Remove databento: cotdata is CFTC positioning, in full

[mspinola]

Companion to marketdata#15, which is where databento now lives. Merge that one first.

ADR-0007 is now implemented rather than mostly implemented: there is no price surface here of any kind. 3,424 lines out, 450 in.

Gone

providers/databento.py · store.write_prices / read_prices · --ingest-databento / --build-databento / --reconcile-databento / --batch / --windowed-n1-stats · the cotdata-prices entry point · the [databento] extra · both parity harnesses · four test files. Full migration table in the CHANGELOG.

This package now has no optional data dependency at all — CFTC positioning is a plain HTTP download of public files, and the databento extra was the last one.

The registry lost its vendor columns

Symbol is down to internal, asset_class, is_equity, report_type, cftc_code, hist_codes. PRICE_SOURCES, resolve_source, default_price_source and $COTDATA_PRICE_SOURCE go with the norgate/yahoo/databento mappings.

A comment I left in the previous PR was wrong, and the code now says so. It justified keeping those columns on the grounds that a deployment might share one registry file between the two packages via $COTDATA_REGISTRY. It cannot: this loader hard-requires cftc_code and marketdata's equities have none, so the two files can never be the same file. Nothing outside this repo read the fields — a fleet sweep found three hits, all prose.

The producer-half machinery is gone

There is one producer. cotdata-cot survives as a bare alias of cotdata-update because the scheduled jobs call it by name; cotdata-prices does not, because there is no price half for it to run. _HALF_ACTIONS and _reject_other_half are removed.

What stays, deliberately

prices and metadata remain declared domains with no writer and no reader, and config.prices_dir()/metadata_dir() survive to serve them. A real store carries entries under both, and an undeclared domain is skipped by --migrate-manifests — undeclaring them would strand those entries in the legacy aggregate forever and make --reconcile resolve their directory by guess instead of by declaration.

The seam tests were rewritten, not repointed

Their premise changed. They were about two live halves; they are now about one live half beside retired ones. So prices appears throughout as legacy data written by hand — there is no writer — and the property under test becomes: a COT write leaves that file byte-for-byte alone.

Two were replaced outright rather than adapted: one asserted cotdata-prices refuses a COT action (no such entry point), and one asserted every action flag is assigned to a half (no such halves). Both were replaced with guards for what actually matters now — that the retired flags are refused rather than silently ignored, and that no price-writing function can come back.

Verification

• ruff check src tests scripts clean; 199 pass.
• CLI smoke: the nothing-to-do message names marketdata-update and every retired flag exits non-zero.
• npf's full suite against this branch plus the ported marketdata: 38 failures, identical to the pre-port baseline, none new — all pre-existing sibling-checkout gaps in this sandbox.

---

Generated by Claude Code