Fix three fleet-wide test invariants broken by sibling landings on main (ungated HF upload #1739, bind-census skips #2477, dangling import roots #2379/#2474)

Goal

Restore the three fleet-wide test invariants that sibling landings on main currently break, so every session's Step 9c gate stops carrying pre-existing red. Each red is a violation of an EXISTING invariant by newly-landed code — not a gap in the invariant — so the fix belongs in the offending file, not in the test.

Surfaced by task #2336 (the shared atomic-write sweep) while running its batch-0 verification. All three are byte-identical to origin/main and touch none of #2336's files; #2336 is not the cause and is not blocked by them.

Evidence

Confirmed red on origin/main (93cee4593f) by a direct pytest run 2026-08-24. Attributions independently verified: for each, the named commit exists, touches the named file, and the file is present on origin/main.

1. tests/test_no_ungated_upload_call_sites.py::test_no_new_ungated_upload_call_sites

AssertionError: NEW direct HF upload call site(s) without the secret gate:
    scripts/issue1739_r2v2_run.py

Owner: #1739, commit 067bd0300c ("task #1739 claim4 r4 (crash-fix): row-index pushdown kills the 52 GiB split-copy stack (128 GB-cgroup OOM) + leg-keyed sentinel + in-fit RSS crumbs").

The test's own message states the two sanctioned fixes: route the upload through hub._upload / upload_dir_sharded (both already gated), or call secret_scrub.assert_upload_clean(paths, what=...) before the direct call. The test docstring explicitly forbids adding the file to GRANDFATHERED, so that is not an option here.

2. tests/test_argcheck.py::test_bind_fleet_census_positive_coverage

AssertionError: [('scripts/issue2477_base_coherence.py', 935, 'api.list_repo_tree',
  "receiver 'api' bindings not uniformly HfApi()"),
 ('scripts/issue2477_base_coherence.py', 789, 'api.list_repo_tree',
  "receiver 'api' bindings not uniformly HfApi()")]

Owner: #2477, commit 37613acbd9 ("issue-2477: base-coherence decoding-sensitivity follow-up round (#2071)").

Two api.list_repo_tree call sites whose api receiver is not uniformly bound to HfApi(), so the bind census can resolve neither and records them as SKIPPED. The gate asserts census.skipped == [] positively — a skip is a coverage hole, not a pass. Fix is at the call sites: bind api uniformly to HfApi() (or rename the non-HfApi binding so the receiver is unambiguous).

3. tests/test_workflow_lint_prod_import_lockfile.py::test_live_tree_clean

assert len(danglers) == 7   # the 7 class-B dangling issue-stem roots

now sees 11. The four NEW dangling first-party import roots:

root	sites	owner
issue2379_analysis	1	#2379
issue2379_capture	1	#2379
issue2379_mapfit	6	#2379
issue2474_fit	9	#2474 (commit d76e89b387)

The seven pinned/pre-existing roots are unchanged: _issue506_common, issue500_predictors, issue541_personas, issue541_predictors, issue541_upload_lib, issue621_analyze, issue_521_prep_turner_corpus.

Attribution correction worth carrying: this red was first reported as #2474's alone. It is not — #2379 contributes three of the four new roots and #2474 one. Fixing only #2474 leaves the assertion red at 10.

A dangling first-party root means the imported module is absent from main (an unmerged branch, or deleted) — latent breakage, not a third-party verdict. Two dispositions per root: land the missing module, or make the import site tolerant the way #2253 did for its own offenders (waive or try-guard in the same change).

Acceptance

1. uv run pytest tests/test_no_ungated_upload_call_sites.py exits 0 with scripts/issue1739_r2v2_run.py NOT added to GRANDFATHERED.
2. uv run pytest tests/test_argcheck.py::test_bind_fleet_census_positive_coverage exits 0 with census.skipped == [] — the two #2477 sites resolved, not waived.
3. uv run pytest tests/test_workflow_lint_prod_import_lockfile.py::test_live_tree_clean exits 0. If a dangler-count pin change is the right answer for any root rather than a code fix, the new count lands WITH A NAMED DELTA in the same commit — the test's own docstring convention for a legitimate floor break.
4. Each fix is attributed to its owning issue in the commit message.

Notes on routing

Three distinct defects across four owning issues (#1739, #2477, #2379, #2474). Splitting into per-owner children is reasonable if that is cleaner than one sweep — the shared property is only that each reds the whole fleet's gate, which is why they are filed together rather than left as three separate discoveries. Where an owning issue is still live, the arbitration rule applies before editing its files: probe for a live writer, post a file-set claim:, and sequence after any unreleased claim rather than racing it.