EPS Dashboard

Overview / Motivation

Auto-filed by the /daily 2026-07-28 problem sweep (transcript miners over the day's 55 session transcripts). Sources: group-D P1, group-G P1, group-A P2(a), group-J P6, group-F P1 (5 miners, 3 issues).

Goal

Fold five measured 2026-07-27/28 sizing failures into plan-compute-sizing.md (+ the vectorize rule's mid-run trigger).

Workflow gap

Bug observed: (1) #1689 fit_ladder: the 200-draw bootstrap-CI battery was ~99.6% of a 176 GPU-h phase (plan sized 4h; ~30+h spent before the user ordered 'drop bootstrap to 0', ETA 38h→5-6h). (2) The same phase's MEASURED pilot extrapolated ~11x low — the evals-per-pair multiplier was assumed arithmetic, not read off the fit script. (3) #1739: two kernel OOMs (163GiB/170GB) because RAM was sized on an anchor behavior while the hallucination lane's table ran 2.2x larger. (4) #1738: the plan's capture-cost basis was ~4x stale post-pilot with no recorded basis update. (5) #1738's near-dupe screen: pairwise similarity scales with pool^2; the pilot ran below production pool size (4h plan → 20-80h projection, ~5.75h burned), and CPU%-only health reads called the frozen serial screen healthy for hours. (Cost figures are transcript-mined/marker-quoted — verify at plan time against #1689/#1739/#1738 events.)
Why it is a workflow gap: the sizing rule mandates a measured 1-cell pilot but not (a) that the per-cell multiplier be read off the code, (b) draw-count necessity, (c) largest-cell RAM keying, (d) a recorded basis update on pilot deviation, (e) pool-scale pilots for quadratic batteries; the mid-run >=2x trigger forces only the vectorize check (vectorize-many-cell-fits.md:265).
Confidence (emitter): medium-high (each incident marker-backed; exact multipliers transcript-mined)
verified-at-filing: grep -n 'Per-cell fit phases' .claude/rules/plan-compute-sizing.md → line 357 (measured-pilot clause present, no inner-loop-count derivation, no draw-necessity clause — 0 grep hits for 'descope'/'necessity'); grep -n 'deviation' .claude/rules/vectorize-many-cell-fits.md → mid-run trigger forces the vectorize signature check only (line 265) (2026-07-29 UTC).

Proposed change (candidate diff sketch — refine in planning)

Five short additive clauses at the named anchors (§ Per-cell fit phases; § Mid-run trigger; the RAM/RSS routing bullet); plus one monitoring line (output-growth over CPU%) in plan-compute-sizing or gotchas.

Scope / surfaces

Primary targets: .claude/rules/plan-compute-sizing.md, .claude/rules/vectorize-many-cell-fits.md

Constraints / invariants

Workflow-surface only — never experiment code, configs/, or tasks/.
scripts/workflow_lint.py --check-asks passes; ruff on touched files passes.
The spawned session runs under a workflow_fix_target: Provenance line — recursion guard applies (it parks, never auto-routes, its own subagents' workflow-fix candidates).

Provenance

fingerprint: 5b9e2c8a123d
workflow_fix_target: .claude/rules/plan-compute-sizing.md