feat: CLI update command (PR-I2 of 2-PR sequence)

[HumanBean17]

Summary

Implements the update subcommand that refreshes shipped artifacts (skill, agent, MCP entry) after pip upgrade.

Features

• Host detection: Scans project + user config files for java-codebase-rag MCP entries
• Artifact refresh: Overwrites skill/agent files from package data, updates MCP command path
• Index update: Runs incremental LanceDB update if index exists
• Dry-run mode: Previews changes without writing files

Changes

• Add detect_configured_hosts() function to scan for configured hosts
• Add refresh_artifacts() function to refresh skill, agent, and MCP files
• Add run_update() orchestrator for the update pipeline
• Add update subcommand to CLI with --force and --dry-run flags
• Add 12 new tests for update functionality

Test plan

• All 12 new tests pass
• Existing test suite passes
• ruff check . is clean
• update --help prints usage
• update after install completes with exit 0
• update --dry-run reports what would change without writing
• update with no configured hosts exits with code 2

Part of

This implements PR-I2 from plans/active/PLAN-CLI-INSTALL.md.

🤖 Generated with Claude Code

[HumanBean17]

Review: PR-I2

Tests pass (65/65), ruff is clean, and the scope matches the plan. A few issues to address:

Critical

SystemExit leaks through _refresh_mcp_config (installer.py:1020)

resolve_mcp_command(non_interactive=True) raises SystemExit(2) when the binary isn't on PATH. SystemExit inherits from BaseException, so the outer except Exception as e at line 1095 won't catch it. If java-codebase-rag-mcp isn't on PATH, the entire update command aborts with a bare exit 2 instead of returning a graceful ArtifactResult failure.

Fix: catch SystemExit explicitly, or resolve the command path before calling _refresh_mcp_config and pass it in (like merge_mcp_config takes mcp_command as a parameter).

Important

Duplication between _refresh_mcp_config and merge_mcp_config (installer.py:1000-1096 vs 420-465)

The JSON read / compare / atomic-write logic is nearly identical. The only real differences are the force/dry_run branching and the return type (ArtifactResult vs bool). Consider factoring the shared write-atomic logic into a common helper, or having _refresh_mcp_config call merge_mcp_config internally and wrap its result.

run_update swallows partial failures (installer.py:1152-1172)

If some artifacts failed to refresh (warnings printed at lines 1137-1141), but the function then hits an early return at line 1155, 1164, or 1172, it returns 0 despite partial failures. The exit code should reflect partial failure (1) when any artifact failed.

Minor

• EXIT_* constants unused (installer.py:33-35): the refactor added EXIT_SUCCESS/EXIT_PARTIAL/EXIT_FATAL but run_update uses raw 0/1/2 literals. Either use them or remove them.
• test_update_no_index_skips_increment sets up fake_stdout but never asserts on the captured output.
• No test for missing java-codebase-rag-mcp binary inside _refresh_mcp_config — the most interesting edge case for the update flow.

Verdict

Fix the SystemExit leak before merge. Items 2-3 are worth addressing here too. Items 4-6 are optional polish.