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-F P4 + P7 + P11.

Add three 2026-07-28 GCP/Hub seam traps to gotchas.md.

Bug observed: (1) #1738 Phase-4a: a kresample subsample drew from the SPLIT-time id list; 12 rows the capture phase had skipped (over-length cis + pilot-era no-primary rows) reached vLLM and killed the engine (add_request ValueError). The in-session lesson marker covers #1738 only. (2) Ad-hoc gcloud compute ssh/describe probes hardcoded us-central1-a; the router had landed eps-issue-1738 in us-central1-c — not-found reads were briefly misread as instance-gone. (3) HF path_in_repo was fed a glob (URL-encoded %2A visible in the 404) and a not-yet-existing prefix was listed — 4 tool_result firings across 2 sessions. Note #1778 (landed) covers verify_artifacts_exist's hf:// glob crash — this entry covers the ad-hoc-probe class.
Why it is a workflow gap: all three are re-hittable seams whose lessons currently live only in one session's markers/transcripts.
Confidence (emitter): medium (mechanisms read from the sessions' own failure text; the glob mechanism verified via the encoded %2A in the traceback)
verified-at-filing: gotchas.md has no producer-filter/zone/path_in_repo entries (grep at compose time found no matching entries; label: the spawned session re-greps gotchas.md before writing to avoid duplicating adjacent entries).

Three concise entries in the matching gotchas sections (vLLM; GCP; Hub-API).

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).