docs: rewrite top-level docs for v0.4 (Go binary, no Python runtime)

[couragehong]

The six top-level docs still described the v0.3 install flow — a Python venv prepared by scripts/bootstrap-mcp.sh, mcp/server/server.py spawned under .venv/bin/python3, agents/{common,scribe,retriever}/ as the business-logic packages, etc. After PR #129 (remove-python-tree) all of that is gone, but a contributor following these docs would still try to set it up and land in a dead end.

This rewrite anchors every doc on the v0.4 truth: a single Go binary (cmd/rune-mcp) that the host CLI auto-spawns over stdio, runtime preparation done at install time, runtime health visible through the diagnostics + vault_status MCP tools.

Per file

README.md (small)

• drop the dead scripts/check-infrastructure.sh reference
• point Related Projects at envector-go-sdk (the Go FHE SDK we actually depend on now) instead of pyenvector

CLAUDE.md (one-line)

• "Writing Python scripts to /tmp" → "Writing scripts to /tmp" in the routing table NOT-do column (the rule still holds; the Python qualifier was v0.3-specific)

GEMINI.md (medium)

• replace the Plugin Root Detection / Runtime Preparation block, which mandated SETUP_ONLY=1 scripts/bootstrap-mcp.sh and .venv/bin/python3 sanitization, with v0.4 guidance: the binary is auto-spawned, health is observable via diagnostics, no venv exists

SKILL.md (medium-large)

• Execution Model: drop scripts/bootstrap-mcp.sh as the "single source of truth" — the binary itself is now the runtime
• Activation Check: drop the Local Runtime Check (step 0) that required the bootstrap script; drop the envector.endpoint / envector.api_key field requirement (those come via Vault bundle now)
• /rune:configure: drop scripts/check-infrastructure.sh validation step; replace with reload_pipelines + diagnostics rendering
• /rune:status: drop "Python venv" / "MCP servers running" rows; render diagnostics' per-subsystem snapshot instead
• /rune:activate: drop the "check Python environment" step
• /rune:reset: drop "stop MCP servers" — the host CLI manages process lifecycle, the user does not

AGENT_INTEGRATION.md (full rewrite, 278 lines → equivalent shape)

• drop all .venv/bin/python3 + mcp/server/server.py invocations across Claude / Codex / Gemini / OpenAI sections
• replace install.sh / install-codex.sh / ensure-codex-ready.sh references with plugin/skill/extension install commands
• rewrite the OpenAI integration around stdio against the Go binary (the agents-sdk + Responses API patterns just point command= at bin/rune-mcp)
• acknowledge that gemini-extension.json still calls bootstrap-mcp.sh as a known follow-up under the Gemini track (not in this PR's scope to resolve)

CONTRIBUTING.md (medium)

• Prerequisites: Python 3.12 → Go 1.26.2+ (toolchain pinned in go.mod)
• test instructions: pytest agents/tests/ → go test -race ./...
• Cross-Agent Invariant Checklist: replace bootstrap-mcp.sh items with "Go binary is single MCP entry point" / manifest consistency
• Project Structure: redraw the tree against current layout (cmd/rune-mcp, internal/{mcp,service,lifecycle,adapters,domain, policy,obs}, agents/{claude,codex,gemini}/*.md only, scripts/dev/ v04/, etc.) — the old tree showed mcp/server/server.py and agents/common/ that no longer exist
• Release Process: bump example version 0.2.0 → 0.4.0; tests command updated; reference Task Sharing low retrieval correctness case #30 for the binary distribution pipeline

Out of scope

• gemini-extension.json bootstrap-mcp.sh path (Gemini track)
• scripts/ Python-era shell scripts cleanup (Gemini track)
• examples/ patterns/ setup/ config/ docs (separate A4 pass)
• package.json + .claude-plugin/plugin.json version bump (separate A5)