speed: apply speed_max_val to speed_temporal - #1668

speed_temporal declares speed_max_val ("maximum value allowed; larger values will be clipped to this value", default 1000, range 0 to 1000) and stores it in speed_temporal_max_val, but extract() appends the raw score and never reads it. speed_chroma applies the same option with MIN(score, speed_chroma_max_val), so the option works on one extractor and is silently ignored on the other.

What breaks today

speed.c:1643-1645 on master 9e48141b:

err = vmaf_feature_collector_append_with_dict(
    feature_collector, s->feature_name_dict,
    "Speed_temporal_feature_speed_temporal_score", score, index);

speed_temporal_max_val has no other use in the file (grep speed_temporal_max_val speed.c: the struct member and the option offset).

Reproducer, GCC 16.2.1 release build, src01 576x324, frames 1 to 3 (identical under --cpumask -1):

vmaf -r src01_hrc00_576x324.yuv -d src01_hrc01_576x324.yuv -w 576 -h 324 -p 420 -b 8 --no_prediction --feature speed_temporal=speed_max_val=3 -q --json -o /dev/stdout

Fix

MIN(score, s->speed_temporal_max_val) on the appended value, the expression speed_chroma uses. Scores below the bound are unchanged, which is every score at the default of 1000.

Test

test_speed_temporal (new): 256x256 noise against weaker noise through the extractor (vmaf_feature_extractor_context_*), scored with the default bound (the fixture must exceed 0.5) and with speed_max_val=0.5 (the score must be exactly 0.5). The second assertion fails on master ("speed_max_val must clip the score") and passes with this change.

Validation

x86-64 Linux, GCC 16.2.1, master 9e48141b, branch head 21f913f1; -Denable_float=true -Denable_checkasm=true.

• Release meson test: 26/26 here, 25/25 on master (the extra test is the new one).
• -Db_sanitize=address,undefined -Db_lto=false: 23 pass and 3 fail of 26 here, 22 pass and 3 fail of 25 on master; test_predict and test_pic_preallocation (SIGABRT from LeakSanitizer) and checkasm (heap-buffer-overflow in adm_dwt2_16) fail on both and are untouched by this change; test_speed_temporal passes.
• The three Netflix pairs, .engagement/golden_compare.py master binary against this branch: ALL IDENTICAL (default dispatch and cpumask=-1).
• No shipped model uses speed_temporal (the vmaf_v1.0.16* models set speed_max_val=45 for speed_chroma only). speed_temporal_feature_extractor_test.py asserts 18.9411808 and 16.721622 with the default bound; they are below it and unchanged. The Python suite itself was not run.

Not covered

• Frames below the size SpEED needs; a 128x128 noise pair already gives a singular covariance matrix and a score of 0 on both builds, so the test uses 256x256. That is a separate matter (feature/speed: fail init on frames too small for one SpEED block #1620).