Risk register (research-preview honesty ledger)

The synthesis and mathematics pages document their own caveats in place. This page consolidates the load-bearing ones, triaged by severity and status, so the readiness of any given path is legible at a glance. It is the gate list for turning a path into a service or a headline claim.

Severity = how much harm the risk does if a user relies on the path today. Status: open (unaddressed), measure-pending (correct but the C# number is not yet harness-measured), mitigated (addressed; how noted).

High — gate a public service / a beats-benchmark claim

ID	Risk	Where	Status
H1	3D packing density numbers were UNMEASURED in C#. The 33.8 / 35.7 / 38.7 / 53.3 / 49.3 % figures are Python pybullet+VHACD or Rhino-live-mesh, not the net48 Core engines. CoACD != VHACD.	SYNTHESIS_3D Open risks	partially addressed 2026-07-06 — the measurement is now Rhino-free and shipped: GreedyMeshHeightmapPacker + signed-tetra MeshVolume + ObjMeshReader measure honest density on the real ETH subset headless (16/16 placed, rho_honest 0.073 vs rho_bbox 0.195, 2.68x bbox over-report; test ETH C# honest 3D density headless bench). Remaining: container-tuned per-packer comparison at verified-zero interpenetration (issue #13).
H2	The pipeline is only ~74% headless. Core files import RhinoCommon: geology/discontinuity, fabrication (DXF, cut-plan, wire-saw), and much of masonry. A "headless server" would need Rhino.Compute for those, not a plain dotnet service.	grep using Rhino in Core	in progress — 77 of 300 Core files remain Rhino-bound (was 78). The 2D nester (ContactNfpHoleNester) is clean and measured, and the client-side nest demo proves the pattern (net9 WASM, no backend). KinematicAnalysis (geology feasibility) was made Rhino-free 2026-07-06 by inlining a cross product. Remaining paths tracked in issue #14.
H3	The production RBE solver is the weak path. The closed-form QP is skipped whenever friction inequalities are present, so the with-friction (production) case always takes the Dykstra alternating-projection route, which the code comments admit has convergence trouble on the 6-DOF family (500-iter serial tail). A stability verdict served blind could be wrong on hard assemblies.	slm_cards/masonry-equilibrium-cra; math/MASONRY	fail-loud since 2026-07-07 — the checker now runs an INDEPENDENT equality-residual audit after every solve (relative \|\|Aeq f − b\|\|inf gate 1e-3, arithmetic independent of all solver lanes) and refuses a stable verdict on an unconverged iterate; a wrong-stable can no longer ship silently. Also verified: production routes through native OSQP / warm-started ADMM (non-Optimal already fail-loud), not the Dykstra path the audit card described. Remaining (lower urgency): solver back-end upgrade via the IConvexQpSolver seam.

