# Review after the September 8 cleanup Baseline: `42ad0eb`, the last production-code cleanup, through `9deab80`. The review follows the intervening changes by subsystem: graph search and Repeat attachment, memory estimates and parameter ownership, copy traversal and caching, exchange scheduling/relocation, package allocation, FP8 and normalization kernels, and their benchmark/profile interfaces. The changes include substantial new functionality; deleting it wholesale would undo requested features and measured improvements. ## Corrections and simplifications * **Tensor memory and estimated exchange storage are separate.** Previously, `MemoryPeaks.standard` and `.total` included estimated rows. Capacity checks and profile rendering subtracted them again, and refreshing row estimates subtracted the old estimate from an already saturated total. Tensor liveness now owns tensor peaks; row refresh changes only the row estimate. Ranking explicitly combines storage where it previously used a combined objective. Existing capacity decisions and objectives are covered by the memory tests. Profile format version 3 documents the changed peak-field meaning. * **Package selection uses the final placed cost.** Final placement already recomputed the schedule cost, but only logged it. Selection compared costs from provisional placement. The finalizer now returns its computed cost with the artifact; acceptance and ranking consume that result together. A regression supplies equal provisional plans with different finalized costs and verifies that the better finalized package wins. No extra scheduling or costing pass was added. * **Occupied-range merging has one implementation.** Diagnostic package assembly duplicated it for tile-local storage and host descriptors, and the tensor allocator had a third version for lifetime-reused ranges. Those callers now share the union operation. Diagnostic host descriptor allocation also stops at the loader's application limit, like normal package assembly, instead of allowing the unavailable tail of physical SRAM. * **Runtime copy retention follows the selected calls.** A local-copy boolean used to retain all four word/strided-copy helpers, with a separate halfword flag. Retention now records the actual selected symbols, including inside Repeat. Linker dependency handling still retains their referenced support. The duplicated classification and over-retention are removed. * Named the expansion cache bucket/counter fields instead of indexing a four-field mutex tuple. Lookup, collision checking, locking and entry limits are unchanged. * Removed the redundant contiguous-capacity wrapper while retaining its public reservation-aware interface. ## Mechanisms retained deliberately * Compact affine traversal, complete FP8 panels, early multicast ownership, shared contracts and the bounded copy-preparation caches address measured expansion costs. Coordinate-oracle and cached/uncached comparisons remain. The earlier rejected whole-compute caching experiment is already absent. * Incremental exchange encoding, shared speculative histories, matching and schedule replay address the captured large ViT stages. The paired-transfer changes are backed by SDK/hardware investigations, not arbitrary restrictions to restore. The compact stream-order alternative remains an explicit experiment, not an implicit production fallback. * The row-footprint model remains a ranking estimate, separate from the actual encoded table cap. Fragment and shortlist limits still bound compiler work. The 80-KiB table cap remains useful: full-depth complete-panel tables need approximately 70 KiB. Returning to 64 KiB would reject those before placement. * Repeat's body cache caches an unchanged body under explicit boundary and allocation constraints. It is distinct from the rolled-back regional optimizer. Resident-sequence lifetimes, dense sequence allocation and compact consumer-compatible parameter homes fix real storage/relocation errors. Their failure to make one full-depth plan fit does not make those fixes redundant. The ineffective extra allocation ordering was already removed in `9deab80`. * FP8 cast orders, distributed packing and producer fusions represent different data-movement choices, not interchangeable wrappers. Their code paths and numerical kernels have specific shape/precision contracts and measurements. * Padding-removal restrictions around mixed precisions, attention scratch and Repeat aliases are conservative correctness rules. Replacing those with precise range/version analysis would be new compiler functionality, not a safe deletion during this cleanup. ## Remaining limits The full-depth failure is still a real live-memory problem for the selected ownership: weights plus cast input/output exceed the available tensor space on one tile before other allocations. This review does not claim to fix it or to provide a proof that all layouts are infeasible. Better activation/parameter ownership and earlier accounting of final support reservations remain separate planning work. The expensive search retries and imperfect exchange estimates also remain; they should be changed through measured planning improvements, not by adding another unvalidated fallback. ## Validation and size * 228 codegen library tests passed; five ignored. This includes the added finalized-cost selection regression and the updated tensor-memory assertions. * The subsequent cache field rename passed all 32 low-expansion tests (one ignored). Clippy passed after both runs with the existing `too_many_arguments` / `type_complexity` allowances. * A small batch-one, two-layer FP8 ViT with fused QKV and 64 compute tiles passed hardware execution and final-output reference comparison (maximum absolute error 0.135742, diagnostic tolerances 0.2 absolute / 0.05 relative). Detailed profiling readback also passed. Artifacts, log and rendered profile are in `artifacts/cleanup-20260910/`. The package includes the functional changes through `0b8fa3a`; the final cache field rename was checked separately. This is correctness validation, not a full-size performance comparison. * Relative to `9deab80`, the production portions of the Rust changes remove 34 lines net, counting comments/blank lines and excluding test changes. The reduction is modest; the principal gains are removing mixed memory accounting, incorrect provisional-cost ranking and duplicated support classification. * User changes in `TODO` and existing untracked artifacts were left intact.