Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning. From 0.6.2 the repository publishes one PyPI package, mcp-ltspice-qucs, with one version.

Unreleased

Nothing yet.

[0.6.2]: 2026-11-02

Added

• First release to PyPI, as one package: pip install mcp-ltspice-qucs. It installs the mcp-ltspice, mcp-qucs-s, and mcp-rf-analysis commands and the rf_mcp_common, mcp_ltspice, mcp_qucs_s, and mcp_rf_analysis modules. The wheel carries the license text.
• A tag-driven release workflow. release-validation gates it, and it publishes one sdist and one wheel through PyPI trusted publishing.

Changed

• The four workspace packages (rf-mcp-common, mcp-rf-analysis, mcp-ltspice, mcp-qucs-s) are merged into the one mcp-ltspice-qucs distribution. Import names and command names are unchanged. Every module now reports the mcp-ltspice-qucs version. In a checkout, uv sync replaces uv sync --all-packages, and the bundled client configs drop --package.
• Tool descriptions say what spec, sweep, and components actually want.

Fixed

• The sandbox probe checks the same system paths the real sandbox binds.

0.6.1: 2026-07-24

Fixed: dependency security audit

• v0.6.0's tag-triggered release-validation workflow ran for the first time and found 47 disclosed CVEs across 10 locked transitive dependencies: cryptography, idna, joserfc, the mcp SDK, pillow, pydantic-settings, pyjwt, pymdown-extensions, python-multipart, and urllib3. All ten are bumped to their patched versions within the existing dependency bounds. No pyproject.toml constraint changed, including the tight fastmcp>=3.4.4,<3.5 pin. pip-audit now reports zero known vulnerabilities. No API or behavior change; the full test suite was re-verified against the upgraded mcp SDK (1.27.0 → 1.28.1) with zero regressions.

0.6.0: 2026-07-24

Fixed: simulator run isolation and stale-artifact rejection

• Qucsator and Xyce harmonic-balance invocations now execute in unique per-run workspaces. A failed process can no longer satisfy the run with a .dat or .HB.FD.prn file left by an earlier invocation.
• Non-zero exits, empty output, malformed Qucs datasets, and malformed Xyce HB results fail before publication. Validated outputs are copied atomically; the canonical default result remains in its immutable run workspace so repeated or concurrent runs cannot overwrite it.
• Every run writes a JSON manifest with input/output hashes, backend identity and version, command, allowlisted environment keys, timestamps, streams, return code, status, and published artifacts. Qucs-S MCP results now expose the manifest path.

Fixed: S-parameter comparison on mismatched frequency grids

• compare_sparameters now compares the union of both measured grids inside their common frequency span instead of using exact frequency intersection. Shared endpoints can no longer hide every interior difference, partial overlaps never extrapolate, and disjoint ranges fail explicitly.
• Phase comparison now reports the shortest circular difference, so 179° versus -179° is 2° rather than 358°. Results include the overlap, interpolation policy, and original input ranges.

Fixed: single-source package and server versions

• Package __version__ attributes and FastMCP server metadata now derive from installed distribution metadata. The v0.5.0 packages no longer identify themselves as 0.1.0.
• Default shared envelopes use the installed rf-mcp-common version, and consistency tests cover module, server, and envelope surfaces.
• Updated stale architecture, envelope, Qucs implementation-status, and bug template text.

Added: shared simulation workspace contract

• rf-mcp-common now provides SimulationWorkspace, SHA-256 artifact hashing, a minimal subprocess environment, and backend-version probing for reuse by the LTspice/ngspice runners and future durable job API.

Fixed: documentation contract and capability boundaries

• Corrected current tool names, signatures, counts, examples, license text, package relationships, Qucs-S implementation status, and backend coverage.
• Documentation now distinguishes ideal analytical response, legacy catalog snapping with first-order parasitic metadata, and actual simulator validation. The bounded circuit workbench separately supports explicit model attachment and supported LTspice/Qucs/SPICE import; unsupported constructs are rejected. One-sweep S-parameter extraction remains explicitly limited to reciprocal symmetric two-ports.
• The generated tool catalog now reports primary and alias counts and supports --check; CI and workspace tests fail when it drifts from live registries.

Fixed: true two-excitation SPICE S-parameter extraction

• The public extract_sparameters tool now accepts a source .asc, validates the V1/Rs1/RL1 matched-port fixture, constructs a port-2 excitation, and runs both sweeps on the same LTspice or ngspice backend.
• S11/S21 and S12/S22 are measured independently and merged only when the frequency grids match. Missing traces, fixture-current inconsistency, ambiguous rewiring, and a missing second result fail instead of emitting a structurally valid matrix containing copied or zero terms.
• Results include the extraction method, fixture, source/raw/log paths, assumptions, validation residuals, backend, return codes, and isolated workspace. The legacy single-sweep helper now requires an explicit assume_reciprocal_symmetric=True opt-in.

Added: model-preserving passive realization

• Every selected passive now carries a ComponentModel record with provider, immutable source reference, model kind, pin map, operating-range fields, source path/subcircuit identity where applicable, reduction method, and SHA-256 checksum.
• User .lib registration accepts exactly one unambiguous two-pin subcircuit; missing/ambiguous pin maps block registration. Measured .s2p files retain their source checksum and explicitly record the first-order reduction used by analytical paths.
• simulate_realized_filter instantiates curated loss/SRF subcircuits or registered .lib models into isolated port-1/port-2 netlists, runs both sweeps on one backend, and returns a full Touchstone matrix with model manifests and extraction provenance. Changing a supplied library changes both its checksum and the simulated response.

Fixed: topology, transforms, and engineering summaries

• Sweep, optimization, Monte Carlo, comparison, and validation now carry explicit filter kind/topology, reject unsupported combinations, and include exact band edges and point targets in their evaluation grids.
• TDR now builds a uniform DC-anchored spectrum or an explicitly labelled band-pass response, preserves complex data, reports range/time resolution, and rejects transforms whose assumptions are not met.
• Delay extraction requires an explicit band. Equivalent-circuit fits report residuals, convergence, bounds, validity range, and identifiability warnings. Passband checks interpolate complex values at exact edges and reject partial frequency coverage.
• The schematic renderer accepts only package-generated LC fixtures it can reconstruct exactly. Arbitrary .asc content fails explicitly.

Added: confined execution, budgets, capabilities, and durable jobs

• Simulator input dependency trees are imported into immutable, checksum-addressed workspaces. Include/path/symlink escapes are rejected, ngspice has a no-network bubblewrap profile, and processes use timeout/cancellation tree cleanup.
• Expensive calls enforce component, point, iteration, core, trace, and combined-work limits. Large outputs stream to bounded JSONL artifacts.
• All servers expose cached capabilities:// resources and probe_backend tools that distinguish installed, launchable, validated, and unavailable states through tiny known-answer checks.
• mcp-ltspice exposes durable workspace/artifact/job primitives for import, parse/validate, submit, progress, cancel, retry, recovery, listing, and bounded artifact reads.

Fixed: MCP contracts and release engineering

• Error envelopes become structured protocol ToolError failures, and every tool carries safety annotations. Underscore names are canonical; 59 dotted compatibility aliases carry deprecation/removal metadata.
• Real in-memory and stdio initialize/list/call/error tests cover every server. Tool schemas, annotations, and metadata are hash-snapshotted.
• Codex .codex/config.toml and Claude Code .mcp.json project fixtures are executable and contract-tested.
• Dependencies have tested upper bounds. CI exercises locked and minimum-direct sets, builds wheels/source distributions, and installs wheels in isolation. Weekly self-hosted jobs cover Qucsator, Xyce, native and Wine LTspice, current Codex, and current Claude Code.
• Tag validation audits exported requirements, builds strict docs/contracts, emits an SPDX SBOM, and creates build-provenance attestations.

Added: auditable component provenance

• Catalog records distinguish generic technology models from orderable MPNs and include provider, source document/revision/retrieval time, license, checksum scope, package/tolerance/ratings, and test conditions.
• A provider protocol and checksum-addressed local-directory provider support deterministic user-model indexing and refresh. User filenames are never promoted to claimed manufacturer part numbers.

