signal-integrity-review

Read runtime access before choosing between explicit-path MCP, a live bridge session, and native/headless CLI. Their availability is separate.

Signal-integrity review

RAG: engineering memory by default — shared workflow. Use the indexed transmission-line, grounding and EMC material to justify the model, assumptions and return-path checklist; solver output does not replace that basis.

Keep analytical equations, geometry heuristics, and external solver output in separate evidence classes. No result is fabrication sign-off. Use public tools/list for exact callable names and get_capabilities for document/configured feature and adapter availability.

Scope selection

• microstrip: supported single-ended analytical model for an outer-layer trace.
• differential_microstrip: supported coupled analytical model with explicit gap.
• symmetric_stripline: supported single-ended analytical estimate within its published range.
• Differential stripline: unavailable; return blocked_by_capability.
• run_openems_stripline_analysis: a registered typed external adapter, not a bundled solver.
• run_ngspice_simulation: a registered batch adapter for a caller-supplied netlist, not a PCB geometry extractor.

Workflow

1. Call diptrace_status, get_capabilities, and get_document_info; freeze the PCB SHA-256.
2. Read get_stackup and get_route_details. Require explicit net, layer, dielectric height/permittivity, copper width/thickness, reference conductor, and target/tolerance when a pass/fail judgment is requested.
3. Use calculate_impedance for standalone inputs, analyze_stackup_for_impedance for complete outer-layer microstrip stackups, and validate_impedance_constraints or analyze_controlled_impedance for named routed nets.
4. Use analyze_return_path only as a caller-radius geometry heuristic. Disclose boundary-only pour geometry, layer-transition ambiguity, and confidence limits. If native refill evidence is needed, use the PCB headless profile on a copy and retain its exported geometry and verdict; do not call all native work unavailable based on the MCP tool list. Even refilled geometry does not turn this heuristic into a field solution.
5. Invoke ngspice/openEMS only when runtime discovery says the configured adapter is available. For openEMS retain request SHA-256, solver version, convergence, result SHA-256, and resources. For ngspice retain the netlist SHA-256, job status, return code/log summary, and resources; its current API exposes no executable version, convergence field, or result SHA-256, so record those fields as unavailable instead of inventing them.
6. Emit ../shared/result.schema.json. Put impedance, effective permittivity, length/skew, frequency, and return-path coverage values in measurements, each with a unit, evidence class, and evidence IDs; do not hide numeric evidence in prose.

Published and implementation boundaries

• Coupled microstrip validity requires 0.1 <= width/height <= 10 and gap/height >= 0.01; outside it, report unavailable/low-confidence metadata rather than forcing the permittivity-order invariant.
• Single-ended microstrip effective-permittivity validity uses 0.01 <= width/height <= 100.
• Symmetric stripline uses width/free_height < 0.35 and thickness/free_height < 0.25.
• Finite copper thickness is not corrected in the coupled branch. Do not use its wide-gap decoupling asymptote as a thickness validation.
• Every distance is millimetres; impedance is ohms.

These bounds and citations are returned by impedance.py. Adapter behavior is implemented in external_adapters.py and mapped truthfully in ../capability-map.json.

Use analytical, heuristic, and external_solver exactly as appropriate; never collapse them into one confidence label.