[Eval track] E1 — acceptance-length eval harness: EvalConfig + EvalCache + gates#639

Draft

maocheng23 wants to merge 19 commits into

dataflow-up-30-e0-layoutfrom

dataflow-e1-eval-harness

Commits on Jul 1, 2026

[DataFlow runtime] Phase D — training managers (no_sync, full resume, checkpoint/eval)

maocheng23
and
claude
committed
[DataFlow runtime] E0 — layout consolidation: runtime/ is substrate-only (move-only)

maocheng23
and
claude
committed
Merge branch 'dataflow-up-28-colocated-lightweight' into dataflow-up-29-training-managers
maocheng23
committed
[Phase D review fixes] per-rank resume state, data-stream seek, DP-safe Evaluator, durable best tracking

maocheng23
and
claude
committed
Merge branch 'dataflow-up-28-colocated-lightweight' into dataflow-up-29-training-managers
maocheng23
committed
[Phase D review fixes 2] rank-agreed best verdict, sample-based resume position, bound comm device

maocheng23
and
claude
committed
fix(eval): normalize loss/accuracy to 0-dim before accumulating

maocheng23
and
claude
committed
Merge branch 'dataflow-up-29-training-managers' into dataflow-up-30-e0-layout

maocheng23
and
claude
committed
[Eval track] E1 — eval harness: EvalConfig identity + on-disk EvalCache + gates

maocheng23
and
claude
committed
fix(e1): token-weighted scalar accuracy, injective cache key, relocatable pointers, retained eval stream

maocheng23
and
claude
committed
fix(eval,checkpoint): token-weighted scalar accuracy; relocatable best/latest pointers

maocheng23
and
claude
committed
Merge branch 'dataflow-up-29-training-managers' into dataflow-up-30-e0-layout
maocheng23
committed
Merge branch 'dataflow-up-30-e0-layout' into dataflow-e1-eval-harness
maocheng23
committed

Commits on Jul 2, 2026