fix: critical bugs in exec pipeline, map cache, and MCP install

[cursor]

Summary

Weekly critical bug investigation found and fixed 4 high-severity issues in recent code paths.

1. UTF-8 panic in ctx exec filters (crash)

• Impact: ctx exec git blame (and aws/docker/kubectl filters with truncate_lines_at) could panic when a line contained multi-byte UTF-8 and the byte cut landed inside a character.
• Root cause: truncate_lines sliced at byte index without checking char boundaries.
• Fix: Truncate at the nearest valid UTF-8 boundary before slicing.

2. Stale ctx map cache on dirty repos (silent wrong context)

• Impact: After editing uncommitted files, ctx map could return a stale cached map from before the edits — agents receive outdated signatures/structure.
• Root cause: git_commit_key for dirty repos hashed only the directory path (first 32 bytes), not working-tree state.
• Fix: Include git status --porcelain output in the dirty cache key.

3. MCP config overwrite on install (data loss)

• Impact: ctx install --agent claude-desktop (or claude-code) overwrote an existing mcpServers.ctx entry if it lacked the _installer: "ctx" marker — permanent loss of user config.
• Root cause: insert_mcp_server only no-op'd for ctx-marked entries.
• Fix: Skip insertion when any same-named entry already exists.

4. Failed commands masked as success in ctx exec (agent misinformation)

• Impact: git pull auth/network failures could return stdout "ok" while exit code was non-zero — agents may treat the operation as successful.
• Root cause: on_empty fallback applied regardless of subprocess exit code.
• Fix: Only apply on_empty when exit_code == 0.

Validation

• Added unit tests for each fix
• cargo test — all passing
• cargo clippy --all-targets --all-features -- -D warnings — clean