Build nested typed evidence rows directly while retaining every original field and validation. Complete serialized bytes match the prior writer on recorded d8 evidence and regression fixtures.
Fresh-head combined qualification: 459 tests passed. Cloud performance measurements do not claim device loop-rate recovery or a successful learning trial.
Merge bounded control/output FIFO heads using original publisher headers. Preserve original age, native, source, overflow and malformed-input checks; retain producer order after joins.
Validated with actual pinned cloud dependencies and fresh-head regression tests.
Symmetric native probes could qualify a coordinate even when only the opposite direction changed its encoded command. Carry each direction’s measured visibility and physical effect into selection and interrupted-study recovery so a quantized direction cannot borrow the other direction’s evidence.
Keep existing trial bounds, slew, deadlines and independent acceptance criteria. Validate the previous selector failure, full FLM suite and recorded d5 producer/learner replay. Real-drive improvement requires post-update recorded evidence.
The d5 Ioniq 6 logs repeatedly reset source tracking when Card repeats a command that already arrived on time. Discard only a recent unchanged repeat after the existing native CAN verifier succeeds; add no sample, application time or permission.
Use independently admitted pending CarState brackets to complete original poses without waiting for that command to reach Card. Classify stale control source clocks as transient stale evidence, preserving the original bounded study deadline. Replace Python character loops with equivalent strict text predicates.
Validation: focused regression suite and paced d5 replay with original Card command references; actual controller/CAN trial application and worsening-trial rollback. No timing/acceptance threshold changes or claimed road improvement.
Prabhaav Pillai