Added: bounded general circuit workbench

• Added versioned CircuitDocument 1.0 with explicit pin-to-net connectivity, ports, analyses, dependencies, provenance, drawing geometry, transformations, electrical fingerprints, and source-located unsupported diagnostics.
• Added bounded import/export for SPICE-family netlists, supported LTspice ASC symbols/wires, Qucs/Qucs-S schematics, and Qucsator netlists. Supported fixtures round-trip without connectivity/value loss; unknown electrical constructs block export and simulation.
• Added common ngspice, LTspice, Qucsator, and Xyce adapter contracts, normalized complex result datasets, readiness negotiation, and stable tolerance policies. A real asymmetric two-port agrees across ngspice and Qucsator for all four S-parameters.
• Added MCP workspace/import/parse/export flows with checksum-addressed artifact:// resources. The live catalog now contains 140 primary tools (76 LTspice, 29 Qucs-S, 35 RF analysis) plus 59 deprecated LT aliases.

Added: constrained component search and generic IR optimization

• Component search now applies hard value, package, availability/orderability, Q/SRF, tolerance, rating, bias, temperature, model-kind, and provider constraints. Unknown requested data fails instead of being treated as a match.
• Selected model records attach exact hashes, provenance, pin maps, validity, and source dependencies to the IR. SPICE/Xyce compilation stages checksum-verified sources and honors logical pin order; Qucsator rejects unsupported cross-dialect models instead of substituting ideal parts.
• Added deterministic bounded/discrete topology-preserving optimization with hard constraints, named SPICE temperature/parameter corners, yield estimation, cooperative progress/cancellation, exact model attestation, independent-backend validation, and JSON/Markdown design-change reports. Fixed models cannot be nominally retuned; explicit instance parameters and generic lumped approximations can be.

Fixed: validate_against_spice no longer emits a fabricated two-port

• The general two-sweep S-parameter fix above closed the fabrication in the public extract_sparameters/simulate_realized_filter tools, but validate_against_spice's output_spice_s2p still wrote the single-sweep network that mirrors S22 from S11 and S12 from S21: silently wrong for any asymmetric or non-reciprocal ladder. It now runs the same honest two-sweep extraction for the emitted file; the fast |S21| verdict still comes from the cheap single sweep. The tool description and output_spice_s2p field now disclose the two-sweep method and its V1+Rs1/RL1 fixture requirement.

Added: durable simulation jobs for mcp-qucs-s

• mcp-qucs-s gains the same simulate-through-IR job pipeline as mcp-ltspice: circuit_validate, simulation_submit, job_get, job_cancel, job_retry, job_list_artifacts, and artifact_read (22 → 29 tools). simulation_submit compiles the imported artifact, runs it, parses the result into a normalized ResultDataset, and validates it against the requested analysis; analysis.kind routes sparameters/ noise to Qucsator and harmonic_balance to Xyce. This wires the QucsatorAdapter/XyceAdapter backend adapters, which existed with full compile/run/parse coverage but no caller.
• Qucsator and Xyce have no verified OS sandbox profile, so these jobs run unsandboxed on the immutable per-run workspace snapshot the adapter's run creates; only submit trusted local inputs.

[0.5.0]: 2026-07-23

The release that emptied the issue tracker. Drivers exist for ngspice, LTspice, qucsator-RF, and Xyce; required CI validates ngspice while the other simulator integrations are capability-gated. The filter portfolio gains the full elliptic family and all five distributed microstrip topologies on an exact coupled-line TEM model; the coexistence chain runs closed-loop from victim bands to a catalog-snapped analytical design; mypy runs with zero disabled error codes; and CI covers Linux (3.11–3.13) plus native Windows discovery. Highlights below, in reverse-merge order.

CI: windows-latest job for native simulator discovery (#40)

A windows-latest job now runs the suite on real nt: it asserts _needs_wine() is False for a native .exe without the os_name test hook, exercises find_ltspice() walking the C:/Program Files candidates with no WINEPREFIX, and shakes out Windows Path semantics across the tests (LTspice itself is proprietary and not installed; simulator-gated tests skip cleanly). First Windows landmine fixed on the way: the render server-tool test used /proc/... as an "unwritable" path, which Windows would happily create: replaced with a file-as-parent-directory, which fails on every platform.

Chore: mypy ratchet complete: zero disabled error codes (#37)

The last three deferred type-completeness codes are re-enabled and their 54 accumulated errors fixed: no-any-return (numpy/JSON returns wrapped in float()/int() or precisely cast(...)), type-arg (bare dict/list/ndarray generics parameterized with the real types), and no-untyped-call (all 10 were genuinely third-party: schemdraw element constructors and matplotlib's PdfPages.savefig: handled by a targeted two-module [[tool.mypy.overrides]], not per-line ignores). The global disable_error_code list is now empty: type correctness AND completeness are both enforced. Also fixed test_runs_headless from #58 to delete DISPLAY and render, instead of asserting the environment happens to be headless.

Tests: EMC closed-form anchors + schematic_render coverage (#34, #36)

• predict_radiated_emissions_loop is now pinned to the textbook far-field small-loop expression (Ott's E = 131.6e-16·f²·A·I/r; the module's exact constant η₀·π/c² = 131.68e-16 agrees to 0.005 dB), with dimensional-consistency checks (double area → +6.02 dB, double distance → −6.02 dB) and a new near-field guard: measurements inside the radian sphere (r ≤ λ/2π) now raise instead of returning a silently wrong regulatory number, and the docstring states the regime.
• predict_conducted_emissions is pinned to a hand-computed CISPR 16 LISN divider at 1 MHz (|Z| = 49.53 Ω, 100 µA → 73.90 dBµV, the full complex arithmetic written out in the test docstring) plus current scaling.
• schematic_render.py (previously zero tests): ladder + elliptic designs rendered to SVG and PNG headless, every refdes and formatted value asserted present in the SVG text (catches silent mislabeling a size check would miss), the .asc round-trip, and both server tools' ok/error envelopes.

Added: synthesize_for_coex_target closed loop (#14, mcp-ltspice)

One MCP call now closes the loop the engineer used to orchestrate by hand: place transmission zeros on the victim-weighted harmonic centroids (#12) → synthesize the elliptic LPF and aim its traps (E24-snapped) → substitute real vendor parts → evaluate the realized ladder's rejection analytically → run the GNSS-aware coex matrix (#15) → escalate the order by 2 until the worst-case desense meets the target or max_order is reached (best-so-far returned with converged=False and the full iteration log either way).

Realism handled explicitly rather than hidden:

• The SRF spec targets the highest placed zero with the issue's 1.2 margin, degrading gracefully (1.2 → 1.0 → unchecked) with the margin used reported per iteration. A 14 nH 0402 with SRF ≥ 3.3 GHz does not exist, and a turn-key loop should say so, not die.
• Zeros default to the dominant 2H/3H PA products (4H/5H sit ≥ 50 dBc down pre-filter and their broad FCC landings would otherwise both outrank 2H/3H and push the SRF spec beyond what 0402 can build); the matrix still evaluates all harmonics through 5H.
• The fundamental sees the filter's in-band rejection, each harmonic's dBc is deepened by the rejection at n·f0, and GNSS victims get the realized filter's rejection at their frequency injected into the ΔC/N₀ model automatically.

This is the one place mcp-ltspice depends on mcp-rf-analysis (one-directional, no cycle). Tests drive the real pipeline end to end: real synthesis fits, real vendor tables, real matrix, no mocks.

Added: GNSS-specific desense in check_coex_matrix (#15, mcp-rf-analysis)

RX entries may now set victim_type: "gnss", switching that victim to the GNSS-specific model with the industry ΔC/N₀ (dB-Hz) metric instead of a generic blocking margin:

• Broadband PA noise (the dominant real mechanism, evaluated once per TX×RX pair): I₀ = PA_noise − filter_rejection_at_victim − antenna_iso; ΔC/N₀ = 10·log₁₀(1 + 10^((I₀−N₀)/10)) against N₀ = −174 + NF. PA noise comes from the victim's pa_broadband_noise_dbm_hz_at_offset, else the TX's broadband_noise_dbm_hz, else a documented −150 dBm/Hz default.
• In-band CW landings (fundamental/harmonics falling inside the GNSS band): the correlator spreads the tone over the code rate, I₀ = J − 10·log₁₀(chip_rate) (GPS C/A 1.023 Mcps default, Q = 1 assumed and stated in the per-entry assumptions).
• Entries carry mechanism (fundamental / harmonic_n / broadband_noise), delta_cn0_db_hz, i0_dbm_hz, the noise floor, and full assumptions; desense_margin_db for GNSS = a 1 dB C/N₀ budget minus ΔC/N₀ so mixed matrices stay worst-first sortable. Generic victims are unchanged. All model points are pinned by hand-computed tests (I₀ = N₀ → exactly 3.010 dB, etc.).

