[jak-pan]

Keeps FLOAT and DOUBLE range bounds exact when the optimizer moves an inverted-index condition to a forward filter. An exclusive comparison then continues to exclude documents equal to its bound. Closes #817.

Problem on main

On d7b478df9843ddfda23a0021ec09f16c105835a5, price > 1.0000001 and weight > 1.0000000001 each return 1,001 documents in the reproduction instead of 951. The collection contains 50 documents at the bounds and 951 above them, split into a sealed segment of 1,000 and a writing segment of one.

The optimizer converts the binary bound back to text before the forward filter parses it again:

zvec/src/db/sqlengine/planner/optimizer.cc

Lines 130 to 135 in d7b478d

	// 2. Revert numeric buf to numeric text
	std::string numeric_text{""};
	if (QueryInfoHelper::data_buf_2_text(right->text(), data_type,
	&numeric_text)) {
	right->set_text(numeric_text);
	}

data_buf_2_text() uses StringHelper::ToString() for FLOAT and DOUBLE:

zvec/src/db/sqlengine/analyzer/query_info_helper.cc

Lines 106 to 114 in d7b478d

	if (data_type == zvec::DataType::FLOAT) {
	*text = ailego::StringHelper::ToString(*(float *)data_buf.data());
	return true;
	}
	if (data_type == zvec::DataType::DOUBLE) {
	*text = ailego::StringHelper::ToString(*(double *)data_buf.data());
	return true;
	}

Those overloads use std::to_string(), which prints both reproduction bounds as 1.000000 in this build:

zvec/src/include/zvec/ailego/utility/string_helper.h

Lines 164 to 172 in d7b478d

	//! Convert floating-point number string (double)
	static std::string ToString(double val) {
	return std::to_string(val);
	}
	//! Convert floating-point number string (float)
	static std::string ToString(float val) {
	return std::to_string(val);
	}

Changes

• Print FLOAT and DOUBLE bounds with %.*g and std::numeric_limits<T>::max_digits10 significant digits: 9 for float and 17 for double. For example, 1.0000001f prints as 1.00000012, which parses back to the same float.
• Add FloatRangeFilter.ForwardScanKeepsExactBound, the collection-level reproduction from the issue. Both filters must return 951 documents.

Tests

Local Release verification on macOS 27.0.1 arm64, Apple clang 21.0.0, CMake 4.3.2. Base production sources matched d7b478df9843ddfda23a0021ec09f16c105835a5; fixed production sources included only this PR's helper change. The regression test was added to both builds.

Test	Base	This PR
ForwardScanKeepsExactBound	0 passed, 1 failed	1 passed, 0 failed
Seven existing suites	274 passed, 0 failed	274 passed, 0 failed

The regression fails on the base because each query returns 1,001. With the fix each returns 951.

Commands (binaries under build/bin):

cmake --build build --parallel 6 --target collection_float_range_filter_test optimizer_test simple_rewriter_test query_info_test sqlengine_test forward_recall_test invert_recall_test collection_test
build/bin/collection_float_range_filter_test --gtest_filter=FloatRangeFilter.ForwardScanKeepsExactBound
for suite in optimizer_test simple_rewriter_test query_info_test sqlengine_test forward_recall_test invert_recall_test collection_test; do
  build/bin/"$suite"
done

Both builds use the compiler cache and six jobs. Tests run serially. Changed-line formatting passes Apple clang-format 21.0.0 with the repository configuration; git diff --check passes.

Compatibility and cost

From code reading: the integer branches and public StringHelper::ToString() overloads keep their behavior. The conversion is shared by the optimizer and the constant-node conversion helper. A source search found no caller of that constant-node helper.

Each converted FLOAT or DOUBLE bound replaces one decimal conversion with snprintf into a 32-byte stack buffer. The optimizer's choice of scan and engine operations are unchanged. This adds no persistent state, on-disk format, dependency, or public API.

Not covered

• Linux, other architectures, Python tests, and the full repository suite were not run. Verification covers the regression and the seven listed C++ suites.
• The pre-commit formatter pins 18.1.8; only Apple clang-format 21.0.0 was available locally.
• delete_by_filter("price > 1.0000001") on this data fails earlier, on both base and fixed sources, with ExecBatch::ToRecordBatch mismatching schema size. That is a separate defect; this PR does not change it.
• No public CI result is available before posting.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-16T07:47:06.699468Z	c66ff5c	New commits

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[egolearner]

LGTM

Thx for the fix.

[jak-pan]

The CI fail doesn't look like coming from me is it?

[egolearner]

The CI fail doesn't look like coming from me is it?

Merged.