Reflection across y-axis: x→-x, output preserves min<max ordering.

tune subcommand is accessible under the 'tune' name.

--tracker, --gt-dir, and --detections-dir are required and parsed.

reset() must clear state and restart tracker IDs from zero.

After reset(), the very next estimate() call returns identity.

Cross-axis residual above 1e-6 must take the freeze branch.

Cross-axis residual below 1e-6 takes axis-aligned (P-update) branch.

Result contains all float summary, integer count, and array fields.

tune() returns 0 when Tuner.run() completes without error.

ImportError (e.g. optuna not installed) causes tune() to return 1.

Invalid tracker ID causes tune() to return exit code 1.

FileNotFoundError from Tuner (missing sequence files) returns exit code 1.

OSError while writing output file returns exit code 1.

Exception from tuner.run() causes tune() to return exit code 1.

xyxy → xywh → xyxy round-trip recovers the original box without drift.

xyxy → xcycsr → xyxy round-trip recovers the original 1-D box.

xyxy → xcycsr → xyxy round-trip preserves the original 2-D shape.

Baseline enqueue only includes tunable search_space keys, not fixed_params.

reset() must be called once per sequence per trial.

compute_hota_metrics produces expected HOTA fields across scenarios.

Test Identity metrics computation for various scenarios.

A tracker with search_space key not in __init__ raises ValueError.

Registered trackers' search_space keys are valid __init__ params.

SportsMOT / DanceTrack-style data (all pedestrian, conf==1) must be passed through unchanged, so existing parity is preserved.

Single-metric evaluation returns only the requested metric.

Single-sequence aggregation reproduces per-sequence HOTA scores.

Single-sequence aggregation preserves raw counts; recomputes IDF1.

Accessing deprecated .trackers must emit FutureWarning with correct message.

A single predict() step must not move the bbox of a freshly created tracklet.

--split values are split on commas and whitespace-stripped.

--split handles malformed comma inputs gracefully.

--split accepts comma-separated values.

--split with spaces around commas strips whitespace.

BoTSORTTracker constructs and updates with any supported estimator.

table() shows HOTA and DetA; MOTA absent when only HOTA computed.

IoU threshold controls whether a 0.4-similarity pair matches.

`predict()` advances `time_since_update` for unmatched tracks.

Unconfirmed tracks gain real IDs once minimum_consecutive_frames is reached.

A confirmed track is pruned after maximum_frames_without_update empty frames.

A track first seen on frame 2 must start with tracker_id == -1.

A confirmed track stays alive across a short detection gap.

Before the first update, no tracked objects are exposed.

Track is removed from tracked_objects once lost_track_buffer is exceeded.

After enough consistent frames the track is mature and visible.

Two mature, simultaneous tracks are both exposed with valid IDs.

Track stays in tracked_objects while absent from update() during occlusion.

Resetting one tracker instance must not make another instance reuse a live ID.

update() must not assign tracker_id on the caller's sv.Detections.

update() with empty detections must not mutate input.

Trackers should use the configured IoU implementation for matching.

Trackers with empty search_space skip validation.

Trackers without search_space are still valid.

predict() must increment age by 1 each call.

predict() must increment time_since_update by 1 each call.

predict() must always return finite bbox coordinates.

Construction counts as the first successful update for SORT-like counters.

Every tracklet must start with time_since_update == 0.

update(bbox) increments number_of_successful_updates for SORT-like counters.

update(bbox) must reset time_since_update to 0.

When t=None, only the linear part R is applied (velocity contract).

Unmatched low-confidence detection appears in update() output with tracker_id=-1.