Eager borrowed-read forwarding: 1T instruction evidence

Candidate 2fd8163e568a232072348e8939d534713f3b3a98 removes blanket input materialization before existing standard-operation _read hooks. Owned-only operations and dtype promotion retain their explicit boundaries. CPU kernels, SIMD code, and the linked BLAS overwrite implementation are unchanged.

Protocol and result

protocol.json was written before collection. Three complete independent baseline/candidate N1/N3 pairs were collected for every case, sequentially, with one fixed warmup. All 36 invocations completed successfully. The metric is whole-process Callgrind Ir differencing for warmed eager operations, not kernel-only cost, wall-clock time, bandwidth, or a native speedup estimate. Fixed preparation cancels; per-repeat harness work remains. Identical binary paths share a cached measurement, hence divisor 2. LM's two distinct paths give divisor 4 and a path mean, not an individual-path result.

Workload	Baseline median Ir	Candidate median Ir	Paired reduction range
multiply 2048²	11,106,156.5	5,309,045	52.180–52.215%
GEMM1024 control	524,967,422.5	523,500,184	0.278–0.281%
LM3 12d path mean control	968,053,178.5	950,507,132.5	1.812–1.814%

The predeclared primary ≥20% reduction and both ≤1% regression controls pass. Vendor dgemm_kernel_HASWELL self-Ir is identical in every baseline/candidate pair: 512,386,632 for GEMM1024 and 444,650,842 for LM. This does not establish unchanged wall-clock kernel cost.

For multiply, two 32 MiB input payload copies are eliminated. Their 128 MiB combined logical read/write traffic is arithmetic accounting, not measured DRAM traffic. A recording regression observes two baseline materializations versus zero after this change, with nonzero transposed inputs and unchanged source values. Canonical packing still dominates remaining LM copy work; this change does not claim to remove all eager or extension-boundary copies.

Provenance and reproduction

Baseline tenferro: 57bee76de1a752782bdd985f332050422f885372 (the prior BLAS-full-overwrite candidate, not its older baseline).
Candidate tenferro: 2fd8163e568a232072348e8939d534713f3b3a98.
Both use strided 17e05ffb168d0f826529ea2c3aceeb4ec851448b, with local overrides; benchmark executable source is 976c4b9a459f94086c714ec0f6836978a7bbdfc0.
Existing user-authorized OpenBLAS Docker image, Rust 1.98.1 / LLVM 22.1.8, OpenBLAS 0.3.26 LP64 pthreads, Valgrind 3.22.0, CPU16 (L3 CPUs16–23). The build and analysis image IDs match. Image/library/binary hashes, thread settings and host snapshots are in raw/environment.txt.
configure_cpu_thread_env 1, TENFERRO_MODE=eager, BLAS selected, decomposer disabled. A separate diagnostic intercepted GEMM1024 calls and verified openblas_get_num_threads()==1 for both binaries. The interceptor from ../blas-full-overwrite/verify_threads.c was absent from measurements.
The host still had active Julia workloads. No quiet-host native comparison was attempted, and profiler-emitted time columns must not be used as one.

Build each revision from the benchmark checkout with its matching local strided override, preserving the baseline binary before switching source:

RUSTC_WRAPPER= CARGO_PROFILE_RELEASE_DEBUG=1 cargo build -j 16 \
  --config /tmp/local-patch.toml --release --no-default-features \
  --features system-openblas --bin tenferro-einsum-benchmark

Keep the embedded benchmark fixture path resolvable. The three fixtures and Cargo.lock/local patch used here are retained under raw/. Inside the Docker workspace, run reproduce.sh with BASELINE, CANDIDATE, and a fresh OUT. It reproduces the collection loops (original collection used one invocation per pair). Put its profile/log/annotation outputs in raw/ for summarize.py; the parser accepts plain or gzip artifacts and checks run/thread metadata. The script is syntax-checked; its equivalent collection loops were executed.

Correctness and remaining gates

91 debug library tests, 354 functional integration tests, 174 doctests pass.
Docker release library: 91 tests pass. The focused new regression also passes release Memcheck with zero errors; leak checking was disabled.
Profile fixtures are zero-input execution smoke checks, not independent numerical validation. Nonzero correctness comes from the regression suites.
The UI integration wrapper still fails: --config dependency patches are not inherited by trybuild. Temporary manifest patches resolve dependency compilation, revealing two expected-stderr mismatches (underline formatting and a kache-remapped path). No unrelated snapshots were blessed; temporary manifest edits were restored. Raw failure logs are retained.
Full repository validation, 4T/AD performance validation, native acceptance, and override-free integration remain pending. Publication and dependency-pin integration remain deliberately last. This is an optimization candidate, not a fully validated release.