`search`: add `--json` flag for machine-readable output

[raphasouthall]

`neurostack search "query"` currently prints results as formatted text. This is useful for humans but hard to pipe into other tools.

Add a `--json` flag that outputs results as a JSON array instead. Each element should include at minimum: `path`, `title`, `score`, and `snippet`.

Example:

```bash
neurostack search "memory consolidation" --json
```

```json
[
{
"path": "neuroscience/memory.md",
"title": "Memory Consolidation",
"score": 0.91,
"snippet": "Consolidation transfers information from hippocampus to neocortex..."
}
]
```

Where to look:

• CLI entry point: `src/neurostack/cli.py`
• Search logic: `src/neurostack/search.py`
• The `search` subcommand is defined in `cli.py` using Click or similar — look for the `@cli.command()` decorator for `search`

Acceptance criteria:

• `neurostack search "query" --json` outputs valid JSON to stdout
• Each result object includes `path`, `title`, `score`, and `snippet`
• Default (non-JSON) output is unchanged
• `--json` is documented in `--help`
• At least one test covers the JSON output path

[raphasouthall]

Verified in production on v1.27.0 (LXC 122 install, commit c79bc4e) against a 536-note / 6561-chunk vault.

$ neurostack --json search "test" --top-k 1
[
  {
    "path": "work/.../infra-dr-test.md",
    "title": "DR Test Infrastructure (infra-dr-test)",
    "section": "# DR Test Infrastructure > ## Pipelines (4) > ### 1. ...",
    "score": 0.5828,
    "snippet": "### 1. ...",
    "summary": "..."
  }
]

Acceptance criteria check:

• Valid JSON on stdout
• Required fields present (path, title, score, snippet) — plus bonus section and summary
• Default (non-JSON) output unchanged — pretty-printed result with headers and snippet, same as before
• --json is documented in --help (--json Output results as JSON)

Flag attaches at the top-level parser (neurostack --json search ...) rather than the subcommand (search --json), but functionally equivalent and works across the 51-invocation retrieval benchmark I just ran against this vault (17 queries × 3 modes, all parsed cleanly).

Test coverage for the JSON output path is the only criterion I can't confirm externally — flagging that in case it's still pending.