Repo-wide doc consistency pass + run --snapshot (0.3.3)

[MarouaBoud]

Repo-wide consistency and accuracy pass so the docs match the shipped code, plus a small feature that makes one doc claim true.

Feature

• mockworld run --snapshot <file> boots a server with a saved .mw.json snapshot restored (reproduce a bug from a shared artifact). Verified: save → run --snapshot restores the exact charge.

Docs reconciled

• Trace attributes corrected to the shipped swarmproof.fault.kind (+ swarmproof.run.seed) across CLAUDE.md, AGENTS.md, docs/PRD.md (REQ-OBS-1), docs/RESEARCH.md (Q1), docs/ARCHITECTURE.md, docs/TEST-PLAN.md (INT-4). The old fault.{type,injected,source} triple is gone from prose.
• Stampede §7 (ARCHITECTURE): status banners note the adapter ships in the stampede repo and the "confirmed" framing predates the real integration; §7.5 reworded from "all open items resolved" to the shipped reality.
• Stale numbers: README test count 86→94; ROADMAP version made version-agnostic; CHANGELOG "v0.1–v0.4"→"v0.1–v0.3"; RELEASING verify snippet de-pinned.
• Install name: remaining pip install mockworld → mockworld-mcp in SPEC/PRD/DELIVERY-PLAN/CHANGELOG.
• Status banners added to RESEARCH/DELIVERY-PLAN/TEST-PLAN/ARCHITECTURE (SPEC/PRD already had them); SPEC banner qualified ("shipped except deep stampede integration").
• Mock count framed consistently as "6 built-ins = 5 services + a hello example."
• Swarm misuse figure in swarm.py docstring aligned to the real ~33%.

Bumps to 0.3.3 so the published package includes run --snapshot and the corrected README (PyPI long-description).

Verification

94 tests pass; twine check passes; a fresh grep finds no residual stale trace attrs, versions, test counts, or pip install mockworld (non-mcp) in prose.