EPS Dashboard

Goal

Clear the one bare HF Hub call on origin/main that keeps workflow_lint.py --check-live-hf-retry-routing (and therefore the no-flags default run, and therefore tests/test_workflow_lint.py::test_workflow_lint_default_exits_zero) red for every session's Step 9c gate.

The single site

scripts/issue1901_singleturn_retrieval_final.py:133   hf_hub_download(

Added to main by commit 94d6f8838f3 ("Add deduplicated single-turn retrieval evaluation"). Measured on origin/main at 2e3480c2df70: --check-live-hf-retry-routing → FAIL (5 error(s)), of which this is the only one resident on main (the other four are an untracked, main-absent scripts/issue2474_whiten_csls.py belonging to a live session — see § Do not conflate, below).

Fix

The established one-line idiom (the same one #2649 applied at 10 sites):

path = hub.retry_transient(lambda: hf_hub_download(...), what="<desc>")

hub.retry_transient is synchronous (return fn()), so wrapping an EAGER call like hf_hub_download needs no waiver and no materialization trick. (Contrast list_repo_tree, which is a lazy generator — there the consumption must be materialized INSIDE the thunk; not applicable here.)

Why this is filed rather than folded into #2649

#2649 (the main-red repair that just landed as PR #2129, merge commit 942ad66d452) enumerated 34 sites across 12 files and cleared all of them; main is verified green on its whole scope. This site landed on main AFTER #2649's branch base 59ebc5a6b27, mid-round. #2649's approved plan makes "editing outside the 12-file work table" an explicit must-ask boundary, so absorbing a new arrival post-merge would have been a silent scope widening. Filed instead, with the provenance recorded on #2649's epm:merged.

Dispatch guidance (cost)

This is ONE line. Running a full /issue cycle for it — planner, critic ensemble, code review, a ~40-min Step 10d gate — costs far more than the fix. Filed with --no-dispatch deliberately: batch it with the other open main-red / lint-routing tasks rather than spawning a dedicated session. Related open siblings at filing time: #2586, #2648, #2650 (thread-caps / torch-before-dotenv), #2573 + #2631 (Step 5a/10d sync-import attribution), #2550 (the structural selector arm).

The recurring-arrival problem this instance illustrates

Every newly-landed scripts/*.py carrying a bare Hub call re-reds main for the whole fleet, so instance-fixing never converges — #2649 cleared 34 sites and a new one arrived before it could finish merging. tasks/REGISTRY.json holds a double-digit count of fix tasks in this class. The durable arm is a land-time gate (pre-commit / CI on the adding commit) or #2550's directory-membership selector proposal, not another repair task. Recording that here so this task is understood as triage, not as a fix for the class.

Acceptance criteria

uv run python scripts/workflow_lint.py --check-live-hf-retry-routing reports no error naming scripts/issue1901_singleturn_retrieval_final.py on a pristine origin/main worktree.
uv run python scripts/workflow_lint.py (no flags) reports no error naming that file.
uv run pytest tests/test_workflow_lint.py green (or, if red, every failure attributed to a cause other than this file, named).
No blanket grandfathering and no snapshot regen used to silence the check (--regen-hf-routing-snapshot is NOT the remedy here — the call is genuinely unrouted, not a stale-snapshot artifact).

Do not conflate

scripts/issue2474_whiten_csls.py (4 further errors on the same check) is UNTRACKED and absent from origin/main — a live session's in-flight work that only appears in a root-side live-tree scan. It is that session's to land correctly and is NOT in this task's scope. Do not file it, and do not read a root-side FAIL as main being red on its account.

Provenance

Surfaced by #2649's post-merge acceptance verification on origin/main (2026-08-31 ~05:12Z), immediately after PR #2129 merged. Evidence: /tmp/i2649-acc2b.out.