# Cross-component cleanup coverage

The earlier Rust cleanup is committed through `59862e8`; its full workspace
check passed 394 tests with 7 ignored. This broader pass is complete: each
component below was reviewed, selected changes are committed, and no selected
cleanup remains outstanding. This is an audit disposition, not a claim that
no future simplification can be found.
Generated artifacts and historical benchmark outputs are reference material,
not sources to rewrite.

| Component | Current coverage | Review disposition |
| --- | --- | --- |
| Compiler/planner, storage, exchange | Shared fusion contracts, span geometry, endpoint accounting and package assembly reviewed; forward/reverse view adapters consolidated in low; identity-copy predicates unified | Candidate generation, ranking and baseline lowering also reviewed; retained separate accounting paths because they model different boundaries |
| Package/driver/runtime/CLI | Shared binding extents, host capture, logging and profile interning reviewed; redundant runtime error wrapper removed | Loader, host invocation and CLI dispatch inspected; no further extraction selected |
| ELF toolchain | Hashing/cache and linker paths inspected; instruction field relocations consolidated; relocation bounds constrained to their section | No additional duplication selected from this inspection |
| Device kernels/runtime | Dense packing supervisor loops, GEMM weight loads and runtime word copies consolidated; cast and normalization families inspected | C++ cast/rearrangement scaffolding simplified; other arithmetic families retained where their instruction/dataflow contracts differ |
| Profile viewers | All three viewers inspected; runtime canvas setup and coordinate conversion consolidated | No additional cross-viewer extraction selected: placement uses virtual rows/DPR scaling; runtime uses event-count-bounded chunks and accounting segments |
| Calibration tools | Shared Torch-only F143 module; shared Hessian accumulation and scale selection; shared calibration loop with exception-safe hook removal; offline placement tool inspected (independent constraints intentionally retained) | No additional duplication selected in the pretrained exporter |
| Experiment scripts | Gather/packing affine detection consolidated; frontier, batch/MLP sweep orchestration and SDK summary inspected | Resume/provenance and failure handling reviewed; independent sweep policies retained |
| Diagnostic harnesses | Shared bindings/logical packing reviewed; six standalone kernel binaries share locked runtime loading and use HostSession::finish for deferred output completion; four share timestamp bindings | Terminal-state polling shares the model checker; kernel-specific fixture assembly and numerical references remain independent |

Calibration validation: four tests from `tools/test*calibration*.py`; deterministic
comparison against the pre-extraction quantization functions produced identical
FP8 projections, reconstructed weights, bias corrections and objectives. Both
successful collection and injected-forward-failure hook cleanup were checked.
The core numerical module was also imported without loading Transformers.

ELF validation: five unit tests and workspace checks pass. A synthetic ELF
fixture with two retained sections linked a valid cross-section reference to
the expected bytes; moving the relocation offset beyond its originating
section was rejected. Field tests cover opcode preservation, maximum values,
misalignment, truncation, and `usize::MAX` offsets.

Device packing validation: the SDK compiled both old and new block-major and
transposed-right kernels with identical allocated section bytes and normalized
relocations. All 22 kernel selection tests passed. AMP-left packing retains its
different destination stride and worker frame.

GEMM/runtime validation: the SDK produced identical allocated section bytes
and normalized relocations for FP32 GEMM, FP16 GEMM, all four FP16/FP8 ×
standard/interleaved weight dispatchers, and the static runtime. The shared
weight loader preserves its unrolled order and both instruction widths.
The full codegen suite also passed: 295 tests, 5 ignored.

Exchange experiment validation: ten reconstruction/geometry tests pass.
On 2,000 deterministic copy sequences, both the optimistic loop count and the
forward-only task descriptors match their previous independent implementations.
Both changed scripts and their tests pass Ruff checks.

Diagnostic completion validation: all ipu-tests targets compile. Hardware
unpack checks passed 14 cases / 84,208 bytes; FP8 cast checks passed 1,404
cases / 7,517,232 bytes, including scales, tails, padding and shifted overlap.
The softmax and kernel-equivalence binaries retain their additional context
halt checks after the shared output handshake.

Diagnostic setup validation: all diagnostic targets compile after the shared
loader extraction; 14 unpack cases / 84,208 bytes pass on hardware. The runtime
can only be borrowed from its owner, which drops it before releasing the lock.
Pretrained fixture export and baseline/candidate entry points were also sampled;
no further consolidation was selected from those portions in this pass.

Completion diagnostic validation: model, softmax and kernel-equivalence
paths now share the existing model terminal-state checker, including the
completion marker, InvalidProgramCounter exception and halted workers. Four
softmax hardware cases passed after extraction; all diagnostic targets compile.

View adapter cleanup: logical-range to shard-extent translation now belongs
to low expansion and is shared by forward and reversed views; mid/view.rs
is removed. The low expansion suite passed (see /tmp/ipu-view-cleanup-tests.log).

Runtime cleanup: removed the single-variant wrapper around DriverError and
its thiserror dependency; RuntimeError remains a public alias. Workspace
compilation passes. No callers pattern-matched the old wrapper in this tree.

Identity mapping validation: an omitted-offset versus explicit-zero-offset
copy regression failed before consolidation (exchange row estimate 18,436
versus 4,612 bytes) and passes afterward. Cast motion, producer tracing and
costing now use the same predicate. Full codegen suite: 296 passed, 5 ignored.

Viewer validation: headless Chromium rendered the same incremental profile
with old and new templates. Initial and zoomed canvases matched pixel-for-pixel;
scaled pointer coordinates, hit tests, time conversion and timeline clicks also
matched. Sidecar loading remained enabled. Evidence: /tmp/ipu-viewer-cleanup/.

CLI/loader review: the CLI enum adapters keep clap out of the profiling
library; package-load and host-run commands have distinct completion policies.
The driver already shares phase handshakes and output capture. Additional
wrappers for its short attachment checks were not selected. Both Rust profile
report tests pass after the viewer change.

Final device inspection: constant/runtime dimension selection in rearrange_f16.cpp
now uses constant-folded expressions; cast_f8.cpp no longer has a single-use
field macro. Ten old/new SDK objects (three packing orders, static/dynamic
dimensions, and four cast type pairs) have identical allocated bytes and
normalized relocations.

## Final verification

- Full workspace: 396 passed, 0 failed, 7 ignored (27 suite results).
- Python: 10 exchange-experiment tests and 8 calibration/placement tests passed.
- Changed Rust files pass rustfmt; git diff --check passes.
- Browser, hardware, ELF fixture and SDK object-equivalence evidence is recorded
  above; numerical reference implementations retain their independence.
- 13 commits after `59862e8`; approximately 390 fewer non-test source lines.
  The line count excludes standalone tests and Rust cfg(test) tails, including
  the shared modules added by this pass.
- The only tracked worktree change left is the user's pre-existing `TODO` edit.
  Existing untracked artifacts and review/callgraph files were preserved.

Full test logs: `/tmp/ipu-cleanup-final-workspace.log`,
`/tmp/ipu-cleanup-final-python.log`, `/tmp/ipu-cleanup-final-format.log`.