agg_multimodal_router.sh launched each worker in the background but
then blocked on its health endpoint inside the same loop, so two
workers booted serially even though they could share the GPU. Move
wait_ready outside the launch loop so phase 1 fires all workers and
phase 2 waits for all of them concurrently.

Combined with this, promote the qwen3-vl-2b agg_router topology to
pre_merge. Its cached_tokens > 0 assertion fails closed when MM-aware
routing silently degrades to text-prefix — catching the same class of
regression the post_merge tests/mm_router/* tests do, but in pre_merge.

Wall-clock in mm-vllm-dev (RTX 5880 Ada, --kv-cache-memory-bytes-capped
profile matching CI): 66s pass time / 72s real (was 131-142s in the
last main post_merge run).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…rge-gater

…rge-gater

Shaves ~10s off mm_agg_router_qwen3-vl-2b in CI (2 requests × ~5s
generation). All assertions are prompt-side, so output length is safe
to halve.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…o pre_merge

agg_multimodal_router_chat_processor.sh had the same serial-boot pattern as
the lightseek script (wait_ready inside the launch loop). Mirror the
Phase 1/Phase 2 split so the chat_processor sibling gets the same speedup,
and promote mm_agg_router_chat_processor_qwen3-vl-2b to pre_merge so the
chat-processor MM-routing path is gated before merge alongside lightseek.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Both lightseek and chat_processor mm_agg_router tests peak at 14.9 GiB
(measured at 0.5s nvidia-smi polling locally + 15.0 GiB at 10s polling
in CI run 26596313542). 18.7 was conservative; 16 keeps 1 GiB safety
headroom while freeing the scheduler to pack smaller co-located tests
alongside instead of reserving the whole GPU0 slot.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Switch the 3 agg_router topologies (lightseek pre_merge, chat_processor
pre_merge, frontend_decode post_merge) from Qwen3-VL-2B-Instruct (bf16)
to Qwen3-VL-2B-Instruct-FP8. The router is precision-agnostic — it
operates on token IDs, block hashes, and counts, not float values — so
FP8 produces bit-identical routing behavior (verified: same cached_tokens
0/272/272, same kv_hit_rate 0.944, same 17/18 block overlap on R3).

Drop profiled_vram_gib 16.0 → 13.0 (measured peak 12.16 GiB + 0.8 GiB
margin) and requested_vllm_kv_cache_bytes 1.6 GB → 512 MB (test needs
<100 MB; smaller cap frees the scheduler's accounting without affecting
real peak which is activation-dominated).

Net effect: ~5 GiB more co-scheduling headroom on a 22 GiB L4 vs the
original 18.7 GiB marker, no coverage loss.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The previous commit's sed range matched agg_router topologies across
all VLLM_MULTIMODAL_PROFILES, not just qwen3-vl-2b. Restore the original
profiled_vram_gib and requested_vllm_kv_cache_bytes for
qwen2.5-vl-3b, qwen2-vl-2b, and phi3-vision agg_router topologies.
qwen2-vl-2b's profiled was dropped 16→13 without verification.

Only qwen3-vl-2b's 3 agg_router* topologies should carry the FP8 swap
and tightened VRAM markers measured in this PR.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The chat_processor variant doubles GPU0 queue time at 4-worker scale
without catching distinct bugs — both lightseek (Rust frontend) and
chat_processor (Python preprocessor) share the kv_router downstream,
so a routing regression would surface in either. Keep the lighter
lightseek-path test as the pre_merge gate; chat_processor stays in
post_merge.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>