Umbrella: audit and burn down all open issues (2026-08-10)

[shinaoka]

Outcome

This umbrella replaces the superseded retirement umbrella #199 and inventories every issue that was open on 2026-08-10.

• Audit basis: origin/main at dc0a8e03286c61a84d56446b5cc2c53295f75d76
• Issues originally audited: 34
• Open PRs at audit time: 0
• Integration branch: umbrella/issue-burndown-2026-08
• Current product decision: all nine workspace crates remain maintained, per chore: prepare strided-rs v0.4.0 #226.
• Sequencing rule: performance optimization is the final phase. Earlier phases establish soundness, contracts, CI, API vocabulary, and documentation first.

Merge protocol

1. Create one focused branch per issue, or one explicitly documented tightly-coupled issue cluster.
2. Open implementation PRs with base umbrella/issue-burndown-2026-08, not main.
3. Keep the umbrella branch green after every merge: cargo fmt --all -- --check and cargo test --workspace; run feature, safety, docs, or benchmark gates appropriate to the issue.
4. Resolve conflicts and cross-issue API decisions on the umbrella branch.
5. Do not merge optimization work until Phases 1–3 are green.
6. After the final optimization phase and cross-phase audit, open one final PR from the umbrella branch to main.

Phase 0 — closed or rejected inventory

• Phase 0: deprecate and freeze strided-einsum2, strided-opteinsum, and the adapter crates #200 — obsolete after chore: prepare strided-rs v0.4.0 #226; closed not planned.
• Phase 0: disposition of open issues targeting the retiring crates #201 — obsolete disposition ledger; closed not planned.
• Phase 2: remove the retired crates (final releases, workspace removal, facade slimming, CI simplification) #202 — removal contradicted the nine-crate maintenance policy; closed not planned.
• perf: HPTT-inspired cache-efficient tensor permutation #105 — historical HPTT umbrella delivered by PRs perf: HPTT-inspired cache-efficient tensor permutation (#105) #111–docs: permutation optimization analysis + use strided-perm in finalize_into #113; closed completed.
• perf: HPTT-inspired blocked permutation in pure Rust (Step 3) #108 — requested 2D micro/macro-kernel architecture delivered by PR refactor: rewrite hptt module as 2D micro-kernel architecture #112; closed completed.
• Use cblas_dgemm_batch when available instead of manual batch loop #125 — speculative provider-specific batched CBLAS FFI without need-before-implementation evidence; closed not planned and measurement folded into einsum2 batched GEMM has 3.5x overhead vs Julia for small-m contractions #115.
• strided-perm: copy_into/copy_into_par call copy_nonoverlapping without an aliasing/overlap check (UB from safe code) #225 — claimed safe-code alias construction was not reproducible; closed not planned. SAFETY documentation remains under Marker remediation: SAFETY comments, #[allow] rationales, INVARIANT markers #219.

Phase 1 — soundness and boundary correctness

These block every later phase.

• strided-view: unsound manual Send impl on StridedView (requires T: Sync) #203 — fix the unsound StridedView: Send bound.
• strided-view: diagonal_view with duplicated axes gives out-of-bounds reads from safe code #204 — reject duplicated diagonal axes and validate constructed metadata.
• strided-view: checked arithmetic for shape products and stride construction in StridedArray constructors #205 — checked shape products and stride construction.
• strided-kernel: reduce_axis forms an out-of-allocation pointer when the reduced axis is empty #207 — empty reduced-axis pointer formation and identity-fill semantics.
• strided-kernel: destination injectivity for elementwise RMW ops; consolidate validate_destination_layout; dead cost==0 split guard #206 — unify destination injectivity validation and remove/restore the dead split signal.
• strided-view: panicking element accessors; Index skips the rank check #208 — canonical checked accessors, rank checks, and documented panic/error contracts.

Phase 2 — contracts, API decisions, and CI foundations

These define the execution and public contracts that later implementation must follow.