Added: place_zeros_for_coex (#12, mcp-rf-analysis)

Restricted-band-aware transmission-zero placement: for each harmonic landing [n·f_lo, n·f_hi], the optimal elliptic-filter zero is the severity-weighted centroid of the victim-band overlap intervals: TZ = Σ s·w·mid / Σ s·w. Not the landing's geometric centre (which is suboptimal whenever victims overlap the landing asymmetrically; the hand-pinned reference case moves the zero from 1800.0 to 1811.48 MHz). Victims come from the caller with optional severity weights, plus optional auto-loads from the package's GNSS and FCC-restricted tables (a 525 MHz TX's 3rd harmonic finds GPS L1 with no user input). The zero budget assigns highest-aggregate harmonics first, spare zeros fall back to landing centres (flagged), and the result carries trap-index hints matching mcp-ltspice's elliptic convention (lowest zero → trap 2) for direct composition with place_transmission_zero, a markdown rationale, plus unprotected_victims / victims_not_at_risk lists.

Implemented in mcp-rf-analysis (coex_zeros.py) rather than mcp-ltspice as issue #12 sketched: the band data and harmonic lookups live here, and mcp-ltspice carries no runtime dependency on this package: the trap-hint contract is the composition point instead.

Added: combline BPF (#27 complete, mcp-qucs-s)

synthesize_combline_bpf finishes the #27 distributed-filter set: N coupled lines shorted at the same end, each tuned by a lumped capacitor at the open end (C = Y_r·cot(θ0)/ω0; pure TEM combline genuinely needs the caps), resonator length θ0 default 45°. Couplings hit k = Δ/√(g_i·g_{i+1}) by inverting the exact pair transcendental ωC = (Y_r ± y_m)·cot(θ(ω)) (new combline_pair_split); the tap comes from the loaded-resonator slope parameter b = (Y_r/2)(cotθ0 + θ0·csc²θ0), which reduces to the interdigital formula at θ0 = 90°. segmented_array_sparams gains cap_loads (lumped C on any node), the graph netlister generalises to generate_coupled_array_netlist (ideal C elements at the loaded nodes; generate_interdigital_netlist kept as an alias), and the design self-reports achieved metrics like the interdigital tool.

Validation: qucsator graph netlist (TLIN + TLIN4P + C) matches the exact solver at numerical precision; on the 2 GHz / 10% / 0.5 dB N=3 design: 11.39 mm resonators (half the interdigital length), 1.137 pF loading caps, band centre 1.994 GHz, BW 11.7%, and the topology's selling point confirmed: first upper spurious at 3.75·f0 with the 1.5–3·f0 stopband below −39 dB (edge-coupled spurs at 2·f0).

Added: interdigital BPF on exact N-line TEM machinery (#27, mcp-qucs-s)

The N-conductor obstacle (a physical line cannot sit in two MCOUPLED elements) is bypassed exactly: a same-velocity TEM array's 2N-port Y-matrix is linear in the characteristic-admittance matrix, so a tridiagonal array decomposes into ordinary TLIN stubs plus floating TLIN4P connector lines: probe-verified to reproduce the CTLIN coupled section at 0.0000 mdB. New multiconductor module:

• segmented_array_sparams: exact S-parameters of stacked commensurate array segments with per-line short/open/port terminations and interior tap ports (validated against the closed-form coupled section at 1e-9 and against the qucsator graph netlist at 0.0001 mdB).
• interdigital_pair_k / mutual_for_k. The interdigitally-terminated pair's resonance split is closed form (cosθ = ±y_m/Y_r), giving the coupling coefficient without book tables.

synthesize_interdigital_bpf: N coupled λ/4 resonators, alternately shorted, tapped I/O. Couplings hit k = Δ/√(g_i·g_{i+1}) exactly on the TEM model; the tap angle comes from the shorted-λ/4 slope parameter θ_t = arcsin(√(π·Y_r/(4·G0·Qe))). Unrealizable Δ/Z_resonator combinations (stub admittance ≤ 0, Qe below the end-fed minimum) are rejected with explanations. The design self-reports its achieved response (achieved: band centre, −3 dB bandwidth, worst in-band return loss) computed on the exact model. The tapped-feed isolated-resonator approximation degrades ripple vs the prototype (measured: band centre 2.003 GHz on a 2 GHz/10% design, BW 11.8%, worst in-band S11 −5.9 dB) and an MCP consumer should see that number, not assume the spec. generate_interdigital_netlist emits the exact ideal-TEM graph netlist (stubs to gnd for shorts, TLIN4P per coupling, ports on tap nodes). Physical per-pair (W, S) via the existing Garg-Bahl inversion, flagged first-cut.

Added: hairpin microstrip BPF (#27, mcp-qucs-s)

synthesize_hairpin_bpf folds the edge-coupled filter into the Cristal-Frankel hairpin-line. The fold is a circuit-graph no-op apart from the U-bend connector: each half-wave resonator (the two arm halves the diagonal MCOUPLED chain already joins) gains the bend's electrical length θ_b, so every coupled section is shortened to 90° − θ_b/2: with one bend length (an MLIN at the mean arm width, default 3×W) this keeps every resonator at exactly 180° at f₀ in closed form. Zero bend degenerates to the edge-coupled design bit-for-bit.

Validated in real qucsator via the new generate_hairpin_netlist (MCOUPLED cascade + bend MLINs at the internal junctions): with a 2 mm bend (8.1° at 2 GHz) the compensated design centres at 1.976 GHz (−1.2%, matching the unfolded filter's −1.1%) with 9.0% bandwidth, while the uncompensated variant sags to 1.891 GHz (−5.5%). The differential test proves the compensation mechanism rather than just plausibility. Corner discontinuities and cross-arm self-coupling are documented as the field-solver residual.

Added: edge-coupled (parallel coupled-line) microstrip BPF (#27, mcp-qucs-s)

synthesize_coupled_line_bpf implements the coupled-section BPF core that hairpin / interdigital / combline all build on. Input is the LPF prototype g_coefficients vector (exactly what the lumped synthesis tools return); order N yields N+1 quarter-wave coupled sections via the J-inverter constants (Pozar §8.7 eq. 8.121). The electrical math is pinned to Pozar 4e Example 8.7: Z₀e/Z₀o = 70.61/39.24 and 56.64/44.77 Ω reproduced to 0.05 Ω.

Physical (W, S) realisation comes from a new coupled_microstrip module: quasi-static Garg-Bahl even/odd-mode capacitance analysis (single-line fringe reused from the Hammerstad-Jensen code, exact elliptic-integral gap term via scipy) plus numerical inversion. Unrealizable couplings (coupler-grade Ze/Zo ratios) are rejected with an explanation rather than returning a bound-clipped corner.

Validation, probe-verified against qucsator-RF 1.0.7:

• qucsator's ideal coupled line is CTLIN (vacuum velocity; node order line1-near / line1-far / line2-far / line2-near, established by terminating all four ports and reading through/coupled/isolated = 0.9573/0.2849/0.0140 against theory 0.9585/0.2856/0).
• The diagonal-connected section (others open) matches the closed-form section ABCD to 0.0000 mdB; the new coupled_line_section netlist element and coupled_section_sparams cascade agree with qucsator at numerical precision across the whole sweep.
• End-to-end through qucsator's real Kirschning MCOUPLED model (new generate_coupled_microstrip_netlist, diagonal chaining): on a Rogers-class laminate the Pozar design centres at 1.979 GHz (−1.1% vs design), −3 dB bandwidth 9.4% vs the 10% spec, −1.76 dB midband loss. On FR-4 the measured 4.05 dB midband loss matches the classical 4.343·Σg/(Δ·Qu) estimate: real dielectric loss, noted in the tests.

