fix(agents): launch panic + parallel concurrent agent dispatch (v0.1.12)

[ste-bah]

Summary

THE central failure that's dragged on for 2 weeks. Discovery worked, listing worked, permissions worked — agents never actually launched. Now they do, AND N concurrent agents per turn.

Root cause #1 — panic on launch

crates/archon-memory/src/embedding/openai.rs::request_batch called reqwest::blocking::Client::send() WITHOUT a block_in_place boundary. reqwest::blocking constructs a tokio runtime per call; from inside an async tokio task this triggers tokio's nested-runtime panic: "Cannot start a runtime from within a runtime / thread is being used to drive asynchronous tasks." Every agent dispatch that touched memory crashed.

Root cause #2 — no parallel dispatch

crates/archon-core/src/subagent.rs::run iterated pending_tools in a serial for loop. When the LLM emitted N tool_use blocks in one turn, they ran one after another — no concurrency.

Fix

• A. Wrap OpenAIEmbedding::embed body in tokio::task::block_in_place. Panic gone.
• B. Replace serial for-loop with claurst's proven three-phase pattern (/tmp/claurst/src-rust/crates/query/src/lib.rs:1815-1944):
  • Phase 1: sequential pre-hooks (permission gate)
  • Phase 2: futures::future::join_all over Vec<Either<Left ready, Right execute>> — concurrent execution, order preserved natively
  • Phase 3: post-hooks + result assembly
• C. No changes to AgentTool::execute. Verified reentrant by code inspection: BACKGROUND_AGENTS uses per-agent_id inserts, status/result_slot/cancel_parent are per-call Arcs, no global serialization. AGS-104/105/107 contracts preserved unchanged. Added barrier-based regression test for deterministic concurrency proof.

Test plan

• A: archon-memory regression test — embed() inside multi_thread runtime returns 401 from fake key, no panic (0.79s)
• B: subagent parallel dispatch test — 3 stub tools complete in 0.40s vs ~0.90s serial baseline, results in input order
• C: archon-tools barrier test — tokio::sync::Barrier::new(2) rendezvous proves both subagents in flight simultaneously, BACKGROUND_AGENTS count = 2 at rendezvous (<1ms wall-clock)
• D (E2E): bin agent_parallel_smoke — 3 concurrent AgentTool calls complete in 0.50s vs ~1.50s serial
• cargo nextest run --workspace: ALL PASS, zero failures
• cargo fmt --check: PASS
• cargo build --release --bin archon -j1: PASS
• TUI binary smoke (Steven) — verify in real session