feat: swarm dispatch — autonomous decompose-and-fan-out

[adbarc92]

@

Summary

Adds a daemon-orchestrated swarm: POST /swarms points the engine at a doc/spec path inside the target repo, a planner agent autonomously decomposes it into independent lanes, and the daemon fans out one agent unit per lane — each running the existing SP1 dispatch→oracle→build/review→PR pipeline unchanged, on its own branch/PR. Bounded by a hard lane cap and a per-swarm worst-case USD ceiling.

Built as a pure, sync admission core + two seam traits with fakes, layered over the untouched per-unit driver. Spec hardened through 3 adversarial critique rounds; design + plan committed under docs/superpowers/.

What landed

• Engine preconditions (P1–P4): next_id/next_swarm seeded from the store on startup (restart-safe id allocation); committed_spend (counts non-terminal reservations as MAX(usd_cap,cost), terminal cost, and planner cost) replaces the cap-blind spend_since on both admission paths — each as a single check+insert critical section; one authoritative TERMINAL_PHASE_STRS consumed by reconcile + rollup.
• Pure core (swarm.rs): admit_lanes dual-guardrail (lane cap + usd_budget − planner_cost envelope) and a sanitizing slug.
• Seams: Planner (FakePlanner / ClaudePlanner) and DocSource (FakeDocSource / GitDocSource), selected by mode so a demo swarm never makes a paid call.
• Persistence: swarms + swarm_lanes tables, units.swarm_id, atomic commit_lane_unit (row + back-link in one txn), swarm_rollup.
• Orchestration & API: run_swarm (plan→admit→fan-out, states planning→fanning_out→running / failed / empty; running→done computed at read); POST /swarms (sync-validated, returns immediately), GET /swarms, GET /swarms/:id.
• Restart-safety: planning→failed; fanning_out→idempotent resume.
• Security: GitDocSource validates against argv flag-smuggling and path traversal (https-only, no leading -, no ../absolute doc paths, post-clone canonicalize+containment).

Test Plan

• cargo test — fleet-core 30 + fleetd 71 pass, 0 failed; 3 Docker/git-gated integration tests ignored by design
• cargo clippy -p fleet-core -p fleetd --all-targets -- -D warnings — clean
• P2 concurrent-admission regression test (two missions racing the cap → exactly one admitted), verified non-flaky
• Manual: real-mode swarm against the sandbox repo (requires Docker + gh + ANTHROPIC_API_KEY)
  @