hf_rows_modelscope(cache=False) still returns cached data #229

Bug and scenario

Verified on merged main 8585b3fee5841ebe3389371838f8c9ccb0fd4bc6 (Python 3.12.14). A caller explicitly requesting fresh ModelScope rows with hf_rows_modelscope(cache=False) can receive old cached bytes instead. This is a public loader contract issue, independent of the A/B CLI.

Deterministic offline reproduction

Run from the repository root with the project dependencies installed. The URL reader is fully mocked and the cache is temporary; no provider or dataset endpoint is contacted.

PYTHONPATH=. python - <<'PY'
import json
import tempfile

import agentdescent.dataloader as dl

with tempfile.TemporaryDirectory(prefix="hf-modelscope-repro-") as root:
    dl.CACHE_ROOT = root
    state = {"answer": "old"}
    calls = []

    def fake_read(url, timeout):
        calls.append(url)
        return json.dumps([{
            "question": "Q",
            "answer": state["answer"],
            "context": [["T", ["S"]]],
        }]).encode()

    dl._read_url = fake_read
    first = dl.hf_rows_modelscope(
        url="https://example.invalid/hotpotqa.json", limit=1, cache=True
    )
    state["answer"] = "fresh"
    bypass = dl.hf_rows_modelscope(
        url="https://example.invalid/hotpotqa.json", limit=1, cache=False
    )

    print(
        "first_answer=", first[0]["answer"],
        "cache_false_answer=", bypass[0]["answer"],
        "mock_fetches=", len(calls),
        "expected_fresh_fetches=2",
    )
PY

Observed:

first_answer= old cache_false_answer= old mock_fetches= 1 expected_fresh_fetches=2

Expected: first_answer=old, cache_false_answer=fresh, and two reads of the mocked source. The explicit bypass should not silently reuse a populated cache.

Source

hf_rows_modelscope accepts cache but unconditionally calls fetch_bytes; that helper reads an existing cache entry. The default-cache regression test_modelscope_rows_are_cached passes but does not cover bypass.

Acceptance criteria

• cache=False bypasses previously cached bytes and fetches the current response.
• Repeated uncached calls observe changing mocked responses.
• Preserve default cache=True reuse and context normalization.
• Add offline regressions for both a prepopulated stale cache and repeated bypass calls, with no real network access.
• Keep bypass behavior consistent with the existing loader conventions; document any intended cache-write behavior explicitly.

Dependencies and prior evidence

Introduced with merged #204; no provider account, live dataset, or other feature is required to repair it. Scope: agentdescent/dataloader.py and loader tests. The separate A/B pairing and seed-wiring issues are tracked in #228.

Previously reported in the cache-bypass review comment and the final exact-head follow-up. Reproduced again on actual merged production code without overlaying a fix.

Dedup: current open work and all-state ModelScope/cache-bypass issue and PR searches found no separate tracking issue or repair PR.