Response to

Reviewer MjTC

• Noise or dropout on clean bank latents
• Co-visibility ignores occlusion; multi-room scenes
• O(n) scoring cost

Noise or dropout on clean bank latents

We had already tried it once: noise up to sigma 0.5 on 85% of samples raised seen Revisit slightly (0.405 -> 0.42-0.44) but cut quality (0.560 -> 0.49), and sweeping sigma at inference changed nothing, which suggests the noise was too strong. [PENDING: G1] a milder variant (sigma <= 0.2, half the samples clean) at 200 s: [fill in Revisit, VBench quality vs the plain 200 s model].

Co-visibility ignores occlusion; multi-room scenes

Correct: frustum overlap measures containment, not visibility, so two rooms separated by a wall can look overlapping. We list this in Limitations (Appendix A.4). We also measured it. The UE capture has no usable depth, so we used monocular depth (Depth Anything 3) with the known poses on 399 chunks sampled as in training (spans up to 200 s). Among each chunk's top-6 retrieved frames, 24-49% of the content that falls inside the frustum is occluded (depending on the depth tolerance, tau = 0.25 to 0.1), versus 38-65% for random frames that pass the same frustum test; 79-94% of the chunks that have a frustum hit have at least one retrieved frame with visible overlap >= 0.3. This is indirect evidence only: the depth is monocular, not ground truth, and tau is our choice. [VERIFY wording before posting] Retrieval also uses pose distance with a weight, which discourages far-away frames.

O(n) scoring cost

Correct, and we have clarified that only attention cost is constant. Measured: 32 ms per chunk at 400 s of history, 70 ms at 800 s, on one CPU thread; it is about 1% of the 3.2 s chunk time (measured above). A spatial index would make it sub-linear; we have not needed it at these lengths.