Added: stepped-impedance microstrip LPF (#27, first distributed topology, mcp-qucs-s)

synthesize_stepped_impedance_lpf maps a lumped LPF ladder (the same composition contract as lumped_to_distributed) to alternating high-Z / low-Z microstrip sections per Pozar §8.6: series L → high-Z section with βl = ω_c·L/Z_h, shunt C → low-Z section with βl = ω_c·C·Z_l, widths and lengths from the existing Hammerstad-Jensen synthesis. Sections exceeding the βl < 45° short-line approximation are flagged in notes. The math is pinned to Pozar 4e Example 8.6. The six electrical lengths (11.86°, 33.76°, 44.27°, 46.12°, 32.41°, 12.36°) and the section widths (11.27 mm / 0.430 mm vs the published 11.3 mm / 0.428 mm) reproduce the book design.

Validation plumbing, all probe-verified against qucsator-RF 1.0.7:

• generate_ladder_netlist gains a series_tline element (ideal line; qucsator's TLIN propagates at c, so the emitter encodes electrical length as L = θ/360 · c/f_ref: established with a λ/4-inverter probe netlist that returned |S11| = 0.6000 exactly).
• New generate_microstrip_ladder_netlist emits a shared SUBST card plus an MLIN cascade (Hammerstad model, Kirschning dispersion), so designs run through qucsator's real microstrip model.
• New tline_cascade_sparams computes the analytical ideal-line cascade; it matches the qucsator TLIN netlist to numerical precision (< 1 µdB).
• End-to-end on the Pozar example: analytical −3 dB at 2.418 GHz (−3.3% vs design. The known approximation shift), real-microstrip MLIN run −3 dB at 2.398 GHz (−4.1%) with 21 dB stopband at 4 GHz, matching the published response.

Remaining #27 topologies (hairpin, interdigital, combline) ship separately.

Added: elliptic band-pass / band-stop synthesis (#26, mcp-ltspice + mcp-qucs-s)

synthesize_lc_bpf and synthesize_lc_bsf now accept filter_type="elliptic" (odd order ≥ 3), completing the elliptic family started by the HPF in the previous release. The Pozar §8.5 band transforms act on each L and C of the LPF prototype separately, so every shunt series-LC trap becomes a four-element composite branch. A series-LC in series with a parallel-LC tank, to ground. That shorts at the two images of the prototype zero: ω = ω₀(√(b²+1) ± b) with b = ω_z·Δ/2 (BPF, notch pairs straddling the passband) or b = Δ/(2ω_z) (BSF, pairs inside the notch; the main-path tanks add a zero at ω₀ itself).

Supporting plumbing, since no existing element type could express that branch:

• mcp_ltspice.extract gains the shunt_composite_trap element ({L_s, C_s, L_p, C_p}), and components_dict_to_elements pairs the four-key refdes group {Lk_s, Ck_s, Lk, Ck} emitted by the synthesis.
• mcp_qucs_s.netlist.generate_ladder_netlist emits the branch as four uniquely-named parts (L3s/C3s/L3p/C3p).
• The v0.2.0 math-consistency invariant generalises: reported transmission_zeros_hz are the roots of u²·L_sC_sL_pC_p − u·(L_sC_s + L_pC_p + L_pC_s) + 1 = 0 computed from the physical component values, and tests pin them to the closed-form mapped zeros to 1e-9 relative.
• Validated against real qucsator-RF at orders 5/7/9 for both kinds: analytical ladder and simulator agree to within 0.02 mdB across a 100 MHz–10 GHz sweep.
• The synthesize_lc_hpf_filter / _bpf_filter / _bsf_filter MCP tool descriptions no longer claim elliptic is unsupported (stale since the HPF landed), and all three now return transmission_zeros_hz.

Elliptic BPF/BSF designs are always series_first (the prototype extraction is T-form only); a shunt_first request is coerced.

Added: register_user_vendor_dir (#11, mcp-ltspice)

Engineers routinely have third-party or measured .s2p files: Würth, AVX, TDK, distributor exports, in-house lab data. That live on no public URL and so could not be used as substitution candidates without hand-editing .asc files or forking the curated table.

register_user_vendor_dir scans a directory of .s2p (and, by filename, .lib) models and registers them under a namespace, after which substitute_real_components(inductor_vendor="<namespace>", ...) uses them like any curated series. Each part's kind, value and SRF are recovered from the measured reactance: via Z = 2·Z0·(1−S21)/S21 for the series-through fixture that is standard for two-terminal parts, not from Z11 (singular for a series element), and cross-checked against the filename shorthand (part_L_3n3.s2p → 3.3 nH). Value is averaged over the extracted L/C rather than the reactance, since a capacitor's reactance goes as 1/f and averaging it directly biases the result. A file whose measurement contradicts its name is indexed by the measurement with the conflict reported; per-file errors never abort the scan; re-registering a directory refreshes the index; and namespaces cannot shadow a curated catalogue.

lookup_part and lookup_part_with_srf_margin now filter a vendor table by component kind instead of classifying it from one sampled entry, so a user directory holding both inductors and capacitors resolves correctly.

Added: validate_against_spice (#16, mcp-ltspice)

The whole pipeline: synthesise, place zeros, substitute real vendor parts, optimise, Monte Carlo: can run end to end on the closed-form ladder without a single SPICE run. That is the right default for a fast inner loop, but it silently drops everything a real simulation captures (vendor .lib subcircuits, DC-bias/temperature behaviour, ground and supply non-idealities, measured vendor S-parameters via .include), so a reported yield could be an artefact of the lumped-element approximation.

validate_against_spice closes the loop: it runs the schematic through a real simulator, extracts the S-parameters, computes the analytical response for the same components, and returns a per-region |S21| divergence plus a verdict: agree, minor_disagreement, disagree, or spice_unavailable. Passband and stopband are separated by the analytical response itself, each judged against its own threshold. A passband disagreement adds an explicit warning that the analytical margin should not be trusted. With no simulator installed it returns the analytical response and spice_unavailable rather than failing, so it is safe to call unconditionally.

Added: elliptic high-pass synthesis (#26, partial: HPF)

synthesize_lc_hpf("elliptic", ...) now works; it previously raised NotImplementedError. An elliptic HPF is the frequency mirror of an elliptic LPF: the LPF→HPF map ω → ω_c²/ω inverts each prototype element into its dual. A series inductor becomes a series capacitor, and a shunt series-LC trap stays a shunt series-LC trap with L and C swapped, so a finite zero at ω_z moves to ω_c²/ω_z, landing in the lower stopband.

Validated at orders 5/7/9: equiripple passband above ω_c that mirrors the LPF to 0.05 dB, deep notches at the mirrored zeros, and the v0.2.0 math-consistency invariant (each reported zero equals 1/(2π√(L·C)) of its trap) holds to 1e-9. Cross-checked end-to-end against real qucsator: the analytical ladder and the simulated response agree to within 0.05 dB across the band.

Elliptic BPF and BSF remain unimplemented and #26 stays open for them. The bandpass/bandstop transforms turn each shunt trap into a four-element composite branch (a series-LC in series with a parallel-LC tank) with no existing element type or .asc netlister support. A materially larger change than the HPF mirror, and not worth rushing into a subtly-wrong filter.

Added: Qucs-S noise parameters (#25, mcp-qucs-s)

extract_noise_parameters was the last scaffolded tool in the suite. It now runs a real noise analysis and returns the four classical noise parameters per frequency: NF50, Fmin, Γopt and Rn. It also evaluates the noise figure at any source reflection coefficient via

F(Γs) = Fmin + (4·Rn/Z0)·|Γs − Γopt|² / ((1−|Γs|²)·|1+Γopt|²)

which is the number that actually matters once a real LNA input match replaces the ideal Γopt.

Anchored on an identity with no fitting in it: a passive network's noise figure equals its insertion loss, exactly, at T₀ = 290 K. Matched pads of 3, 6, 10 and 20 dB all report NF equal to their loss to within 1e-4 dB, with Fmin = NF and Γopt = 0.

Two Qucs conventions were established empirically rather than assumed, since either would produce plausible-but-wrong numbers:

• Rn is in absolute ohms, not normalised to Z₀. Recomputing F at Γs = 0 from Fmin/Rn/Γopt reproduces the reported NF50 to 4e-16 in ohms, and is off by 0.01 in noise factor if treated as normalised.

• Noise temperature is per-component, not an analysis setting.

