feat(init): force CodeGraph usage guidance into each agent's always-on instructions

[cvetty]

Context

whygraph init --agent <name> already copies a per-agent asset tree into the target repo (WhyGraph's own subagents, skills, slash commands, rules) but ships no CodeGraph usage guidance. Because WhyGraph sits on top of CodeGraph — every whygraph scan builds/refreshes the .codegraph/ index — a developer in a WhyGraph-initialized repo has the codegraph_* MCP tools available, but their agent has no instructions on how to use them, so it tends to grep instead, or dump huge codegraph_explore output into context.

CodeGraph's own Claude Code installer solves this by injecting a guidance block into .claude/CLAUDE.md. This PR does the equivalent for every agent WhyGraph supports, so the guidance is forced — it lives in each agent's always-on instruction surface, not a lazily-triggered skill. (A skill / description-triggered rule may not fire before the wrong reflex — grep — wins.)

Per-agent always-on surface

Agent	Surface	Mechanism
Claude Code	.claude/CLAUDE.md	new append-merge file (assets_merge_files)
Codex	AGENTS.md	extend existing merge file
Copilot / VS Code	.github/copilot-instructions.md	extend existing merge file
Cursor	.cursor/rules/whygraph-codegraph.mdc	new rule with alwaysApply: true

The Cursor catch: its existing four rules are alwaysApply: false (lazy, like skills) — not forcing. CodeGraph forcing needs a dedicated alwaysApply: true rule.

Content

• Source of truth: CodeGraph's own block (verified byte-identical across the repo's .claude/CLAUDE.md, the user global, and the npm template).
• Claude Code: vendored verbatim ("Explore agent" / "main session" are native Claude Code concepts).
• Cursor / Codex / Copilot: same forcing rules in neutral wording (drop Claude-isms; keep "prefer codegraph_* over grep", "don't pull large explore output into context", lightweight-tool table).

Changes

• Only code change: assets_merge_files=("CLAUDE.md",) on the _CLAUDE target in agents.py. The existing _merge_block machinery handles fresh-write / replace-in-markers / append idempotently — no machinery changes.
• New bundled claude-code/CLAUDE.md + cursor/rules/whygraph-codegraph.mdc; extended codex/AGENTS.md + vscode/copilot-instructions.md.
• One-sentence note in the project CLAUDE.md.
• Tests: count/manifest assertions updated (8→9 files, +cursor rule), marker assertions on the Claude/Cursor init tests, and a new test_init_agent_claude_merges_existing_claude_md.

Verification

• uv run pytest → 451 passed.
• Manual init per agent in a scratch repo: Claude gets the verbatim block, Codex/Copilot the neutral section, Cursor the alwaysApply: true rule — all under <!-- BEGIN whygraph --> markers (frontmatter for Cursor).
• Re-running init --agent claude keeps exactly one WhyGraph block (idempotent, non-destructive of user content).

Known tradeoff

If a user also runs CodeGraph's own installer, its block (under <!-- CODEGRAPH_START -->) and WhyGraph's (under <!-- BEGIN whygraph -->) coexist as two near-identical blocks. Acceptable — WhyGraph's path is self-sufficient by design. No de-dup logic in v1.