• Adopt the shared rule set by reference and add the strided-specific rules #216 — adopt shared rules by reference and retain only strided-specific policy.
• Suffix vocabulary for the retained kernel surface (accumulating ops and _view) #220 — decide suffix vocabulary before broad documentation changes.
• Unify the parallel threshold and extend the ExecutionPolicy contract and rayon guard test to strided-perm #209 — one threading threshold and a policy-aware strided-perm.
• strided-kernel: scalar tiled transpose kernels are gated on the parallel feature #210 — make scalar tiled-kernel selection independent of the parallel feature.
• strided-kernel: make copy_into_col_major policy-aware or remove the duplicate API #212 — make copy_into_col_major policy-aware or remove the duplicate API.
• Nested parallelism: Strided thread splitting vs BLAS internal threading #6 — define the Rayon/provider threading contract.
• strided-einsum2: allow multiple GEMM backends to coexist and support runtime selection #86 — complete backend coexistence, explicit selection, and ActiveBackend migration policy.
• CI: feature matrix, run clippy, deny rustdoc warnings #215 — feature matrix, clippy, and rustdoc warning gates.
• Track Faer typed MaybeUninit overwrite API for strided GEMM #198 — external Faer typed-MaybeUninit tracker; remains blocked until upstream exposes a sound overwrite API.

Phase 3 — documentation, organization, and repository hygiene

Complete this phase before performance work so benchmarks and optimization PRs use stable rules, APIs, and CI.

• Docs: move benchmark tables to strided-rs-benchmark-suite; fix stale versions and faer_design.md claims #214 — move benchmark results to the benchmark suite and correct stale docs.
• Docs remediation: # Errors sections and runnable doctests #217 — concrete # Errors sections and runnable doctests.
• Test organization: extract inline cfg(test) modules; resolve strided-perm's test-gated implementation modules #218 — extract inline tests and resolve test-gated implementation modules.
• Marker remediation: SAFETY comments, #[allow] rationales, INVARIANT markers #219 — verified SAFETY/allow/INVARIANT markers; no blind comment insertion.
• Add Debug impls to the remaining public types #221 — re-audit all nine maintained crates and add missing Debug implementations.
• Build-artifact hygiene: no documented stale-artifact pruning mechanism #229 — repository-level stale-artifact pruning policy.

Phase 4 — performance optimization (final phase)

No item in this phase lands before Phases 1–3 are green. Apply the shared need-before-implementation and paired-experiment protocol; microbenchmarks alone are insufficient.

• strided_perm::copy_into is 4x slower than Julia permutedims! for high-rank binary tensors #114 — high-rank, small-extent permutation bottleneck.
• Adopt OMEinsum.jl's pre-permutation dimension fusion in einsum2 #116 — pre-permutation contraction-group fusion/materialization avoidance.
• einsum2 batched GEMM has 3.5x overhead vs Julia for small-m contractions #115 — prepared/reusable einsum planning and measured small-m/batch overhead.
• strided-kernel: parallelize typed reduce_axis through the existing reduction partitioner #211 — policy-aware parallel typed reduce_axis.
• strided-kernel: incremental offsets in erased reduce and indexed replay hot loops #213 — incremental offsets in erased/indexed hot loops.
• perf: avoid finalize copy in einsum2 by deferring output permutation (lazy permute) #128 — decide lazy/eager/heuristic output permutation from current end-to-end evidence after strided_perm::copy_into is 4x slower than Julia permutedims! for high-rank binary tensors #114/Adopt OMEinsum.jl's pre-permutation dimension fusion in einsum2 #116.

Dependency summary

• Phase 1 blocks all later phases.
• Adopt the shared rule set by reference and add the strided-specific rules #216 → Docs remediation: # Errors sections and runnable doctests #217/Marker remediation: SAFETY comments, #[allow] rationales, INVARIANT markers #219/Add Debug impls to the remaining public types #221/Build-artifact hygiene: no documented stale-artifact pruning mechanism #229.
• Suffix vocabulary for the retained kernel surface (accumulating ops and _view) #220 → Docs remediation: # Errors sections and runnable doctests #217 and public API documentation updates.
• Unify the parallel threshold and extend the ExecutionPolicy contract and rayon guard test to strided-perm #209 → strided-kernel: parallelize typed reduce_axis through the existing reduction partitioner #211/strided-kernel: make copy_into_col_major policy-aware or remove the duplicate API #212 and all policy-aware parallel work.
• Nested parallelism: Strided thread splitting vs BLAS internal threading #6/strided-einsum2: allow multiple GEMM backends to coexist and support runtime selection #86 establish backend and threading policy before einsum2 batched GEMM has 3.5x overhead vs Julia for small-m contractions #115.
• strided_perm::copy_into is 4x slower than Julia permutedims! for high-rank binary tensors #114/Adopt OMEinsum.jl's pre-permutation dimension fusion in einsum2 #116 → perf: avoid finalize copy in einsum2 by deferring output permutation (lazy permute) #128.
• Phase 4 starts only after Phases 1–3 are green.

Each retained issue has a current-state audit comment and an implementation handoff pointing to this umbrella and branch.