test: add 247 tests across workspace, boost coverage to 79%

Phase 2 coverage improvement (71.95% -> 79.04%):

Integration tests (48 new):
- Agent tool loop: output, error markers, max iterations
- Tool errors: blocked, confirmation, sandbox, IO error
- Streaming: chunks, empty response, redaction
- Memory: persistence, summarization, history
- Commands: /skills, /skill, /feedback
- Builder: security, skill reload, shutdown

LLM crate (55 new):
- claude.rs: request serialization, SSE parsing, split_messages
- ollama.rs: host parsing, message conversion
- orchestrator.rs: task classification, routing, fallback
- any.rs: enum dispatch for all methods (from 0%)

MCP + Skills + Memory (89 new):
- executor: instruction parsing, error handling
- manager: transport clone/debug, server lifecycle
- registry: tool hashing, UUID generation, search
- qdrant_matcher: construction, content hashing
- watcher: file change detection (from 0%)
- matcher: ranking, backend delegation
- semantic.rs: summarization, message count
- config.rs: all sections, 20+ env overrides

Security coverage:
- shell.rs: 96.63% -> 98.76%
- vault.rs: Secret deserialization added

Total: 586 -> 833 tests, all passing

Author: bug-ops

crates/zeph-core/Cargo.toml

@@ -29,6 +29,7 @@ zeph-skills = { workspace = true, features = ["qdrant"] }
 zeph-tools.workspace = true
 
 [dev-dependencies]
+serde_json.workspace = true
 serial_test.workspace = true
 tempfile.workspace = true