fix(claude): honor CLAUDE_CONFIG_DIR when listing session history

[pjdoland]

Summary

The Claude CLI relocates its entire config dir, including the session transcripts under projects/, when CLAUDE_CONFIG_DIR is set. NBI's session history surfaces (the chat-sidebar resume picker and the launcher tile's session list) always read ~/.claude/projects, so on deployments that set the override (JupyterHub images commonly point it at a persistent volume, e.g. /home/jovyan/workspace/.claude) both pickers came up empty even though claude --resume saw every session.

Solution

The load-bearing change is in claude_sessions.py: get_sessions_dir and list_all_sessions now default claude_home to a new util.get_claude_config_dir() helper (CLAUDE_CONFIG_DIR when set, else ~/.claude) instead of a hardcoded Path.home() / ".claude". The skills directory (config.py) and spinner-verbs lookup (extension.py) already carried their own inline copies of the same expression; they now route through the helper too, so the override semantics cannot drift between surfaces.

Two deliberate choices, both verified against the CLI's behavior: the helper does not expanduser the env value (the CLI doesn't expand ~ either, so a literal-tilde value should fail identically in both), and an empty string counts as unset (matching the CLI's Node-style falsy check). The docs filesystem table and the troubleshooting diagnostic listing now mention the override, since ls ~/.claude/projects/ is exactly the misleading step on an affected deployment.

Testing

• Four new tests in tests/test_claude_sessions.py (TestClaudeConfigDirDefault): get_sessions_dir honors the env var and falls back to the home dir (HOME and USERPROFILE both pinned, so the assertion holds on POSIX and Windows); list_all_sessions discovers transcripts under $CLAUDE_CONFIG_DIR/projects with no claude_home argument (the exact production call shape); and a negative test proving ~/.claude is not consulted when the override is set. All existing tests pass claude_home explicitly, so the suite stays insulated from a developer's exported CLAUDE_CONFIG_DIR (verified by running the full file with a junk value exported).
• Full verification green: pytest (1177 plus the claude_client suite separately), tsc --noEmit, lint:check, jest (352).
• Live verification: started JupyterLab with CLAUDE_CONFIG_DIR pointing at a synthetic transcript store; GET /notebook-intelligence/claude-sessions returned exactly the synthetic session for both scope=all and scope=cwd, and none of the real sessions in the actual ~/.claude/projects, confirming the override replaces the home dir the same way the CLI does.
• The SDK subprocess inherits os.environ, so NBI-created Claude sessions land in the same override dir the picker now reads, and Jupyter terminals inherit the server env, so the generated claude --resume command finds them.

Risks / follow-ups

• No behavior change when CLAUDE_CONFIG_DIR is unset: the helper's fallback is the same ~/.claude path as before (via expanduser, which Path.home() also uses).
• Two adjacent surfaces have the same class of gap but need different relocation logic, so they are excluded here and tracked separately: claude_mcp_manager.py reads ~/.claude.json (the CLI moves that file inside $CLAUDE_CONFIG_DIR when set, but its default location is the home dir itself, not ~/.claude), and plugin_manager.py falls back to ~/.claude/plugins for the plugin cache.

Fixes #373

[pjdoland]

Resolved conflicts.