refactor: share propagation between sam3_track_frame and sam3_propagate_frame

[PABannier]

What / why

sam3_track_frame and sam3_propagate_frame repeated the same code almost line for line: propagating active masklets, propagating pending masklets (including memory encoding and storing the object pointer), memory encoding for active masklets, sam3_update_tracker, the add_mask_to_result lambda, and the resolve_overlaps / fill_holes / remove_sprinkles post-processing. This PR moves that shared code into static helpers that both public functions call. sam3_propagate_frame is now 9 statements.

Helper split

• sam3_propagate_masklets(tracker, state, model, masklets, fi, pm, po): propagates each masklet that has memory, updates score / last_seen / MDS, and writes the binarized mask into pm. With po set (active masklets), the raw outputs are kept. With nullptr (pending masklets), memory is encoded right away, one masklet at a time as before.
• sam3_store_propagated_memory: sam3_encode_memory + sam3_extract_obj_ptr_cpu + sam3_store_obj_ptr.
• sam3_commit_active_masklets: memory-encodes the active masklets in po, then calls sam3_update_tracker.
• sam3_add_mask_detection (was the lambda) and sam3_add_masklet_detections (loops over a masklet list and emits masks found in pm).
• sam3_finish_tracked_frame: resolve_overlaps, fill_holes, remove_sprinkles, then frame_index++.

Call order is unchanged in both functions. In sam3_track_frame, pending masklets are still memory-encoded before PCS detection, and new pending masklets are still memory-encoded before the active masklets. The no_obj_ptr F16/F32 read block has not changed at all. The fprintf and SAM3_LOG calls stay in the public functions, so __func__ in the logs is unchanged.

Verification

Parity tool (byte-compare against deterministic origin/main baselines; covers encode_image, segment_pvs, visual tracker add_instance + propagate_frame over 3 frames):

PARITY OK   edgetam_f16 gpu0
PARITY OK   edgetam_f16 gpu1
PARITY OK   sam2.1_hiera_tiny_q8_0 gpu0
PARITY OK   sam2.1_hiera_tiny_q8_0 gpu1

Full Release build: no errors, no new warnings. The only warnings come from ggml/stb and were already there.

sam3_track_frame is only compile-tested. There is no full SAM3 model locally. I checked its diff line by line against the original instead.

Existing discrepancy (kept as is, likely a bug)

In sam3_track_frame, the pending-masklet loop never wrote the propagated mask into pm. sam3_propagate_frame does. So in sam3_track_frame, pending masklets from earlier frames are never emitted in the result, and their all_tracked_masks entries passed to sam3_match_detections are empty. This also applies to masklets promoted to active on this frame. This PR keeps that behavior on purpose: track_frame passes a scratch pending_pm map, with a comment at the call site. Fixing it (passing pm) should be a separate PR.

Diff

 1 file changed, 118 insertions(+), 204 deletions(-)

🤖 Generated with Claude Code

https://claude.ai/code/session_01M2qLmQ9ag49hT61V6i9qAq