  .SP … Temp="16.85" does not change device noise. Qucs defaults

  components to 26.85 °C, where a 10 dB pad correctly reads 10.13 dB

  1 + (L−1)·T/T₀. build_noise_netlist therefore applies the IEEE reference 16.85 °C to resistor lines by default, leaving any explicit Temp alone and never rewriting non-resistor device cards.

Added: harmonic balance via Xyce (#24, mcp-qucs-s)

run_harmonic_balance was scaffolded: it detected Xyce and then returned error("not yet implemented"). It now runs a real analysis.

• mcp_qucs_s.harmonic_balance builds the .HB netlist, runs Xyce, and parses the frequency-domain output into a single-sided spectrum. One tone gives harmonic distortion; two give the intermodulation products, IM3, OIP3 and IIP3.
• sweep_compression_point is a new tool: sweep drive level and locate P1dB by interpolating where gain falls 1 dB below its small-signal value.
• Missing Xyce still returns an error() envelope, now pointing at the source-build instructions rather than at binaries that do not exist for Debian/Ubuntu.

Validated against a closed form rather than a plausible-looking number: for a behavioural cubic V(out) = a1·V(in) + a3·V(in)³ the third-order intercept is exactly A = sqrt(4·a1/(3·a3)), and the computed IIP3 matches it to 0.001 dB. IM3 tracks the textbook 3:1 slope (30.00 dB per 10 dB of drive) and the intercept does not drift with drive level. A diode pair is the more realistic-looking test circuit but a poor validator. Its I-V is exponential, not cubic, so its products never follow a 3:1 slope.

Three Xyce behaviours found the hard way and now documented in docs/installation.md:

• .HB needs one NUMFREQ entry per tone; a single value with two tones aborts the run.
• ^ in a behavioural B-source expression makes the HB startup transient diverge; explicit multiplication works.
• .PRINT HB_FD emits a two-sided spectrum, so a positive-frequency bin folds to twice its printed magnitude (DC does not).

Fixed: documentation (docs/installation.md)

The Xyce section pointed at "pre-built debs" from Sandia. Those do not exist: Sandia's Linux binaries are RHEL 8 RPMs that explicitly do not run on Debian/Ubuntu, and the team no longer ships open-source binaries at all. Replaced with the verified source build (Trilinos 14.4 + Xyce 7.10, ~17 min on 8 cores), including the GCC-15 incompatibility that otherwise fails the Trilinos build at ~30%.

Fixed: three silent-correctness bugs (#31, #32, #35)

• .asc I/O assumed UTF-8/LF (#31). LTspice XVII writes UTF-16LE with a BOM, so read_components decoded mojibake, matched no SYMBOL line, and returned {} as though the schematic held no parts. Worse, update_component then rewrote the file as UTF-8/LF: silently re-encoding a user-authored schematic it had never successfully read. Encoding and line endings are now detected and preserved byte-for-byte outside the edited line; an undecodable file raises AscDecodeError instead of looking empty; and editing a refdes that is not present raises rather than rewriting the file unchanged.
• The ngspice netlister mis-netlisted every non-lowpass ladder (#32). It read each element's position from its refdes letter, so a highpass (series C, shunt L) was emitted as its dual and ngspice returned S-parameters for a circuit nobody designed. z0 was hardcoded to 50 Ω besides. New mcp_ltspice.asc_netlist netlists from schematic geometry using the same coordinate rules LTspice applies, which is correct for any topology: including schematics this package did not generate, and reads the port impedance off the terminating resistors. On a generated lowpass it reproduces LTspice's own netlist node for node. Unknown symbol kinds raise NotImplementedError naming the symbol instead of being skipped.
• cascade_networks extrapolated silently (#35). It found the common grid with numpy.intersect1d, which needs exact float equality; two instruments never produce bit-identical grids, so real inputs hit the fallback that resampled every other network onto network 1's grid: extrapolating past the end of their measured data with no warning. deembed_network raised in the same situation. Both now share one documented policy in common_frequency_grid: restrict to the overlap of the measured ranges so resampling is always interpolation, raise when the ranges are disjoint, and report any trimming through the envelope's warnings.

Fixed: Qucs-S could not be driven at all (mcp-qucs-s)

Running qucsator-RF 1.0.7 for the first time showed the Qucs-S backend was unreachable end to end, for three independent reasons.

• Nothing could produce simulator input. run_sp_analysis, export_touchstone and extract_noise_parameters all take a path to a file no code in the package could generate, so using them meant hand-authoring a schematic in the Qucs GUI first. New mcp_qucs_s.netlist generates qucsator netlists, and a new simulate_lc_ladder MCP tool goes design → netlist → simulation → Touchstone in one call.
• find_qucs_s() returned the GUI. Discovery tried qucs-s first, which is the Qt application, not the engine. Passing it -i netlist -o dat opens a window and blocks forever on a headless box. Discovery now looks for qucsator_rf / qucsator and never falls back to the GUI; when only the GUI is present the error says so and points at the missing --recurse-submodules submodule build.
• parse_qucs_dat could not read Qucs output. It called float() on every line, but Qucs writes complex values as +1.23e-10+j6.18e-08, so every real .dat raised ValueError. It also expected each S-parameter split into .r/.i sections, a format qucsator-RF does not emit. Both layouts now parse.

Elements are specified by explicit position (series_l vs shunt_l) rather than inferred from the refdes letter, so highpass and bandstop ladders netlist correctly. The failure mode behind #32 in the ngspice netlister, deliberately not repeated here.

Verified against real qucsator output: lowpass and highpass ladders track the closed-form Butterworth response to within 0.001 dB, and a shunt LC trap notches at its resonance. Regression cover includes a real qucsator-RF .dat committed as a fixture, so the parser cannot drift from the format the simulator actually writes.

sch_path is renamed to netlist_path throughout: qucsator consumes a netlist, not the GUI's .sch, and the old name sent users looking for the wrong file.

Fixed: generated schematics were electrically disconnected (mcp-ltspice)

Found by running the pipeline against real LTspice 26.0.2 (Wine) and ngspice 44.2 for the first time. Both simulator paths were broken end-to-end: run_simulation reported success and extract_sparameters then failed, so no user could get S-parameters out of either.

• generate_lpf_asc never emitted a single WIRE. The _wire helper was dead code, and LTspice determines connectivity purely from coordinates, so it netlisted every component onto its own NC_* nodes . No filter at all. The AC analysis returned No. Points: 0 and a 692-byte .raw, which the artifact-presence success policy accepted. Extraction then died with "No plots found in the RAW file". Symbols are now placed by solving for the anchor that puts their pins on the intended nodes, using pin offsets from the stock symbol library and the R90 convention (dx, dy) → (-dy, dx) verified against LTspice.
• S-parameter extraction required an I(Rs1) trace that ngspice never writes, so every ngspice run raised IndexError from spicelib. The port-1 current is now derived from V(p1) by Ohm's law, which is exact under the port convention the function already assumed. This also removes a sign-convention trap: SPICE orients a device's current by netlist node order, and LTspice emits Rs1 p1 N001 for our schematic, so its I(Rs1) runs out of port 1: pinning S11 at 0 dB for a perfectly matched filter. When the trace is present it is now used only as a consistency check, warning if magnitudes disagree.
• get_trace raises instead of returning None, so the "Missing required traces" ValueError was unreachable. It now reports the missing and available trace names.

Verified against both simulators: |S21| tracks the closed-form doubly-terminated Butterworth response within 0.15 dB from 0.1–4 GHz, with passband |S11| below −94 dB.

Regression cover for the class of bug, not just the instance: test_asc_connectivity.py reimplements LTspice's coordinate-based connectivity rule in pure Python, so CI catches a floating pin with no simulator installed; the end-to-end tests now compare against theory instead of asserting raw_path.is_file(), which a zero-point stub satisfied.

Fixed: LTspice first run blocks batch mode under Wine (mcp-ltspice)

Recent LTspice releases open a modal "Anonymously Share LTspice Usage Data" dialog the first time they run in a Wine prefix. It appears even under -b, so batch runs hang until the caller's timeout with an empty log and no .raw. The symptom points nowhere near a consent prompt. run_simulation now warns up front when LTspice.ini is absent from the prefix, and converts the resulting TimeoutExpired into a RuntimeError naming the dialog and giving both remedies.

