Repository navigation
Conversation
|
I added The new cases cover:
Several cases intentionally require a contract decision before they can be registered in CTest—for example, a dedicated unsupported-result-topology failure and a distinction between exact-in- |
|
Diagnosed I added root |
|
Added an apply-ready manual patch series under
Apply from the repository root in order: git apply --check patches/0001-exact-patch-interior-witness.patch
git apply --check patches/0002-verify-patch-probes-exclude-holes.patch
git apply --check patches/0003-regress-annular-patch-witness.patch
git apply patches/0001-exact-patch-interior-witness.patch
git apply patches/0002-verify-patch-probes-exclude-holes.patch
git apply patches/0003-regress-annular-patch-witness.patchThe series replaces the incomplete centroid witness heuristic with exact vertical decomposition, strengthens independent verification to exclude hole interiors, and adds a through-column annular regression before realization. |
|
I’m implementing the next outstanding tracker component, P0 — freeze product contracts, on |
|
P0 — Freeze product contracts is complete and Landed schema-3 product policy/result/provenance/replay contracts, immutable shared-lifetime exact/result ownership, explicit backend maturity and qualification rules, representation-specific capability checks, fail-closed fallback categories, strict bounded decoders, public support-boundary documentation, fixed option/replay golden vectors, and focused Component 1 tests. Validation: the definitive focused workflow built the repository and passed The next outstanding tracker component is P1 — publish a durable authoritative exact result. |
|
Implemented tracker component P5.1 — stable provenance and attribute transfer. Highlights:
Commits:
Validation on the final commit is green:
The final remote tree matches the locally validated tree exactly. |
|
Implemented tracker component P5.2 — the one-call conservative Boolean service. Highlights:
Commits:
Validation on the final commit is green:
|
Temporarily install the self-restoring registered workflow that validates and publishes the reviewed Component 07 coplanar event-topology slice to PR #90.
…ure guide Reduce the frozen fuzz-duration floor from 24 CPU-hours to 10 minutes (600 aggregate CPU-seconds) across the qualification matrix constant, the P6.10 campaign driver, the frozen candidate fixtures, and the Plan 16 / qualification documentation. This is the leadership cost-saving relaxation; fail-closed evidence, false-success blocking, and minimization/promotion requirements are unchanged. Add docs/MeshBooleanBetaTesting.md so end users can report field failures unambiguously (typed code/subcode/message-key, replay binding digest, serialized operands, and environment) and reference it from the product contract documentation.
…report Add qualification_defect_kind, make_qualification_demotion_report, and qualification_report_authorizes_promotion so a discovered false success, unexplained disagreement, schema incompatibility, or material platform defect explicitly revokes or demotes a previously qualified profile. Add focused tests. Commit the reproducible P6.11 qualification report naming the exact repository state and recording that the retained candidate campaign remains incomplete/blocking; no profile is qualified, qualified_default stays fail-closed, and all profiles remain experimental or candidate.
…mpaign mode Add --available-toolchain to scripts/run_mesh_boolean_p610_campaign.sh so the non-deferred frozen-manifest step and the six runnable fuzz allocations are re-enabled with only the toolchain present on the invocation host. The frozen matrix dimensions that need unavailable tools (oldest-supported compilers, libc++, ThreadSanitizer, independent AArch64) and the two TSan fuzz allocations are recorded as documented known limitations (a new non-blocking step status) instead of blocking anomalies, and a completed run finalizes with the distinct complete_limited_toolchain status. Add the built-in non-deferred inventory dispatcher scripts/run_p610_nondeferred_inventory.sh and clear campaign run/assess instructions in tracker.md. Extend the driver self-test for the new inventory.
…ings In --available-toolchain mode always use the built-in non-deferred inventory dispatcher instead of an external P610_NONDEFERRED_CAMPAIGN_COMMAND, so a stale environment variable cannot silently substitute /bin/true and skip the frozen non-deferred manifest. Record the limited available-toolchain campaign findings in the tracker instructions: the reproducible B0-B8 frozen canonical-output identity drift (root cause narrowed to a post-P12 productization commit changing canonical bytes content), the instrumented Debug build cc1plus terminations at -j32, and the Clang 7.0.1 libstdc++ exact_point3 destructor link failure.
…digests - Give exact arithmetic/kernel value types out-of-line defaulted destructors so Clang 7.0.1/libstdc++ emits both complete-object (D1) and base-object (D2) destructor symbols, resolving the link-time 'undefined reference to exact_point3::~exact_point3()' on the Clang profiles. - Re-freeze B0-B8 output stage semantic digests after review: the selected_exact_boundary schema bump (3->4) and preparation-provenance fields from P2.1/P5.2 intentionally change the selection/realization/ output digest chain while all structural counters and canonical byte counts are unchanged.
- Fix dangling references in property tests: binding a reference to context->performance()->stage(...) on a temporary shared_ptr leaves a dangling reference once the temporary is destroyed. Copy the performance_counter_snapshot/stage snapshot by value instead. - Fix the independent symbolic-registry verifier: it indexes a.vertices[id] / a.curves[id] by canonical id without bounds checks, so a malformed (mutated-id) artifact caused a heap-buffer-overflow instead of a clean rejection. Reject ids that do not match their storage indices (and null upstream pointers) before any dereference, mirroring the structural invariant in valid().
ASan/UBSan (and TSan/MSan) slow the exact-arithmetic Boolean tests by roughly an order of magnitude, so the fixed 300-600s timeouts turned six correct tests (GlobalArrangement.Properties, CellClassification.Properties, Selection.Properties, Approximate.Adversarial, Replay, PerformanceBaselines) into false timeout failures. Scale every mesh Boolean CTest timeout by 6x under WITH_ASAN/WITH_TSAN/WITH_MSAN so the timeout stays a safety cap rather than a false failure.
…ions Document the six campaign findings (B0-B8 digest drift, Clang D1/D2 ABI, instrumented-build resource pressure, dangling test references, verifier out-of-bounds, sanitizer test timeouts) and their in-tree resolutions, and record the available-toolchain limited campaign as candidate evidence.
… no-promotion outcome
2 participants
No description provided.