent refactor (needs dispatch to send agents)

[justrach]

Blocks: #258, #259, #260, #263, #264, #265, #266, #267, #268, #269