planner: fix model-name casing and kv_hit_rate namespace so engine metrics resolve

[Pernekhan]

Two fixes that together make the planner's Prometheus metrics resolve. Both were silent — no error, just n/a on every tick.

1. Model-name casing (worker_info.py)

resolve_worker_info preferred the MDC-derived model name over the operator-supplied config.model_name. Discovery normalizes that name to lowercase, but engine metrics carry the model's real casing:

model=deepseek-ai/DeepSeek-V4-Flash-0731-roce-disagg -> accept_length 3.67
model=deepseek-ai/deepseek-v4-flash-0731-roce-disagg -> None

One character-case difference emptied every engine-metric query. _clamp_accept_length then falls back to 1.0, so the ITL estimator divides forward-pass time by 1 instead of the real ~3.9 — inflating the estimate ~4x (265ms against a measured 39ms), breaching the ITL SLA every cycle and pinning the planner at SCALE_UP.

Fix: config model_name wins, and the log names both when they disagree.

2. kv_hit_rate namespace spelling (traffic_metrics.py)

get_avg_kv_hit_rate sanitized the namespace to underscores unconditionally. The router emits it that way, but some scrape paths relabel it back to the hyphenated k8s form (nvidia.com/dynamo-namespace):

dynamo_namespace=deepseek_ai__DeepSeek_V4_Flash_0731_roce_disagg -> EMPTY
dynamo_namespace=deepseek-ai--DeepSeek-V4-Flash-0731-roce-disagg -> 5 series

DeepSeek-V3.2 carries both forms, so neither spelling is universally correct. Fix tries sanitized first, then raw, and names both in the miss log.

Verified live

before:  kv_hit_rate: n/a     accept_length: n/a     summary: SCALE_UP (every cycle)
after:   kv_hit_rate: 0.362   accept_length: 3.079   summary: HOLD
         "No prometheus data" warnings: ZERO

ITL estimate went 265ms -> 37.9ms against a measured actual of 38.9ms.

Notes

• Based on a3e3a561, the commit the deployed dynamo-planner image is built from — not main — so it applies cleanly to what actually runs.
• Shipped as a thin overlay image (FROM <planner image> + COPY), since the planner is pure Python.
• A third, companion fix lives in deepinfra/backend PR [BUG]: dynamographdeployment operator didn't handle multinode spec update operation ai-dynamo/dynamo#4749 (the planner config was also missing namespace entirely). All three are needed.

🤖 Generated with Claude Code

https://claude.ai/code/session_017rS1uFFKT2xV4x63Amfvtw