Rebuild: strip the agent harness, pivot to a tmux plugin

[apriendeau]

Throws out the RPC/provider agent harness and repoints the repo at a much smaller tool: a
tmux plugin that spawns isolated coding-agent instances in a polyglot monorepo.

41,319 deletions, 647 insertions. Rust goes from ~34k lines to 538.

Why

The harness was solving model orchestration. The actual problem is running several agent
CLIs (Claude Code, Codex, Grok, whatever) in parallel against one monorepo without them
colliding — and the parts of that worth building are worktree lifecycle and environment
materialization, not panes or provider APIs.

tmux already owns panes, input, resize, scrollback, and detach/reattach.
opensessions already owns the sidebar and
agent-state display. Neither materializes a usable instance of a monorepo, and that gap is
the whole product.

Design

Full write-up in DESIGN.md. Two properties are load-bearing:

• No daemon. Every invocation is transactional and exits. No server, no socket, nothing
  with a lifecycle.
• tmux is the state store. Instance metadata lives in window options (@phx_worktree,
  @phx_branch, …) and the window id is the primary key. No registry file, nothing to desync.

Shape is a phx.tmux shim binding four keys (new, done, discard, rename) plus a
small binary those bindings invoke. Arguments arrive via tmux format substitutions; there is
no human-facing CLI.

The core is materialization — APFS clonefile for node_modules and target, copied
.env and MCP config, and per-instance env for ports, COMPOSE_PROJECT_NAME, and sccache.
Without copy-on-write cloning, worktree-per-agent isn't viable on a real monorepo.

Removed

• All fourteen superseded src/ modules plus crash.rs
• The five example plugins and the cargo workspace around them
• DESIGN.md and SPAWN_AGENT_SPEC.md (superseded), plugin/rpc/miri justfile recipes
• Eight dependencies: ratatui, crossterm, reqwest, syntect, uuid, chrono, and friends
• beads entirely — .beads/, the bd prime SessionStart and PreCompact hooks, the
  bd-init recipe, gitignore entries, and the tracker sections of CLAUDE.md / AGENTS.md.
  Work is tracked in Jira and Linear; the repo carries no local tracker. This also removes
  the old mandatory-push session protocol.

Kept

src/worktree.rs and its tests — create/remove/list/diff/merge, zero internal dependencies,
all six tests still green. It maps directly onto the new design.

CI

• Miri job dropped: no unsafe remains, and Miri can't spawn the git subprocesses the
  tests need
• crates.io publish dropped: TPM is the distribution channel now
• Windows release target dropped: tmux is unix-only
• CI now runs on macOS as well as Linux, since clonefile is macOS-specific
• Release assets keep the phx-<target>.tar.gz name — that's the contract with phx.tmux,
  which downloads a binary on first load

Not in this PR

src/main.rs is a clap skeleton; every subcommand returns "not implemented". Per
DESIGN.md §11 the next step is materialize standalone, proven against a real monorepo —
that's where all the risk sits. Four open questions are tracked in §9, the sharpest being
clone freshness when lockfiles drift.