[AzeezIsh]

Summary

vision-fidelity has been red on main: integrity 9/10 on multi-choice-image, "choice(s) [1] did not answer about the image". It fails on every vision model I ran: Qwen3.6-35B-A3B-FP8 (the gate subject), Qwen3.8-Flash-Next NVFP4 and FP4-Local. Main's last committed record (.benchmarks/vision-fidelity/2026-09-29-d5472821c8.json) passed 10/10.

Bisect (sparky GB10, spark benchmark run vision-fidelity --pull-request-gate, first-parent, 2026-10-21):

• d5472821c8: re-tested on this box first, 10/10, so the failure is not environmental.
• 1d33249896 good, f3ea6f0bb0 good.
• f95b729a09 bad, 50c6124e5b bad.
• First bad: b9a7edad8 (qwen3.8-flash-next: two-GB10 TP=EP=2 (Enntity series + RC16) with RDMA GID and cuBLASLt startup fixes #174).

Mechanism:

• chat_blocking::run_blocking_path sends one scheduler request per choice and gave the image pixels to choice 0 only. Choices 1.. carried no media and silently prefilled their image pad tokens from choice 0's leftover encoder rows.
• qwen3.8-flash-next: two-GB10 TP=EP=2 (Enntity series + RC16) with RDMA GID and cuBLASLt startup fixes #174's ep_broadcast_vision_state_for_seq(enabled = false) clears that per-request vision state, in single-rank serving too, for any request without media. That clearing is the right guard, since a request without media should never splice a stale image, so choice 1 now gets nothing.
• The server logs show it: on f3ea6f0bb choice 1's first token equals choice 0's (13602); on b9a7edad8 it does not (8725).

Fix: every choice carries its own pixels and is encoded on its own (crates/spark-server/src/api/chat_blocking.rs). The benchmark's doc comment (media_integrity::multi_choice) is updated to match. n=1 is unchanged.

Evidence

• sparky, vision-fidelity --pull-request-gate, default subject Qwen/Qwen3.6-35B-A3B-FP8, alternating, 180 s cooldown.
• main 6b09b40aa (binary sha256 94a1b8c0a3043dfe) vs this branch (4c7fd94187268ca3).

run	record	integrity	multi-choice-image	verdict
main #1	run-1792575629041714855	9/10	FAILED	FAIL
fix #1	run-1792575886530913767	10/10	both choices answered about the image	PASS (14 geometry cells, 3/3 probes, control held)
main #2	run-1792576144127996550	9/10	FAILED	FAIL
fix #2	run-1792576398694970202	10/10	both choices answered about the image	PASS

Test plan

• cargo fmt --all -- --check (1.93.1), cargo clippy -p spark-server --tests
• vision-fidelity on main vs fix, ×2 each (above)
• Required-gate campaign at the final tip

Generated with Devin