[jmh530]

Fixes #494.

Repeatedly adding or subtracting a two-value pairwise Summator overruns its partial-sum buffer on the 65th merge. The merge inserts source partials through put, then incorrectly adjusts the destination counter to count the source’s original observations. This leaves the counter inconsistent with the destination’s binary tree.

Keep the counter maintained by put, treating each inserted partial as a leaf in the destination tree.

Also preserve the source during self-addition and self-subtraction. Inserting partials can otherwise overwrite values that the same merge has yet to read: an accumulator containing 1, 2, 3 produces 18 rather than 12 after s += s. Copy the source only when it aliases the destination; ordinary merges retain the existing traversal without a copy.

Public signatures and other summation algorithms are unchanged.

Regression coverage

Add @safe pure nothrow @nogc tests for float, double, and real, covering:

• Repeated accumulator addition and subtraction beyond the original failure threshold.
• Empty and varied-size batches, with scalar insertions between merges.
• Exact expected sums, source preservation, and the binary-tree state invariant.
• Self-addition and self-subtraction followed by further insertion.

Validation

• Confirmed the buffer overrun and incorrect self-addition against unchanged source.
• Focused optimized regressions pass on Windows DMD 2.113 and LDC 1.43, including DIP1000.
• Windows DMD unoptimized module harness: 27 modules passed.
• Windows LDC optimized module harness with DIP1000: 27 modules passed.
• Focused optimized regressions pass on Linux LDC 1.43 and 1.28.1.
• Diff whitespace checks pass.

Two broader optimized harness failures also reproduce with unchanged source: Windows DMD fails the reduce test at iteration.d:914, and Linux LDC fails the existing precise-summation residual assertion. Neither is changed here.

No full repository suite or performance benchmark was run. Correcting the tree can change floating-point grouping after merges; bit-for-bit agreement with the previous invalid merge state is not promised.

AI-Assisted-By: Codex (GPT-6 Astra Medium)

[jmh530]

Added two follow-up commits:

• b60588e adds the seven-element-source regression: merging it twice into a distinct accumulator previously underflowed the partial-buffer index. Coverage includes addition and subtraction for float, double, and real.
• 90238ee adds a fast path for sources with one stored partial. The value is read before insertion, so even self-merging needs no snapshot. Multi-partial self-merges retain snapshot protection.

Local Windows validation passed with DMD and LDC, including the 27-module test harness and additional checks for compile-time evaluation, cancellation, infinities, NaNs, and self-merges against explicit snapshots.

Local optimized LDC microbenchmarks showed that the fast path substantially reduces the self-merge overhead. In the tested double-addition cases, repeated ordinary merges were about 2% slower than the original implementation, while one-partial self-merges were slightly faster. Results vary by compiler and workload; these are not application-level performance guarantees.

A final branch review found no further actionable issues. Smaller snapshots were also investigated, but their mixed performance did not justify additional changes.