feat: GavelFGD+ — Placement-Opportunity-Aware LP + FGD scoring improvements

[Turquoise-T]

Overview

This PR introduces GavelFGD+, which improves upon base
GavelFGD through
coordinated changes at both the placement and allocation
layers.

Key Insight

Fragmentation is caused by placement decisions, not
allocation decisions.
The previous LP fragmentation penalty was architecturally
unsound: penalizing
allocation to fragmented GPU types cannot reorganize jobs
already placed on
those nodes. GavelFGD+ fixes this by replacing the
backward-looking penalty
with a forward-looking Placement-Opportunity-Aware (POA)
bonus.

Changes

Placement layer:

• Paper-accurate FGD sigmoid scoring
• Popularity threshold raised to 95% (more aggressive
  packing)
• Buddy tiebreak for equal-score placement options

Allocation layer (POA LP bonus):
After each round's lease-extension phase, compute per-type
fit fractions:
poa[type][demand] = fraction of nodes with ≥ demand free
GPU slots

LP objective becomes:
max min_j(throughput_j) + μ · Σ_{j,t} poa[t][demand_j] ·
x[j,t]

This steers allocation toward GPU types where jobs can
currently be placed,
creating a correct feedback loop between the allocation and
placement layers.
The bonus naturally vanishes at high utilization when fit
fractions → 0
everywhere, reverting to pure Gavel fairness under
saturation.

Results (Alibaba cluster, 7 load levels, 2 seeds)

At 85.4% GPU utilization:

• Frag rate: 2.01% (GavelFGD) → 0.47% (GavelFGD+), 76.7%
  reduction
• JCT: 22,275s → 21,810s, 2.1% improvement

Effect is load-dependent by design: dormant at low
utilization (POA scores
uniform everywhere), strongest at high utilization (POA
scores differentiated).