fix(v4): report expert-matmul phase in serve telemetry — web dashboard (#890)

[JustVugg]

#890: DeepSeek V4's "Turn wall time by phase" is 100% "other" — I/O and matmul both 0%.

Root cause (two layers): v4_prof_emit hardcoded every phase but disk to 0.000, with a comment claiming the runtime "has no per-phase instrumentation beyond disk-read time". That's not true — and it misses the real problem: on a box with enough RAM the model sits warm in page cache, so disk_sec ≈ 0 as well, and the compute phase (the actual time when the disk is warm) was never measured at all. Everything folded into "other".

Fix follows GLM's pattern (its m->t_emm timer): a counter on the shared store instance — not a per-unit global — so it works across the per-COLI_V4_UNIT builds without a linking problem. matmul_sec sits beside disk_sec in V4ExpertStoreState, accumulated around the expert forward in moe_token_pipeline, and read per-turn in v4_serve_one exactly like disk_sec. The dashboard is unchanged — it already parses these fields; V4 just fills them.

Verified:

• builds clean across the unit split (the extern is declared in both the block unit that calls add and the serve unit that reads the getter — the warning I hit and fixed);
• deepseek-v4-tiny-check token-exact → the timing hooks don't touch model output.

Honest scope:

• Runtime confirmation of the non-zero matmul field on the live dashboard is best done on a box that can hold the 284B checkpoint (this dev box is RAM-starved as I write). The logic mirrors disk_sec, which already renders correctly, so the risk is low — @acedogblast, if you pull this, the Profiling tab should show I/O + matmul instead of all-other.
• attention still folds into "other" (it lives in a struct the store doesn't reach); a follow-up can add it the same way.

Refs #890 — will close it once the matmul field is confirmed live.

🤖 Generated with Claude Code