026-G11 diagnostics and failure modes bootstrap¶

G11 remains NOT_STARTED. This bootstrap inventories existing mechanisms and maps historical evidence; it does not requalify another gate or close G11.

Inventory¶

Area	Classification	Current mechanism
Schema, material, element and load validation	`PUBLIC_FAIL_CLOSED`	`InputValidationError` with field-level messages
Mesh, connectivity, geometry and active-DOF checks	`PUBLIC_FAIL_CLOSED`	`MeshValidationError` and deterministic mesh reports
Singular linear systems	`PUBLIC_FAIL_CLOSED`	`NumericalConvergenceError` with singular-system message
Nonlinear, contact and tangent failures	`INTERNAL_DIAGNOSTIC` plus public failure	`NonlinearFailureReason`, diagnostics payloads and rejection logs
Eigenvalue/mode and buckling boundaries	`INTERNAL_DIAGNOSTIC` plus public failure	backend/buckling diagnostics and explicit convergence errors
Trial/commit/rollback	`INTERNAL_DIAGNOSTIC`	transaction equality and `rollback_before_retry` evidence
Unsupported combinations	`PUBLIC_FAIL_CLOSED`	schema/mesh/route rejection tests
Unified cross-route failure envelope	`OWNER_APPROVED_BOUNDED`	focused runner and four native adapters; broad cross-route campaign remains future work
Observed uncontrolled failures	none in reviewed evidence	no uncontrolled failure was promoted or hidden

The reviewed evidence includes the six G04 invalid-input cases, G05/G06 inverted-element expected failures, G06 J2 rollback, historical buckling and contact diagnostics, and the TL failure/cutback/rollback zoos. These are mapped as diagnostic evidence only and are not automatically requalified.

Owner contract review¶

The machine-readable contract is in qualification/0_2_6/g11_requirements.json. It defines eight requirements: invalid inputs, singular systems, unsupported combinations, numerical non-convergence, deterministic diagnostics, rollback/state integrity, provenance, and no silent fail-open. Each records failure classes, expected behavior, oracle, evidence references and the Owner review decision KEEP.

All eight requirements are retained. The review requires fail-closed behavior, determinism, no silent NaN/Inf continuation, no silent PASS, exploitable diagnostics and provenance; state preservation is mandatory whenever mutable trial state exists.

The transverse failure envelope is OWNER_APPROVED_BOUNDED as a qualitative evidence schema. It requires FAILURE_CLASS, ROUTE, EXPECTED_BEHAVIOR, ERROR_TYPE_OR_CODE, STATE_PRESERVED, DETERMINISTIC, NO_NAN_INF, NO_SILENT_PASS and EVIDENCE_ID. Route-native exception types remain allowed; this schema does not require a common Python exception or imply a successful qualification result. No numerical threshold is introduced.

Current mapping is in qualification/0_2_6/g11_evidence_mapping.json. The four cases remain planned specifications in qualification/0_2_6/g11_adversarial_cases.json; focused native-route execution is recorded separately in qualification/0_2_6/g11_native_execution_evidence.json. The execution covers one singular system, one explicit unsupported combination, one controlled non-convergence, and one rollback/state-integrity rejection/retry. It is not a transverse campaign and does not close G11. Any aggregate threshold or maturity promotion requires a separate Owner review.

The route-neutral runner is src/solveur/verification/g11_runner.py, with route-native adapters in src/solveur/verification/g11_native_adapters.py and focused tests in tests/unit/test_g11_runner.py and tests/unit/test_g11_native_adapters.py. It accepts injected route-native adapters, emits the approved envelope, checks repeatability, NaN/Inf, silent PASS and state preservation, and archives JSON provenance and diagnostics. It does not change solver implementations. The four-case focused execution is recorded; the full G11 campaign remains unexecuted.

The focused cross-route extension is specified in qualification/0_2_6/g11_cross_route_cases.json and recorded in qualification/0_2_6/g11_cross_route_execution_evidence.json. It exercises one fail-closed case each for geometric_nonlinear_static, modal, linear_buckling and linear_static_contact. These routes remain PARTIAL_WITH_RUNTIME_FAILURE_EVIDENCE; one failure mode per route is not a qualification of all route capabilities.

Coverage boundary¶

The contract is transverse, but evidence remains route-specific. Linear static input rejection, nonlinear transaction rollback and provenance controls are the strongest mapped areas. Modal/buckling, contact, geometric-nonlinear route execution, deterministic cross-route diagnostics and a unified no-silent-PASS assertion across every route now have focused runtime evidence, but remain partial pending broader route coverage. G04, G08, G09, G07/TL and Agent A evidence is reused only at its historical diagnostic level.

The focused blocker-closure aggregation is recorded in qualification/0_2_6/g11_cross_route_aggregation.json. It joins the four previous native cases, the four partial-route cases and the additional HEX8 mutable case by route, failure class, expected behavior and runtime result. The four formerly partial requirements are assessed as SATISFIED_BOUNDED for this evidence envelope only; their official contract statuses remain PARTIAL, and the prior nonlinear rollback case remains the bounded mutable/retry proof.

Boundary¶

G04, G08, G09, G07/TL and Agent A work are untouched. The contract remains open and G11 remains NOT_STARTED: focused native evidence is recorded, while the broader adversarial/cross-route campaign and any maturity decision remain future Owner work.