Fixed: documentation (docs/installation.md)

• The Qucs-S build recipe cloned without --recurse-submodules, so qucsator-RF: the actual simulation engine: was never built; the result surfaces much later as "Qucs-S not installed". Also adds the missing flex, bison, gperf and dos2unix build deps (gperf fails at cmake time, dos2unix only at ~78% of the build) and corrects libqt6charts6-dev → qt6-charts-dev.
• Documents the LTspice first-run dialog and the wineboot -u / msiexec /qn silent-install path.

Fixed: simulator portability (mcp-ltspice)

Reported by @cr4i50n in #28 and #29 after trying to run the LTspice server on their own machine.

• LTspice under Wine exits 1 on a successful run, so the runner's strict return-code check rejected every good simulation. Success is now decided by whether the .raw artifact was produced; the return code is retained in RunResult and logged as a warning. Any stale .raw is deleted first, so the artifact cannot be a leftover. Same policy now applies to ngspice, matching mcp-qucs-s.
• Wine does not translate POSIX paths in argv, so -Run /home/… reached LTspice as a malformed drive spec. The .asc path is now converted with winepath -w.
• A .exe no longer forces Wine on native Windows, and find_ltspice() learned the Windows install locations and honours WINEPREFIX. Windows users previously always got "LTspice.exe found but Wine is not installed".
• A set-but-invalid LTSPICE_PATH silently fell through to ngspice; it now warns.
• New MCP_LTSPICE_SIMULATOR (ltspice | ngspice) pins simulator selection for ngspice-only deployments. Read by detect_simulator() and run_simulation(): deliberately not by find_ltspice(), which must keep reporting what is actually installed.

Fixed: correctness (all packages)

Surfaced by enabling mypy, which had never actually run (see below).

• ladder_sparams_from_components violated reciprocity. S12 was computed as 2·(a·d − b·c)/denom. That determinant is identically 1 for a cascade of series-Z / shunt-Y two-ports, but evaluating it numerically overflows on long ladders: a 9th-order bandstop produced inf at one bin, and the finite-value guard rewrote S12 to 0 while S21 stayed finite.
• series_lc_parallel leaked divide-by-zero / invalid-value RuntimeWarnings and its clamp could not fire at the anti-resonant bin: it tested the quotient (NaN there) instead of flooring the denominator before dividing, as the neighbouring trap branch does.
• RF_MCP_LOG_LEVEL=debug crashed every server at import. The raw env string went straight to Logger.setLevel, which only accepts an int or an exact uppercase name. Case is now normalized, numeric levels are accepted, and an unusable value warns and falls back to INFO instead of taking down the server.
• mcp-rf-analysis error envelopes were unattributable. _wrap reported func.__name__, but most tools pass a lambda, so failures read "<lambda> failed: …". They now name the tool that was called.
• Unrecognized band filters returned an empty list, so a misspelled region read as "there are no such bands". list_lte_bands, list_gnss_bands, and list_ism_bands now raise with the available values, matching list_5gnr_bands / list_halow_channels.
• list_spec_templates used .suffix on a Traversable, which only exists for filesystem-backed packages: AttributeError when zip-installed.
• richards.py / runner.py called .group() on a possibly-None regex match, crashing on any refdes without a digit.
• render_response accepted an unvalidated list as freq_range_hz, building a wrong-arity tuple from a 1- or 3-element list.
• network_from_dat skipped the validation its sibling performed, turning a partial Qucs-S .dat into a bare KeyError. Both loaders now share a checked path that also rejects ragged component arrays.

Fixed: CI and tooling

• CI had been red since 2026-05-13. tests/test_workspace_smoke.py failed both ruff format --check and ruff check, and never ran: the top-level tests/ directory was missing from testpaths.
• mypy had never checked a single line. It aborted with "Duplicate module named conftest" before reaching any source file, and CI hid that with || true. Both fixed; the suppression is gone and the workspace is clean under the enforced rule set.
• The docs site never deployed: GitHub Pages was not enabled, so the deploy job failed on every push while build passed.
• Fork PRs never got CI. Approval was required for all first-time contributors, so neither #28 nor #29 was ever validated.
• Simulator test gating matched too broadly. "ngspice" in item.keywords also matches test names and parametrize ids, so a case parametrized over the string "ngspice" was skipped as though it needed the binary. Now uses get_closest_marker.
• uv.lock was left stale by the v0.4.0 version bump.

Added: tests (458 → 533)

• mcp-rf-analysis had no server-layer tests at all; all 33 tools are now swept for the shared envelope contract, with the tool list discovered by introspection so new tools are covered automatically.
• mcp-qucs-s had no conftest.py, so its registered qucs marker gated nothing. Added, plus xyce. sparams.py. The only module that runs once Qucs-S is installed, and one that needs no binary to test: went from zero tests to full parse / round-trip / malformed-input coverage.
• check_coex_matrix's only test asserted nothing. It was named for a 2H collision, conceded in its own comment that the chosen pair does not collide, and checked only that two keys existed. Repointed at LTE B3 DL, which HaLow 2H = 1830 MHz genuinely falls into, plus a clean-pair negative case.
• Portability coverage for the runner: LTSPICE_PATH precedence, Wine path translation and its fallback, native-Windows invocation, the simulator pin, and both halves of the artifact-vs-returncode policy.

[0.4.0]: 2026-05-13

Changed

• License: Apache-2.0 → AGPL-3.0-or-later (all four workspace packages: mcp-ltspice, mcp-qucs-s, mcp-rf-analysis, rf-mcp-common). Aligns with the eng-mcp-suite toolkit-wide AGPL move (the AGPL closes the "wrap as a paid SaaS without contributing back" gap by extending copyleft to network use). Underlying tools: Qucs-S (GPL), LTspice (proprietary), scikit-rf (BSD): are runtime-invoked, not redistributed by these wrappers, so the wrapper's AGPL license is independent of theirs. See the LICENSE_SUMMARY for the toolkit-wide rationale.

(no changes yet: next stream picks up from the v0.3.0 roadmap)

[0.3.0]: 2026-05-06

Correctness: silent failures resolved (all packages)

A five-agent code review identified six silent-failure paths the v0.2.0 correctness-honesty pass missed. All are fixed; the user-visible impact:

• mcp-ltspice.extract_sparams_from_raw previously filled only the diagonal of the S-matrix, leaving S21 and S12 as zero. Every Touchstone file produced via the LTspice → ngspice → extract_sparameters flow claimed −∞ dB through the network, which silently failed any evaluate_filter_spec check. The fix recovers the full S column from a single AC sweep (S11 from (V₁ − Z₀·I_Rs1)/(2√Z₀·a₁), S21 from V₂/(√Z₀·a₁)) and fills column 2 by reciprocity + symmetry for the passive ladder filters this package synthesises. New flag assume_reciprocal_symmetric (default True) for the rare case where a user wants to merge two separate sweeps.
• mcp-rf-analysis.cispr_limit_at log-interpolated across the 5 MHz step in the CISPR 22/32 Class B QP table (56 dBµV → 60 dBµV), reporting ~58 dBµV at 12 MHz instead of the actual 60 dBµV. The same flaw was present in the FCC 15B mirror table. Both tables now encode the step with a paired (5e6, 56.0), (5.001e6, 60.0) entry, mirroring the pattern already used in radiated.py. Dead _FCC_15_109_B_AT_3M table removed.
• mcp-ltspice.runner never asserted proc.returncode on either the LTspice or ngspice path, so a non-zero exit (convergence failure, missing model lib) that left a stale .raw from a previous run was reported as success. The RunResult.log_path field was constructed but the file was never written for the LTspice path. Both fixed: returncode is now checked explicitly with stdout/stderr tail in the exception, and the log file is persisted on every run.
• mcp-ltspice.srf_check silently continued past components whose vendor lookup raised ValueError/KeyError, then returned severity='ok' even though half the BoM was unaudited. Components that can't be audited are now surfaced via a new unaudited field and a warning per refdes; severity rises to at least caution.
• mcp-ltspice.synthesis.lc_filter unpacked scipy.optimize.least_squares results without checking res.success, so an unconverged elliptic fit silently returned wrong L/C values. Now raises RuntimeError with the optimiser status when convergence fails.

