Summary

Fix a replacement omission in the fixed-size hashsig heaps and add a regression test.

Reproduction

For an input with more than 127 runs, hashsig_heap_insert() takes the full-heap replacement path. Before this change, it decremented h->size, moved the last element to the root, and sifted down, but discarded the newly accepted val. This leaves the heap short and causes later values to be admitted unconditionally.

A regression test creates 600 distinct runs and compares the same runs in forward and reverse order. Since the signature is order-independent, the expected similarity is 100. Before this change, the test reported 49.

Cause

The full-heap path evicted the root without storing the replacement value.

Fix

Store val at the root and sift it down without changing the heap size.

Impact

Large signatures now retain the intended 127 extrema. Similarity and rename-detection results no longer depend on the order in which runs arrive because of the dropped replacement values.

Validation

• Regression test before fix: failed, similarity 49 instead of 100.
• Targeted core::hashsig suite after fix: 3/3 passed.
• git diff --check: passed.
• Full ctest --test-dir build --output-on-failure: 67% passed (8/12 CTest groups). The four failing groups were environment/configuration-related: the build had no TLS backend, and the existing offline fixture reported a missing object plus an unrelated git_diff_options initializer mismatch. The changed hashsig suite passed.

Unverified

I did not run the online tests with a TLS backend configured.

Fixes #7375

Assisted-by: OpenClaw (opencode-go/gpt-5.6-luna)