docs: record CodeGraph worktree discovery

[joe-crick]

Stage 0 for the CodeGraph worktree-aware integration plan.\n\nChanges:\n- Documents current CCE path-based storage identity.\n- Documents CodeGraph worktree mismatch behavior and machine-readable surfaces.\n- Records selected Git base-ref strategy and MVP limitations.\n\nStack note:\n- Upstream push permission is denied for joe-crick, so later stacked stages are open in the fork with bases chained branch-to-branch: joe-crick#1 through https://github.com/joe-crick/code-context-engine/pull/9.\n\nChecks:\n- uv run python -m pytest -o addopts='' tests/test_project_storage_dir.py tests/integration/test_git_context.py

[rajkumarsakthivel]

Review

Thanks for the work here. The architecture is clean (git identity -> storage layout -> overlay backend -> fusion -> MCP) and the test coverage for the new modules is solid. This is a meaningful feature addition.

A few things to address before merging:

Scope and labeling

This PR adds ~2960 lines of new feature code across 37 files (new modules, MCP server changes, CLI commands, config additions). The title says docs: but this is a feat: PR. Please relabel so the changelog and release notes reflect what actually shipped.

Code concerns

1. WorktreeOverlayBackend doesn't implement StorageBackend formally. The codebase uses getattr checks for count_chunks, needs_overlay_index, etc. throughout cli.py and mcp_server.py. If StorageBackend gains a new required method, this will silently break. Consider subclassing or a protocol.

2. Private attribute access. _count_chunks falls back to backend._vector_store.count(). This pattern is duplicated in multiple files. Should be a proper interface method on StorageBackend.

3. No file size or type filtering. _source_for_worktree_path reads entire files into memory for every modified/added path on each context_search call. Large binary or generated files would be read and scanned with no guard.

4. Blocking subprocess in async context. CodeGraphClient._run_json and git helpers use synchronous subprocess.run but are called from async code paths. This blocks the event loop. Consider asyncio.create_subprocess_exec or run_in_executor.

5. Token estimation. _structural_token_count uses len(text) / 3.3 as a heuristic. The existing codebase has proper token counting. Using a different method here creates inconsistency.

6. Absolute paths in discovery doc. docs/codegraph-worktree-integration-discovery.md references /home/joe/Webstorm_Projects/codegraph/.... Please remove machine-specific paths.

What's good

• Backward compatible: structural_provider defaults to "off", existing users see no behavior change
• Clean module separation with proper __init__.py exports
• Good test coverage for overlay backend, fusion, merge, codegraph provider, and budgeting
• The layered architecture makes sense

Worth pursuing, just needs the above addressed first.