[Altren]

This change uses the reported file size as an allocation hint and reads directly into the buffer. When the size is unavailable or more data follows, it continues in 64 KiB blocks without seeking. Binary and text modes remain distinct, including native text conversion; the hint is not treated as a guaranteed byte count.

The shared save helpers now check stream state after closing, returning an error for failed writes or buffered-output flushes instead of reporting success. Reader stream errors and directories also return errors. Public APIs remain unchanged.

Measurements

Measurements compare upstream main (bb0c1dcb) with the new shared file loaders. Both sides use the same parser and serializer code. Apple Clang 21, macOS arm64, C++20 Release -O3 -DNDEBUG, no LTO. Six process runs with rotated/reversed case order; at least 0.1 s warmup and 0.25 s measured CPU time per case. Tables report median wall-clock milliseconds, including file open/read/close, allocations, and result destruction. Files are in the OS cache; these are not cold-disk throughput measurements. Fixture generation and equality checks are outside timing.

File loading only

Case	File bytes	Main, ms	New, ms	Speedup
Binary/1024	1,024	0.0122	0.0123	—
Text/1024	1,024	0.0127	0.0124	1.03×
Binary/1048576	1,048,576	2.0545	0.1486	13.83×
Text/1048576	1,048,576	2.4888	0.1512	16.46×
Binary/33554432	33,554,432	66.7553	5.0514	13.22×
Text/33554432	33,554,432	79.5736	4.9843	15.96×

Complete file load and deserialization

Case	File bytes	Main, ms	New, ms	Speedup
Cereal/Dataset	36,008,009	88.8233	23.1746	3.83×
YAS/Dataset	36,008,015	88.0490	22.5152	3.91×
JSON/Dataset	70,124,332	214.4066	57.6431	3.72×

Dataset contains 500 groups, each with 3,000 points (double x, y) and 3,000 uint64 IDs. Both loaders reproduce every original field.

Validation

• Added BM_load_file cases to the all-format benchmark target for binary/text files at 1 KiB, 1 MiB, and 32 MiB; all six compiled and ran.
• A Linux-only /dev/full test checks buffered output failure at close. Linux and Windows were not run locally; the text-mode equivalence test is portable and intended to exercise Windows newline conversion in CI.

[Altren]

The only test failure is an apt update issue