Add uninitialized reduction and indexed replay

docs/2026-07-31-issue-187-worklog.md

@@ -0,0 +1,110 @@
+# Issue #187 Worklog
+
+## RED History
+
+- RED: indexed uninitialized destinations required a safe copy-then-update
+  boundary; direct initialized backing views were forbidden.
+- GREEN work introduced private raw-pointer writers and a closure-scoped
+  post-copy receipt.
+- Acceptance work adds differential lifecycle and indexed coverage.
+
+## Architecture and Safety Contract
+
+- Only CopyPlan constructs the private post-copy receipt.
+- Receipt construction is closure-scoped and cannot escape the HRTB helper.
+- Uninitialized writers use MaybeUninit storage and full-overwrite or
+  copy-then-update proofs before any typed read.
+- Reduction terminal writers use raw pointers plus validated extents.
+- Integer erased scatter uses wrapping i32/i64 combine functions; typed public
+  scatter semantics remain unchanged.
+
+## Verification
+
+Commands:
+
+    cargo fmt --all
+    cargo test -p strided-kernel --test issue_187_uninit_indexed
+    cargo test -p strided-kernel --features parallel --test issue_187_uninit_indexed
+    cargo bench -p strided-kernel --features parallel --bench issue_187_uninit_indexed --no-run
+
+Initial fixture pass did not run benchmark timing; later affinity benchmark
+evidence is recorded below with exact candidate-relative upper bounds.
+
+## Typed-storage migration after issue #190
+
+The indexed and reduction fixtures now use concrete `Vec<MaybeUninit<T>>`
+storage and `from_uninit_slice`; typed accessors are used for post-replay
+inspection. The all-dtype macros instantiate the concrete `$ty`, and hole
+tests do not read unreachable elements. The only raw pointer construction
+left in these fixtures is the narrow, documented stale-invalid `Bool` input
+case.
+
+Verification on the issue-187 worktree:
+
+```text
+cargo fmt --all
+cargo test -p strided-kernel --test issue_187_uninit_indexed --test issue_187_uninit_reduce
+cargo test -p strided-kernel --features parallel --test issue_187_uninit_indexed --test issue_187_uninit_reduce
+cargo test -p strided-kernel --test issue_187_source_contract
+cargo check -p strided-kernel --all-targets --all-features
+cargo bench -p strided-kernel --features parallel --bench issue_187_uninit_indexed --no-run
+```
+
+Results: indexed 68/68 and reduction 11/11 passed in both default and
+parallel configurations; source contract 5/5 passed; all-target and bench
+no-run checks passed.
+
+Focused strict-provenance Miri passed:
+
+```text
+MIRIFLAGS='-Zmiri-strict-provenance -Zmiri-symbolic-alignment-check' cargo +nightly miri test -p strided-kernel --test issue_187_uninit_indexed aligned_uninit_lifecycle_all_indexed_families
+MIRIFLAGS='-Zmiri-strict-provenance -Zmiri-symbolic-alignment-check' cargo +nightly miri test -p strided-kernel --test issue_187_uninit_indexed bool_gather_invalid_operand_rejects_before_mutation
+MIRIFLAGS='-Zmiri-strict-provenance -Zmiri-symbolic-alignment-check' cargo +nightly miri test -p strided-kernel --test issue_187_uninit_indexed dynamic_update_hole_layout_preserves_unreachable_bytes
+MIRIFLAGS='-Zmiri-strict-provenance -Zmiri-symbolic-alignment-check' cargo +nightly miri test -p strided-kernel --test issue_187_uninit_indexed gather_validation_errors_preserve_sentinel
+MIRIFLAGS='-Zmiri-strict-provenance -Zmiri-symbolic-alignment-check' cargo +nightly miri test -p strided-kernel --test issue_187_uninit_reduce validation_errors_leave_uninitialized_bytes_untouched
+```
+
+The first Miri pass exposed a fixture bug: an invalid-Bool destination was
+being compared before initialization. It was changed to initialized typed