Medium — address before scaling, safe for interactive use with the caveat

ID	Risk	Where	Status
M1	Masonry numeric hygiene. Four uncoordinated absolute epsilons; no assembly recentering (far-from-origin inputs degrade the Cholesky of Aeq Aeq^T); undeclared units (a mm model silently mixes mm geometry with SI gravity, weights off by 1e9).	slm_cards/masonry-equilibrium-cra numeric stress	partially addressed 2026-07-07 — the worst precision hole is fixed: BlockCenterOfMass now integrates the signed-tetra volume/centroid relative to the first vertex (translation-invariant recentering), so far-from-origin models no longer lose ~10 digits (test: COM at 1e6 offset matches near-origin to 1e-6). Remaining: Ruiz equilibration before the Cholesky, declared units at the boundary, scale-relative epsilons.
M2	Legacy RBE sign convention still live. RbeQpFormulation.Build yields f_n = -mg against lowerBounds = 0, making real assemblies infeasible; only BuildPhysicsCorrected is correct. Two conventions coexist.	RbeQpFormulation.cs:174-212	defused 2026-07-07 — RbeQpFormulation.Build is now [Obsolete] with the failure mode spelled out in the message, so accidental use warns at compile time; production uses BuildPhysicsCorrected (verified: no production callers of the legacy path). The sign-pinning tests suppress deliberately. Full deletion remains as cleanup once those pins are reconciled.
M3	Physics-settle determinism is conditional. BulletSettleService is reproducible only single-thread with a fixed step count; a service under concurrent load may not be.	packing/EQUATIONS 2.6	open — keep settle opt-in; mark determinism P1.
M4	Type-conversion seam. Core Vec3/MeshTriangle vs the MeshSnapshot consumed by CoACD/Bullet; world transforms for verify+settle must survive the seam (world = R*local with R the transpose of the Bullet basis).	SYNTHESIS_3D Open risks	covered 2026-07-07 — headless convention-pinning tests: the Bullet row-vector basis packing (producer) composed with the NboSettle row-major application (consumer) round-trips a known rotation + rigid motion, and MeshSnapshot round-trips vertex/triangle arrays exactly. Either end flipping the transpose now fails the battery.
M5	Voxel seed is below benchmark alone. ~32-34 % (Python) vs the 33.8 % skyline; the density win depends on the settle stage. Not a defect, an expectation.	SYNTHESIS_3D	open — state the seed floor as parity/below in any claim.
M6	Masonry model is no-tension only (dry-stone). The stability checker is strictly frictional Heyman/CRA: FrictionConeBuilder cone anchored at f_n >= 0, tension penalized to zero, NO cohesion / tensile-bond / surcharge term. It cannot represent MORTARED masonry (lime mortar, cementitious rubble-core, Roman concrete) — a whole structural class. A mortared structure analysed with it gets a conservative lower bound (safe: stable-without-mortar implies stable), which can read UNSTABLE for a real structure that stands on its mortar bond (e.g. a steep Maya corbel vault).	FrictionConeBuilder; RbeQpFormulation	open, boundary documented 2026-07-07 — safe to use as a conservative bound if the caveat is stated. Fix = Mohr-Coulomb cohesion extension (\|f_t\| <= c*A + mu*f_n, tension cutoff f_n >= -sigma_t*A); Opus-tier, HITL + re-validate battery + re-prove conservativeness + calibrate c/sigma_t. Tracked in the private development map.
M7	Async scan components can emit stale results. AsyncScanComponent (base of Load Cloud, Read LAS Cloud, Estimate Cloud Normals, Scan Reconstruct, and Cloud ICP once fix/cloud-icp-async merges) carries no input fingerprint / generation ID. If the user edits an input while Run stays true and a task is in flight, the in-flight branch ignores the new snapshot; when the old task completes, its payload is emitted on the next solve against the NEW inputs (result for cloud A shown under cloud B). Also: CancellationTokenSource is replaced without disposal.	AsyncScanComponent (in-flight branch + _ready emission)	open, confirmed 2026-07-10 (external code review + direct source reading). Mitigation today: toggling Run false cancels and discards the in-flight result, and cancelled payloads are never emitted — the race needs an input edit while Run stays true, which is a normal user action. Fix: fingerprint or generation-stamp the snapshot at task start, reject a completed payload whose stamp mismatches the current inputs (auto-restart with the fresh snapshot), and dispose replaced CTS instances. Mechanical, one file; all five subclasses inherit the fix.
M8	Component-level exceptions are laundered into input warnings. FrahanComponentBase seals SolveInstance inside a blanket catch (Exception) that converts ANY uncaught failure — null-dereference bug, index bug, native failure, genuinely bad input — into one orange Warning ("Skipped - input or data issue", message trimmed to 240 chars). GhGuard similarly converts data-access exceptions to "No data". Deliberate UX (a mis-wired canvas reads as waiting, not broken), but it makes software defects indistinguishable from bad user input, with no stack trace, log, or diagnostic trail. A component can silently produce fewer blocks / empty toolpaths downstream.	FrahanComponentBase	open, boundary documented 2026-07-10 (external code review; confirmed verbatim). Counterpoint: the highest-stakes verdict path is NOT laundered — the masonry checker's independent residual audit (H3) fails loud. Fix: classify — known input/data exceptions stay Warnings; unexpected exceptions become hard Errors carrying a diagnostic ID plus a persisted log line (exception type + stack), so defects surface instead of degrading silently.

Mitigated (addressed, kept here for provenance)

ID	Was	Status
X1	Friction pyramid conservativeness: the raw FrictionConeBuilder K=4 exact-coefficient branch is optimistic (outer, under-constrains by up to sqrt(2)).	mitigated — the shipping MasonryStabilityChecker overrides to K=8 inscribed (mu_eff = mu cos(pi/K)); this session machine-proved the inscribed pyramid is a subset of the true Coulomb cone (Z3, math/verification). The card's K=4 finding describes the builder in isolation, not the shipping path.
X2	Honest density needed Rhino (VolumeMassProperties); Core VolumeEstimate was bbox only.	mitigated — MeshPackItem.MeshVolume (signed-tetra) + MeshPackResult.FillRatioMeshVolume added 2026-07-06, tested; Core rho no longer needs Rhino.
X3	Kriging.Predict header comment claimed (sill + nugget) - w^T w.	mitigated — corrected in code to the actual latent variance sill - w^T w.
X4	Unclear whether any GH path routed the coupled CRA certificate out of process.	mitigated — verified: the worker runs only the penalty RBE checker; the coupled certificate has a single in-process call site.

Reading this for deployment

The 2D nesting path (ContactNfpHoleNester: exact NFP-BLF, hole-aware, boundary mode) is the one path that clears every gate above: Rhino-free, deterministic, exactly validated (0-overlap boolean gate), and benchmarked against OpenNest 2.89 with a measured C# number (RESULTS). Everything 3D / masonry / geology carries at least one open High or Medium item and is not yet service-ready.