Test pinning (31 new regression tests)

• New tests/test_hardening_v030.py (mcp-ltspice): contract tests for every fix above, including a monkeypatch-based round-trip of extract_sparams_from_raw that doesn't need a simulator install.
• New tests/test_microstrip_loss_pin.py (mcp-qucs-s): scaling-law tests for dielectric and conductor loss (linear in tan_d, linear in f for α_d, √f for α_c, 1/√σ for α_c) plus order-of-magnitude guardrails across FR-4 / Rogers / Duroid at 1–28 GHz. Catches missing- factor and wrong-conversion bugs without depending on textbook example values that vary across editions.
• HPF / BPF / BSF response now verified at orders 3, 5, 7, 9 (was only 3 and 5 before). BPF additionally checked at narrow (Δ ≈ 0.01) and wide (Δ ≈ 0.67) fractional bandwidth. The regimes where classical LPF→BPF formulas break down.
• Vendor parasitic substitution: lookup_part("coilcraft_0402hp", 4.7e-9) now pinned to return an SRF in the documented 4–8 GHz neighbourhood, not just "non-zero".
• CISPR 22/32 Class B step at 5 MHz pinned with explicit at 5.0 MHz / at 5.1 MHz / at 15 MHz regression tests.

Security

• SECURITY.md now spells out the untrusted-.asc attack surface explicitly. runner.py invokes external simulators on the supplied path with no sandboxing; a malicious .asc can include .SUBCKT, .lib, or .include directives the simulator will load. Document guidance: run inside a container/chroot if accepting external schematics.

Cleanup

• Empty packages/mcp-ltspice/src/mcp_ltspice/resources/ scaffolding tree (specs/ templates/ models/{coilcraft,johanson,murata,tdk}/, all empty) removed.
• tasks/ (private workflow scratch) untracked and gitignored.
• All four packages bumped from 0.1.0 (stale since the suite tagged v0.2.0 without bumping pyprojects) to 0.3.0 to align with the suite tag.

Test count

458 passed, 4 simulator-gated skips, 0 failures (was 427 / 4 / 0 in 0.2.0).

0.2.0: 2026-04-28

Filter synthesis: HPF / BPF / BSF (mcp-ltspice)

Closes the biggest user-visible gap in filter coverage. Three new top-level synthesis functions backed by classical Pozar §8.5 frequency transformations from the LPF prototype:

• synthesize_lc_hpf (filter.synthesize_lc_hpf): high-pass via series-L → series-C and shunt-C → shunt-L. Component count equals the LPF prototype's. -3 dB at the specified cutoff (Butterworth) or equiripple edge (Chebyshev).
• synthesize_lc_bpf (filter.synthesize_lc_bpf): band-pass via series-L → series-LC tank, shunt-C → shunt-LC tank. Component count doubles. f₀ = √(f_low · f_high), Δ = (f_high − f_low) / f₀; each LC pair resonates at f₀ exactly.
• synthesize_lc_bsf (filter.synthesize_lc_bsf): band-stop via series-L → series parallel-LC (anti-resonant), shunt-C → shunt series-LC (resonant). Used to notch a specific band (LO leakage, image rejection).

Currently supports Butterworth and Chebyshev I across all three. Elliptic HPF/BPF/BSF needs a separate transformation for finite transmission zeros and is on the roadmap.

Analytical S-parameter analysis is wired up for all four kinds (LPF, HPF, BPF, BSF). components_dict_to_elements extended with a kind parameter ("lowpass", "highpass", "bandpass", "bandstop"). Three new ABCD element types added to support BPF/BSF resonator topologies:

• series_lc_series: series-LC in main path (BPF series section). Z = sL + 1/(sC): dips at f₀.
• shunt_lc_parallel: parallel-LC to ground (BPF shunt section). Y = sC + 1/(sL): dips at f₀.
• series_lc_parallel: parallel-LC in main path (BSF series section). Z = sL/(s²LC+1): peaks at f₀.

The existing shunt_lc_trap kind (series-LC to ground; elliptic LPF trap) doubles as the BSF shunt section. Verified BPF response: -3 dB at band edges, deep stopband > 50 dB one decade out. Verified BSF response: > 60 dB notch at f₀, lossless passband one decade out.

Substrate preset library + microstrip loss (mcp-qucs-s)

• New mcp_qucs_s.substrates module with 16 curated presets covering FR-4 (4 thicknesses), Rogers RO4350B (3), Rogers RO4003C (2), RT/Duroid 5880 (2) + 6002, PTFE, Isola FR408HR (2), Taconic TLY5. Each carries documented εr / h_mm / t_um / tan_d from the manufacturer datasheet.
• New list_substrate_presets_tool MCP tool returns the full catalogue with descriptions.
• synthesize_microstrip_line and analyze_microstrip_tool now accept a preset name string (e.g. "Rogers4350B_0508") as the substrate argument in addition to the parameter dict. Saves engineers from re-typing {er, h_mm, t_um, tan_d} every call.
• analyze_microstrip now uses the substrate's tan_d (previously accepted but ignored) to compute conductor + dielectric attenuation in dB/mm, surfaced as alpha_d_db_per_mm, alpha_c_db_per_mm, alpha_total_db_per_mm. Pozar §3.8.1 dielectric formula plus skin-effect conductor loss (default σ = 5.8 × 10⁷ S/m for copper; override for gold / aluminium / measured plating). Verified against published values for FR-4 / Rogers / Duroid at 5 GHz.

Monte Carlo trace mode (mcp-ltspice)

monte_carlo_analysis gains a trace=True flag that emits a JSONL file (one record per trial: seed, sampled components, metrics, pass/fail status, failures list). Lets engineers do offline sensitivity / root-cause analysis of yield loss without re-running MC. Default path is mc_trace_<base_seed>.jsonl in cwd; override via trace_path.

monte_carlo_analysis(transmission_zeros=...) default changed from True to None: now auto-infers topology from the components dict (matches the components_dict_to_elements behaviour added earlier in this Unreleased stream). Old code passing True/False continues to work unchanged.

Tests

30 new tests across test_hpf_bpf_bsf_synthesis.py, test_substrate_presets.py, test_mc_trace.py. Total pass count 421 (+30 vs. prior 391 baseline), 0 regressions.

mcp-ltspice: power-supply EMC pre-compliance toolkit

Five new tools fill the gap between SMPS sizing (existing buck / boost / LDO) and a real product passing conducted-emissions:

• design_pi_output_filter (power.design_pi_output_filter): Pi-section LC output filter (C-L-C, 3rd-order LPF) sized for an attenuation target at a given frequency. Returns L, C_in, C_out, resonant frequency, achieved attenuation at f_target / f_sw, plus a damping-resistor recipe and BOM notes (inductor saturation, MLCC DC-bias derating).
• design_dm_input_filter (power.design_dm_input_filter): 2nd-order differential-mode LC input filter sized for a conducted-emissions target. Includes the Middlebrook stability check (|Z_out_filter| < |Z_in_converter| / safety_factor) so the filter doesn't destabilise the converter's loop, and surfaces a damping-branch recipe (R_d, C_d) per Erickson & Maksimović §10.4.
• predict_conducted_emissions (power.predict_conducted_emissions): harmonic decomposition of a trapezoidal switching waveform with two-sinc envelope (duty-cycle and edge-rate cutoffs), LISN-loaded prediction, CISPR 22 / 32 Class A or B limit overlay (QP or AVG detector). Returns per-harmonic frequency / emission / limit / margin arrays plus a worst-margin pass/fail summary.
• design_rc_snubber (power.design_rc_snubber): RC snubber for switch-node ringing. Inputs: parasitic loop inductance, switch C_oss, peak voltage, switching frequency. Returns R, C, ring frequency, achieved damping factor, and per-cycle dissipation. Standard recipe: C_snub = C_oss, R_snub = √(L_par/C_oss) × 2ζ.
• design_cm_choke (power.design_cm_choke): common-mode choke selection from a curated catalogue (Würth WE-CMB, TDK ZJYS / ACT, Murata DLW). Filters by DC current rating, target CM impedance at design frequency, and DM-leakage cap. Returns ranked candidate list plus the highest-margin pick.

