Fleet-wide Step 9c red: two ungated HF upload call sites in scripts/issue1739_r2v2_run.py

Goal

Restore tests/test_no_ungated_upload_call_sites.py to green on main by routing the two direct HfApi().upload_file calls in scripts/issue1739_r2v2_run.py through the gated path (or asserting the secret gate before them), so that any round whose Step 9c selection includes that invariant test stops bouncing on a failure it did not cause.

The problem — standing red on main, not a diff regression

scripts/issue1739_r2v2_run.py calls HfApi().upload_file(...) directly at line 439 (_stage_crumb_path upload) and line 543 (_gate1_crumb_path upload), both wrapped in hub.retry_transient(...) but neither routed through the gated hub._upload / upload_dir_sharded path and neither preceded by secret_scrub.assert_upload_clean(...).

Verified by the #2253 orchestrator (2026-08-21):

• git show origin/main:scripts/issue1739_r2v2_run.py — both call sites present, byte-identical, on pristine origin/main.
• git log 2087dda0c1..origin/main -- scripts/issue1739_r2v2_run.py — zero commits since #2253's worktree base, so this is not a recent regression and is not attributable to any in-flight round.
• Live run: uv run pytest tests/test_no_ungated_upload_call_sites.py -x -q → 1 failed, AssertionError: NEW direct HF upload call site(s) without the secret gate: scripts/issue1739_r2v2_run.py.

Why this is worth its own task rather than a note

The failing test is a WORKFLOW INVARIANT test, so it is stem-mapped into the Step 9c selection of any round touching the upload surface. Every such round bounces its test-verdict gate on a red it cannot fix within its own scope — the #1388 shape, except the red is already landed rather than newly introduced. The cost is paid repeatedly by unrelated tasks until someone gates these two calls.

The test's own docstring is explicit that the GRANDFATHERED allowlist is not the remedy ("Never add to GRANDFATHERED — see this file's docstring"), so the fix is to gate the calls.

Proposed change

Route both call sites through the gated upload path — hub._upload / upload_dir_sharded — or, if the direct call is deliberate for these small JSON crumb files, call secret_scrub.assert_upload_clean(paths, what=...) immediately before each. Follow whichever pattern the file's other upload sites (if any) already use, and whatever hub exposes for a single small file, rather than inventing a third shape.

Both payloads are small JSON stage/gate crumb files written to hub.DEFAULT_DATASET_REPO as repo_type="dataset", so neither is an LFS-path or a bulk-upload case.

Acceptance

• uv run pytest tests/test_no_ungated_upload_call_sites.py PASSes with no GRANDFATHERED addition.
• The two crumb uploads still function (the gate is added, the upload is not removed) — confirm via the file's own smoke path or a dry-run, not by deleting the calls.
• scripts/workflow_lint.py no-flags run stays green; ruff clean on the touched file.

Related

• scripts/issue1739_r2v2_run.py:439,543 (the two call sites)
• tests/test_no_ungated_upload_call_sites.py:176 (the failing assertion)
• #1739 (the owning experiment task — the script is its driver)
• #2253 (surfacing round; its implementer flagged this while running its own gate scope, with the origin/main byte-identity proof)

Provenance

Surfaced by the implementer at #2253 implementation round 1 as one of two orchestrator-routing items, and independently re-verified by the #2253 orchestrator (origin/main blob read, empty commit log since base, live pytest failure) before filing. Filed rather than left as a chat note because the red is standing on main and taxes every future round that selects the test — #2253's own 134-test selection does NOT include it, so this is filed for the fleet, not to unblock #2253.