fix(trainer-rank): price CP2 recompute using each rank’s layouts

[bradhilton]

Agent: Dalinar

At CP2, the two ranks can hold different attention and GDN layouts. Price each rank’s saved boundaries, recomputed mixer, retained attention state and MoE stage from its own rows, then take the larger floor. Pair #1002’s pending adapter-gradient term with that same rank’s floor. This avoids combining one rank’s gradients with the other rank’s peak, which overcounted a short-sequence case by up to 632 MB.

Scope is CP2/TP1 gradient groups using ART’s CP attention without a softmax offset; other configurations retain the old floor. Building the peer layout costs about 40–76 ms initially and under 1 ms from cache, now counted in planning time. Reports retain the layouts for CPU replay.

Recorded trainer-rank, CP and replay tests passed. Local two-H200 Qwen3.6-35B-A3B traces validated the executor mirror/layout ledger within 0.3%. The layout correction left measured peaks and losses unchanged and accuracy roughly unchanged: real-data EP2 raw error was +18.2% cold/+20.2% warm. On the full stack with #1002, cold EP1 errors were +0.9–2.0% and EP2 +3.0–13.4%, with no measured underestimates.

The executor itself has not been tested directly against its mirror; that needs GPUs. CP4 and other executor configurations retain the old floor. GDN width and routing allowance remain unchanged, and routing still accounts for much of the EP2 excess. Stacked on unmerged #963 and includes #1002; review #1002 and #963 first.