All five register under both their flat names and power.* namespaced aliases. Tests in test_power_emc.py (31 cases) cover return-shape, math correctness (resonance / attenuation / dissipation scaling), CISPR class / detector relationships, and edge cases.

mcp-qucs-s: implementation status corrections

Earlier [Unreleased] notes claimed "Implemented all 4 simulator-driven tools with graceful degradation." This is corrected to the actual status:

• Implemented (4 closed-form synthesis tools, no Qucs-S install required):
• synthesize_microstrip_line: Hammerstad-Jensen W/L from Z₀
• analyze_microstrip_tool: Z₀ / ε_eff / wavelength from W
• synthesize_coupler: branch-line / rat-race / coupled-line / Lange (single-line approximation; coupled-line gap synthesis pending)
• lumped_to_distributed: Richards transformation + Kuroda identities
• Implemented (2 simulator-driven tools, require Qucs-S):
• run_sp_analysis: Qucs-S .dat parser + Touchstone exporter
• export_touchstone. Run + export .s2p in one call
• Scaffolded: return error("not yet implemented") envelope (Tier-6 roadmap):
• run_harmonic_balance: detects Xyce, but Xyce-netlist generation and harmonic-content parsing (IM3 / IIP3 / 1 dB compression) are pending
• extract_noise_parameters: detects Qucs-S, but noise-analysis dataset parser (Fmin / Γopt / Rn / NF50) is pending

Both scaffolded tools previously returned ok() envelopes with a "note: scaffolded" field: calling agents could mistake this for success. They now return error() envelopes with clear "not yet implemented" messages.

• The README's tool table previously listed synthesize_microstrip_filter (stepped-impedance / hairpin / interdigital filter realisations). The tool was never implemented. The row has been removed; filter synthesis in the distributed domain is on the Tier-6 roadmap.

Architecture hygiene (mcp-ltspice)

• Vendor catalogue: JOHANSON_L and TDK_MLG were COILCRAFT_0402HP.copy(): alias-only, not real data. Replaced with measured tables for the Johanson L-07W series (0402) and TDK MLK1005S series.
• compare_filter_orders now accepts srf_margin and max_value_drift_pct and passes them through to the inner substitute_real_components call. Previous behaviour ran the comparison with legacy semantics, so winners could differ from a final design that enabled SRF gating.
• Tool namespacing: every flat tool name now has a namespaced alias (filter.* / analog.* / power.* / digital.* / vendor.* / sim.* / coex.*). Old names continue to work but emit a one-shot DeprecationWarning per process pointing to the namespaced equivalent. Removal in a later major release.

Correctness fixes (mcp-ltspice)

• components_dict_to_elements(components, transmission_zeros=...). The transmission_zeros flag now defaults to None (auto-infer from the components dict) instead of False. Prior default silently produced wrong S-parameters when an elliptic ladder was passed without the flag set explicitly. Explicit False on elliptic-shape components now emits a RuntimeWarning.
• place_transmission_zero and trap_lc_for_freq. The preserve_ratio: bool parameter is replaced by mode: Literal["preserve_ratio", "hold_l", "hold_c"]. The old API with preserve_ratio=False and both l_existing and c_existing provided fell through none of the conditional branches and silently substituted L=1 nH. That bug is fixed. Legacy preserve_ratio calls continue to work via a deprecation shim.
• Elliptic synthesis (synthesize_lc_lpf filter_type="elliptic"). The reported transmission_zeros_hz field was inconsistent with the actual 1/(2π√(L_k C_k)) of the synthesised trap pairs at certain orders. Root cause: _fit_lc_to_prototype was an unconstrained least-squares fit over (L, C) pairs that drifted the L·C product away from the target trap resonance. Fix: each trap now has only L_trap as an optimisation variable; C_trap is computed as 1/(ω_zk² · L_trap) so the resonance is pinned exactly to the prototype's transmission zero. New regression test (test_elliptic_synth_consistency.py) covers orders 3 / 5 / 7 / 9 across multiple (fc, ripple, stopband) combinations and asserts the reported and achieved TZ frequencies agree to within 1 %.

Vendor catalogue (mcp-ltspice): Johanson L-07W and TDK MLK1005S replacements

The previous JOHANSON_L = COILCRAFT_0402HP.copy() and TDK_MLG = COILCRAFT_0402HP.copy() aliases were misleading. They suggested broader vendor coverage than the package actually had. Replaced with nominal datasheet-derived tables for:

• Johanson L-07W series (0402): 19 values from 1.0 nH to 39 nH, with SRFs ~10–15 % higher than Coilcraft 0402HP at equal inductance (wirewound construction). Extends the available value range upward.
• TDK MLK1005S series (0402): 19 values from 0.6 nH to 22 nH, extending the catalogue to sub-nH values that Coilcraft 0402HP doesn't carry.

These are first-order parasitic estimates suitable for synthesis-time sims; for design-final precision, fetch a real S-parameter file from the vendor (planned via the vendor.fetch_* tools on the Tier-3 roadmap).

Docs / branding

• Logo SVG (full and mark variants) under assets/
• mkdocs-material site at mkdocs.yml with auto-generated tool API docs via mkdocstrings
• Per-page docs: index, getting-started, architecture, installation, tool-catalog, examples, reference (envelope / Touchstone / E-series), contributing, changelog, security
• .github/workflows/docs.yml builds and deploys to GitHub Pages

Repo conventions

• Issue templates: bug, feature, question (config disables blank issues)
• Pull request template
• CODEOWNERS
• Dependabot config (weekly, grouped dev-tools / runtime / actions)
• Code of Conduct (Contributor Covenant 2.1)

0.1.0: 2026-04-22

Initial release of the four-package monorepo.

rf-mcp-common 0.1.0

• Envelope response model with ok() / error() constructors and Timer for runtime metadata.
• Hz-strict Touchstone I/O wrapping skrf.Network: read_touchstone, write_touchstone, network_to_touchstone, sparams_at.
• E6/E12/E24/E48/E96/E192 component snap helper with signed percent error.
• Frequency, dB, dBm conversion helpers.
• Structured JSON logger + tool_timer context manager.

mcp-ltspice 0.1.0

• 11 MCP tools covering the full synthesis-through-Monte-Carlo workflow:
• run_simulation: LTspice (Wine) or ngspice fallback
• extract_sparameters: .raw → .s2p via voltage/current method
• synthesize_lc_filter: Butterworth + Chebyshev I (closed-form g) + Elliptic (scipy ellipap + weighted least-squares LC fit)
• place_transmission_zero: Move shunt-LC notch to a target frequency, preserve L/C ratio, snap to E24/E96
• find_transmission_zeros: peak-detect notches in S21
• substitute_real_components: vendor parasitic tables (Coilcraft 0402HP / 0603CS, Murata GJM C0G, Johanson L, TDK MLG)
• evaluate_filter_spec: pass/fail per criterion with margin in dB
• optimize_filter: Nelder-Mead, loss = sum of negative margins, E24/E96 snap
• monte_carlo_analysis: joblib parallel Gaussian sampling, yield% + per-metric histograms
• stability_check: Rollett K, |Δ|, Edwards-Sinsky μ
• render_response: Bode PNG with frequency marker lines
• Analytical ABCD-chain S-parameter computation in extract.py (no simulator required for synthesis sanity checks)
• LTspice .asc generator + spicelib-based reader/modifier

mcp-rf-analysis 0.1.0

• 19 MCP tools, all simulator-agnostic (Touchstone in/out)
• Bundled regulatory databases:
• LTE bands 1-71 (3GPP TS 36.101)
• 5G NR FR1 + FR2 (3GPP TS 38.101)
• GNSS (GPS L1/L2/L5, GLONASS L1/L2, Galileo E1/E5/E6, BeiDou B1/B2/B3)
• ISM allocations by ITU region
• 802.11ah HaLow channels per region (US, EU, JP, KR, CN, SG, AU/NZ, IN)
• FCC §15.247 + §15.205 restricted bands
• ETSI EN 300 220 / 300 328 / 301 893 limits
• 3GPP UE spurious + GNSS protection
• Spec template library (FCC Part 15.247, HaLow US LPF)
• Multi-radio coex matrix with predicted desense per aggressor/victim pair
• Link budget tools (Friis, log-distance, antenna isolation, desense)

mcp-qucs-s 0.1.0

• Initial scaffold + status tool. Full implementation lands in Unreleased above.

Examples

• examples/halow_lpf/: full end-to-end design of a 9th-order elliptic LPF for 802.11ah HaLow + multi-radio coex. Passes all 8 spec criteria with 86.6% Monte Carlo yield at 2% component tolerance.