MCP: Slice summary reads with document.node depth and childLimit

[fahrim]

Sliced summary reads for large containers.

vectorcraft://object/{id} returns the whole subtree: on a 2,000-object document a group read is 1.5 MB and a layer read 1.8 MB. The cost left after #328 is the reply size itself. This adds depth and childLimit to document.node {summary: true} so agents can read containers in slices.

Design (no new commands, no URI changes — slicing lives at the command layer, where the MCP spec leaves it; resources/read takes only a URI):

• depth: child levels included (None: all, the old behavior). 0 is the node alone.
• childLimit: children shown per node, top of the stack first (None: all).
• Truncation is never silent: a level that shows fewer children than it has reports childCount. With no truncation the output is byte-identical to before (no childCount appears).
• The full (non-summary) object JSON is never truncated — exactness is its contract.
• A present-but-invalid value (depth: -1, childLimit: "many") is rejected with -32602 rather than silently reinterpreted.
• The object template keeps its full-depth default; its description and docs/mcp.md point large-container readers at run_command document.node {id, summary, depth, childLimit}.

Measured over stdio on a 2,000-rectangle group, release build:

read	time	bytes
full (unchanged default)	30.1 ms	1,516,377
depth: 0	0.43 ms	203 (+ childCount: 2000)
childLimit: 10	0.6 ms	7,813 (~194x smaller)

Quality gates: the default path is byte-identical to current main (engine test equates summary: true with the document.inspect subtree; live A/B of leaf/group reads matches hash-for-hash), and a new engine test covers nesting (depth: 1 shows children while the nested group reports childCount), limits, and invalid values.

cargo xtask ci: all 7 steps pass.

[bflatastic]

Thanks @fahrim, merged! This is a clean follow-up to #328. Keeping the slicing in the command layer, and reporting childCount only when a level is cut, is the right design, and your measurements made it easy to review. I rebased it on main and added one commit on top:

• Rejected: depth/childLimit without summary: true is now an error. Before, a request that looked sliced quietly returned the whole full-JSON subtree.
• Docs: docs/mcp.md has a paragraph on the slice parameters, childCount and the validity rules.
• Tests: fractional and oversized values, limits past the end of the tree (byte-identical to the full summary), and a sliced group read through run_command in the CLI stdio sweep.

Checked end to end over MCP stdio on a 2,000-child group: childLimit: 10 cut the reply from 810 KB to 4.5 KB, with childCount: 2000 reported. A small note on the PR body: through run_command, invalid values come back as a tool result with isError: true rather than a -32602 protocol error. That's the right behaviour for tool calls.