EPS Dashboard

Overview / Motivation

Filed from #1336 (2026-08-14). The memory-aware null-draw chunk cap that fixed a 72.24 GiB CUDA OOM in the SHARED pooled-ridge fit core exists ONLY on the issue-1336-fullcorpora branch. Any pooled fit run launched from main still carries the OOM.

Goal

(a) Land the fit-core memory fix on main correctly, and (b) decide the disposition of PR #1717, whose branch is 12,580 commits behind with ~100 workflow-surface files diverged on both sides.

Workflow gap

Observed: scripts/issue825_fit_cells.py on main lacks _resolve_null_draw_chunk and _free_device_bytes (0 grep hits each). The branch copy adds +212/−6 lines implementing a memory-aware draw-axis chunk cap that bounds peak device memory independent of n_train and λ-grid length. Verified on the live run (SLURM 12874): [fit825] null-draw chunk resolved: step=3 requested=64 draws=20 unit_gb=4.91 free_gb=130.8 factor=6 safety=0.8 dev=cuda:0, and all 5 previously-OOM-killed units completed.
Why it matters beyond #1336: this fit core is imported by #825 and siblings. Without the fix, a pooled/off-arm-shaped fit from main reproduces the failure — a 72.24 GiB allocation against a 139.8 GiB H200 at n_train=149964, d=4096 with 20 null draws.
Why it cannot be a copy: main's own copy has diverged +9/−2 since the merge base 9800ea4a6a, so overwriting with the branch version would revert those. Needs a genuine 3-way merge of the single file, plus the repo-root inline payload lint gate (scripts/*.py), the branch's 5 accompanying tests (tests/test_issue1336_null_chunk.py), and a code review.
Numerics constraint: chunk ≥ 2 must stay bit-identical to unchunked (array_equal-pinned on the branch); chunk = 1 differs at the last fp64 ulp (~4.4e-16, allclose-pinned). The cap's clamp floor is 1 by design and a shared-tenant GPU can select it, so the invariance pins must come across with the fix.
PR #1717 disposition (the second half, needs a decision): branch issue-1336-fullcorpora, 83 branch-only commits, GitHub mergeable: CONFLICTING / DIRTY, 12,580 commits behind, and 101 of 102 branch-side .claude/ changes touch files main also changed (269 main-side). Three defensible options: full merge WITH an explicit workflow-file resolution plan; surgical cherry-pick of the code-only commits (leaving the workflow files to main); or close the branch as artifact-complete, since every artifact #1336's promoted clean-result rests on is already on main (f22b6dc90d, 06e841e479, 49a43b6fe2, eff921f7d8, 3ed54c4e6a, plus the separately-ported width-cap passthrough 50b22bff84). Option 1 risks silently reverting other sessions' workflow edits across ~100 fleet-loaded agent/rule/skill files.
Confidence: high — every figure above was read directly from git this session.

Proposed change (refine in planning)

3-way merge scripts/issue825_fit_cells.py's chunk-cap addition onto main's current copy; bring tests/test_issue1336_null_chunk.py (5 tests incl. the chunk-size-invariance pin and the production-shape memory accounting pin at n_train=149964, d=4096); audit the 26 heldout_r2_sweep consumers for numeric inertness as the branch round did; clear the inline payload lint gate; code review. Then execute the chosen PR #1717 disposition. Do NOT attempt a blind rebase-merge of the branch.

Scope / surfaces

Primary: scripts/issue825_fit_cells.py, tests/test_issue1336_null_chunk.py
Decision surface: PR #1717 / branch issue-1336-fullcorpora
Evidence: #1336 markers v297 (two-mode diagnosis), v302 (both fixes), v311 (merge state)