+Bool sentinel storage; no production code was changed. The corrected run
+passed all five filters.
+
+## Sol high follow-up
+
+The dynamic-update hole coverage now has both an initialized-canary test and
+a separate strict-Miri test with genuinely uninitialized unreachable holes;
+the latter only inspects reachable slots. Scatter extrema coverage executes
+the uninitialized scatter path for i32 and i64 with repeated indices under
+Serial and bounded 1/2/4-thread contexts. The lifecycle coverage executes
+gather, dynamic slice, dynamic update, scatter, reduction, and the
+copy-then-update receipt path; the large threshold benchmark remains outside
+the focused Miri filters.
+
+Fresh coverage: `53/53 files passed` using:
+
+```text
+cargo llvm-cov --workspace --json --output-path coverage.json
+python3 scripts/check-coverage.py coverage.json
+```
+
+Affinity benchmark evidence, all candidate-relative 95% upper bounds at or
+below the 1.20 gate:
+
+```text
+t1 taskset CPU60 upper95: reduce 1.0169, gather 1.0415,
+  dynamic_slice 1.1028, dynamic_update 1.1966, scatter 0.7869
+t4 taskset CPUs60-63 upper95: reduce 1.0456, gather 1.0337,
+  dynamic_slice 0.9308, dynamic_update 1.1312, scatter 0.7856
+```
+
+Sol high focused Miri passed with strict provenance and symbolic alignment:
+
+```text
+MIRIFLAGS='-Zmiri-strict-provenance -Zmiri-symbolic-alignment-check' cargo +nightly miri test -p strided-kernel --test issue_187_uninit_indexed aligned_uninit_lifecycle_all_indexed_families
+MIRIFLAGS='-Zmiri-strict-provenance -Zmiri-symbolic-alignment-check' cargo +nightly miri test -p strided-kernel --test issue_187_uninit_indexed uninit_lifecycle_executes_dynamic_and_scatter_families
+MIRIFLAGS='-Zmiri-strict-provenance -Zmiri-symbolic-alignment-check' cargo +nightly miri test -p strided-kernel --test issue_187_uninit_indexed dynamic_update_uninitialized_holes_are_never_read
+MIRIFLAGS='-Zmiri-strict-provenance -Zmiri-symbolic-alignment-check' cargo +nightly miri test -p strided-kernel --test issue_187_uninit_indexed scatter_integer_extrema_wrap_in_uninit_path
+MIRIFLAGS='-Zmiri-strict-provenance -Zmiri-symbolic-alignment-check' cargo +nightly miri test -p strided-kernel --test issue_187_uninit_reduce axis_holes_negative_stride_and_identity_match
+MIRIFLAGS='-Zmiri-strict-provenance -Zmiri-symbolic-alignment-check' cargo +nightly miri test -p strided-kernel --test issue_187_uninit_reduce uninit_reduce_product_and_nonfinite_match_initialized_replay
+MIRIFLAGS='-Zmiri-strict-provenance -Zmiri-symbolic-alignment-check' cargo +nightly miri test -p strided-kernel --test issue_187_uninit_reduce uninit_reduce_simd_tail_and_sum_squares_match_initialized_replay
+```

strided-kernel/Cargo.toml

@@ -69,3 +69,8 @@ required-features = ["parallel"]
 name = "issue_184_uninit_replay"
 harness = false
 required-features = ["parallel"]
+
+[[bench]]
+name = "issue_187_uninit_indexed"
+harness = false
+required-features = ["parallel"]

strided-kernel/benches/issue_184_uninit_replay.rs

@@ -273,12 +273,14 @@ fn bench_fused(context: ExecContext) {
         "fused_add_mul",
         || {
             plan.execute(&context, &mut initialized, &refs).unwrap();
-            black_box(initialized.data()[0]);
+            black_box(initialized.data_as::<f64>().unwrap()[0]);
         },
         || {
             plan.execute_uninit(&context, &mut uninitialized, &ptrs)
                 .unwrap();
-            black_box(unsafe { uninitialized.data_mut()[0].assume_init() });
+            black_box(unsafe {
+                uninitialized.data_as_uninit_mut::<f64>().unwrap()[0].assume_init()
+            });
         },
     );
 }