feat(mcp): host-neutral assembly joint create/update

[jeffglousher]

Summary

• Expose host-neutral assembly joints as MCP tools (Direct → host::handle), the slice deferred on feat(mcp): host-neutral assembly component tools #62.
• Tools: assembly_create_joint (CreateJointRequestDto) and assembly_update_joint (UpdateJointRequestDto). Query stays assembly_document / assembly_solution (already on main).
• Schemas match the landed host DTOs (kind, topology-backed connectors, limits, advanced, optional grounded body/occurrence).
• New headless golden assembly_joint_create_update_query_roundtrip: two extruded boxes → revolute joint → assembly_document query → rename/limits update → solution still non-empty. No cad_attach.

Why

#62 shipped component/occurrence/pose/grounded and left joints for a follow-up once schemas stayed small. Host already has stable CreateJointRequestDto / UpdateJointRequestDto and assembly_create_joint / assembly_update_joint engine methods.

What is NOT done

• Does not stack on feat(mcp): UI-owned inbox apply (#11 slice) #60 inbox apply or feat(mcp): assembly mutates in shared cad_submit map #63 submit-map.
• No motion studies, contact sets, delete-joint, or mechanism-drag tools.
• No live UI / attach writeback. Not a NIST/print-kit dump.

Test plan

• cargo test --manifest-path mcp-server/Cargo.toml --bin nbcad-mcp -- --test-threads=1
  • 38 passed, including assembly_joint_create_update_query_roundtrip
  • registry / focus-pack matrix now 119 modeling tools (assembly pack 10)
• rustfmt on the two MCP files

Refs

Follow-up to #62. Does not touch #60/#63.

[jeffglousher]

Rebased onto latest main (a0015f0, #67). Clean rebase — no conflicts (mcp-server only; independent of #60). New tip: eaf9f0f. mcp-server 38 green including assembly_joint_create_update_query_roundtrip. Not merging.

[jackControls]

Thanks for expanding the assembly MCP surface. I rechecked the current head (eaf9f0f). One schema/round-trip compatibility issue remains:

The advertised joint tool schemas reject valid joint JSON produced by the server itself. source_surface_frame and the optional joint limit fields are declared as object-only in the create/update schemas, while the Rust DTOs serialize absent Option values as explicit null. A strict MCP client that validates tool input against tools/list therefore cannot pass an unmodified assembly_joint_query result back into create/update. The existing round-trip test calls CadServer::call_tool directly, so it bypasses advertised-schema validation and does not catch this.

Please allow both object and null for these nullable fields (including the advanced joint limit variants) and add a round-trip test that validates the queried payload against the advertised input schema before submitting it.

[jeffglousher]

Addressed the schema/round-trip review on 3a071a6.

Fix: advertised assembly_create_joint / assembly_update_joint schemas now accept object-or-null for the fields that serialized DTOs emit as null:

• source_surface_frame
• primary limits: limits, angle_limits, linear_limits
• advanced limits: secondary_angle_limits, tertiary_angle_limits, secondary_linear_limits

Test: assembly_joint_query_validates_against_advertised_update_schema queries a created joint, validates { joint } against the tools/list assembly_update_joint inputSchema, then submits via tools/call (not only CadServer::call_tool).

mcp-server: rustfmt clean; clippy -D warnings on the new code (pre-existing chunks_exact_to_as_chunks / manual_contains elsewhere unchanged). cargo test 39 green.

Not merging.

[jackControls]

Thanks for the update. The advertised schemas now accept null for every optional joint field emitted by the DTOs, including connector source frames and primary/advanced limits. The regression test validates a queried joint against the tools/list schema before updating it, and the focused test plus MCP CI pass. Approved.

[jeffglousher]

Rebased onto latest upstream/main (0fecc57, #60 merged). Kept object-or-null advertised joint fields (source_surface_frame + primary/advanced limits) and assembly_joint_query_validates_against_advertised_update_schema.

Conflicts were tool-count (MODELING_TOOL_COUNT 119 + 22 helpers) and the shared-mutate classifier test — both kept. New tip: b2d23e6.

Local: assembly_joint_create_update_query_roundtrip, assembly_joint_query_validates_against_advertised_update_schema, tool_registry_is_granular_and_protocol_lists_revolve, every_shared_mutate_is_accepted_by_cad_submit_classifier green. Not merging — rebase dismissed the prior approval.

[jeffglousher]

Closing as superseded by #69 (df875a0 / tip fad8dca).

Main already has this PR’s surface: assembly_create_joint / assembly_update_joint ToolSpecs, object-or-null advertised schemas (source_surface_frame + primary/advanced limits), mutate-map registration (the 08a94235 hunk), Assembly disclosure tags, and both goldens (assembly_joint_create_update_query_roundtrip, assembly_joint_query_validates_against_advertised_update_schema).

Leftover vs this head is only #69 follow-on (stronger replace-all update description + extra adversarial coverage). Rebasing onto current main would re-apply already-landed hunks on files #69 expanded — leftover, not unique work.