Read bars from marketdata, not cotdata (ADR-0007 step 4)

[mspinola]

ADR-0007 makes cotdata CFTC positioning only and moves every bar to marketdata. The provider shipped and was verified against cotdata's own output on the Windows box (mspinola/marketdata#7, mspinola/marketdata#8). This is the consumer side.

COT positioning still comes from cotdata. Only bars moved.

Companion: mspinola/cot-analyzer#26 — these two land together, and this is the one that changes reads.

The three call sites

All adjustment="backadj", which is what ADR-0007 measured, so the reads themselves are a rename:

file	before	after
signals.py:1028	cotdata.get_prices(symbol, adjustment="backadj")	marketdata.get_bars(symbol, "backadj")
options_data.py:366	same	same
CotIndexer.py:719	…, start=start_date	…, start=start_date

Tier names and the returned frame are unchanged.

The part that is not a rename, and would have failed silently

CotIndexer busts its parquet caches when the upstream store schema moves, because the per-symbol guards key on column presence and cannot see a value-only change. The example that guard was written for is named in its own comment: reconstructed volume being promoted — a price schema bump.

Prices are no longer in cotdata. So after this repoint, watching cotdata.schema_version() alone leaves uncovered the exact case the guard exists for: stale cached metrics computed against a superseded bar schema, with no error anywhere.

The marker now records both store versions, as separate keys rather than one combined number — they are independent counters, and a max would hide a bump in whichever store happens to sit lower. A marker written before the split has no marketdata version, reads as 0, and busts once. That is correct rather than merely tolerated: the price source moved underneath those caches.

Two tests cover it, including the pre-split marker case.

Verified rather than assumed

marketdata's futures path returns the stored columns without cotdata's keep-list filter, so the frame is wider — it carries FirstVolume, SecondVolume, FirstContract, SecondContract, Volume_Reconstructed, Volume_Source. All three consumers select columns explicitly (price_data[['Open','High','Low','Close']], df['Close']), so the extra columns pass through untouched. Checked before repointing rather than after a failure.

Testing

238 pass, ruff clean, CI green on 3.10–3.13. Run locally against real cotdata and marketdata installs rather than mocks, so the import swap and the new schema call are actually exercised. The signals test doubles were repointed too (signals.marketdata.get_bars), since a stale monkeypatch target would have silently stopped patching and let the tests hit the real store.

CI needed a marketdata sibling checkout: it is not on PyPI, so the declared dependency resolves only from the workspace, and without the checkout pip install -e . failed before a single test ran. A dummy MARKETDATA_STORE joins the dummy COTDATA_STORE for the same reason the first exists — neither package defaults a missing root, both raise by name, and there are two roots now.

Deployment note

The Linux server needs MARKETDATA_STORE set and a second synced store — wired up in mspinola/cot-analyzer#26. Both stores are synced, not produced there.