[sophisid]

MESSI could return duplicate neighbors or partial squared-L2 distances, and raw SIFT exposed an additional adaptive-breakpoint error: a low positive PAA value was assigned to the highest SAX bin. That invalid lower bound pruned true nearest neighbors. On 20K SIFT vectors / 100 queries / k=10, the previous PR head returned 19–33 wrong top-k distances per run; one query returned a nearest distance of 27102 instead of 25379.

This fixes the affected exact-search paths:

• Deduplicate queued positions, retain the best refinement, and accept only strictly improving top-k candidates.
• Encode PAA values by searching the ordered cuts directly. Both tail bins, shifted cuts, repeated cuts and exact boundaries are handled without reading the previous cardinality's table or assuming Gaussian signs.
• Copy strided Python MESSI queries to contiguous float32 storage and reject mismatched dimensions / unbuilt indexes. Both the integer-k and SearchConfig APIs are covered; ordinary fvecs views previously misread row headers as vector coordinates.
• Preserve inclusive range pruning and continue refinement past partial sums equal to the radius, including radius zero.
• Guard the optional Coconut Python binding so BUILD_COCONUT=OFF produces an importable module.

Regression tests compare every returned distance and ID/distance pair against exhaustive ground truth and reject duplicates. No recall threshold below 100% is accepted. Exact ties permit any distinct IDs at the correct distances. Coverage includes Gaussian/adaptive modes, one/four workers, k=1/10/100/1000, repeated queries, streaming inserts, second-index isolation, SAX lower bounds, NumPy layouts, and inclusive range boundaries. tests/check_messi_sift.py provides a reproducible optional audit against FAISS IndexFlatL2 and the published SIFT1M ground truth.

Validation completed locally: isolated Release C++/Python build; all 21 focused CTest cases passed; 8 adaptive/shared-index cases passed; 5 shared-SAX Astronomy checks passed; Coconut streaming and range checks passed with Coconut enabled. On the original 20K SIFT reproduction, wrong distances and record lower-bound violations both fell to zero. Twelve real-SIFT range checks also passed with exact comparisons.

The full 100-query SIFT1M static/streaming audit completed: all 160 cases passed, checking 4,444,000 returned ID/distance pairs, with zero wrong distances, invalid pairs or duplicate rows. Full-dataset references match the published ground truth. Coverage includes static 20K/1M builds and streaming 750K→875K→1M, k=1/10/100/1000, both breakpoint modes, one/four workers and two query repetitions, using unchanged raw descriptors and header-strided fvecs queries. Independent streaming groups ran in parallel. The original validator's final two checks were rerun in a fresh process after swapped-memory I/O waits; completed cases were preserved and no performance claims are made.

The regular C++ regressions are covered by Simple CI; the Python layout test additionally requires BUILD_PYTHON=ON and NumPy. GitHub Release/Debug, dry-run and CodeQL checks passed; the complete Simple CI suite is still running.