Hub-suppression on/off ablation (close the v2 open question)

[nikolay-e]

Context

Promised in v2, not delivered. Known incoherence: in-degree damping exempts Semantic edges — the exact channel through which canonical hubs (Logger/Config/Utils) accumulate mass; only the out-fan rule partially covers them. Whether the net configuration suppresses hub dominance is empirically open.

Tasks

• On/off cells: in-degree damping, out-fan rule, both, neither — at B=8000 on the frozen manifest
• Per-cell telemetry: mass share of top-1% in-degree nodes
• Decide the deployed configuration from the data; document the exemption rationale or remove it

Acceptance

• ablation table + decision committed; paper v3 section replaces the "open question" wording

[nikolay-e]

Planned as the immediate next paper cell — the only v2-protocol ablation the v6full sweep did not cover. Runs on the ready template: python -m eval run-final --winner results/final/v6_arms/new.json --manifests-dir datasets/eval-splits/v1 --extra-env DIFFCTX_HUB_SUPPRESSION=0 ... (exact env knob to be confirmed against config/graph_filtering.rs before the run; if suppression has no env override yet, a one-line read_env_* addition lands first). Compare against results/final/v6full/ego_grid_new b8000 rows, paired. ~30 min of compute.

[nikolay-e]

Deferred to post-v3 (#256): not one of the 16 matrix cells in the v3 plan. v2's promise stays open — reinstate as an appendix cell if the cycle has room.