[moe] Great 10T: ablate shared expert

[dlwh]

Description

TL;DR: Run the full 10T ablation for the shared expert as part of the great gate rather than relying on a single spot check.

Hypothesis or Goal

We want to know whether the shared expert still belongs in the promoted recipe under a stricter comparison bar.

Links

• Parent sweep: MoE Sweep: Hyperparameters, Routing, Architecture, and Optimizer #3469\n* Gate: [moe] Great 10T gate for #3469 #4014

Results

[github-actions]

🤖 I'm looking into this issue. I'll post a task list shortly with my plan and progress.

[claude]

🤖 Plan:

• Research codebase and check for duplicate work
• Post research + proposed fix comment
• Create branch and implement experiment script
• Write test
• Run pre-commit and tests
• Open PR
• Verify CI

[claude]

🤖 Plan:

• Research codebase and check for duplicate work
• Post research + proposed fix comment
• Create branch and implement experiment script
• Write test (5/5 passing)
• Run pre-commit and tests - all passing
• Open PR: [moe] Add multi-budget shared expert ablation for great 10T gate #4062
• Verify CI (checks queued, waiting for runners)