ADR-0007 step 2: record that the marketdata futures provider shipped

[mspinola]

Description

Starts ADR-0007 step 2 — the market-data extraction out of cotdata. No cotdata code changes here. The code landed in mspinola/marketdata#7. This PR is the cotdata-side record: a new handoff, plus the close-out of the 2026-08-04 work order that claimed the step.

Deletion of cotdata's price code is §7.5 of that work order and comes after the consumers are repointed. Until then cotdata is still every consumer's only working path, so it is untouched on purpose.

What shipped in marketdata

providers/norgate.py (both tiers, volume reconstruction, data-driven finals gate, NDU-down guard, contract specs), a store tier axis, adjust.ratio_adjust for propadj, the futures path in get_bars, contract-spec read/write/upsert, 49 futures registry symbols, and --domain / --metadata / --full on the producer CLI. 89 tests, ruff clean.

The store could not hold the result

The work order's §4 established that crowdmon needs both stored tiers because propadj is derived from the pair — read as a statement about what the producer must fetch. It is also a statement about what the store must hold, and marketdata's store could not: bars/<domain>/<source>/<symbol>.parquet gave both tiers the same path, so the second write would have silently replaced the first.

The one-frame-per-symbol layout rests on a property equities never violate — corporate actions are dated events the vendor hands over with the bars. Norgate's back-adjustment is roll splicing the vendor performed, and the stitched spread appears in no other series, so backadj cannot be derived from unadj or the reverse. That is a schema change (v1 → v2), not a provider detail.

Measured before writing anything: on v1, get_bars("ES", "backadj") raised tier must be one of ('split', 'raw', 'total'). check_tier accepted the futures tier and adjust() then rejected it, so the declared futures domain was a promise about error messages rather than a working path.

crowdmon was deprecated three days after the work order was written

Checked because CLAUDE.md mentions it in passing. It is not a passing matter — the work order's §4 is the section it says it exists to record, and it reasons entirely from crowdmon.

crowdmon/DEPRECATED.md decides it 2026-08-07, one day after that work order's own crowdmon companion. Its §3 was resolved 2026-08-08: both launchd jobs unloaded, cot-analyzer's /damage page removed (that repo's PR #22), the open work order closed unstarted, and the package left with no consumers at all.

The both-tiers rule is not relaxed — the provider enforces it three ways. Its argument moves to npf/books/treasury_seasonal.py, a live book that sets RETURN_TIER = "propadj" after a void first run in which 15 of ZB's 100 verdict-window trades had their sign inverted, and which reads backadj and unadj together for roll detection regardless.

§7.3 is void. Repointing a frozen package with no consumers is waste, and DEPRECATED.md §2 wants its live pins neutralised rather than given a post-freeze dependency. That costs something worth naming: crowdmon was to go first because it fails loudest, so dropping it drops the check that would have proved the provider before the easy consumers leaned on it. Nothing exercises propadj until the deferred npf pass, which makes the Windows-box comparison the main evidence the port preserved the numbers.

Convention note

The work order body is preserved verbatim and the corrections are appended as its §8, per the handoff register's rule — amending a body after execution destroys the record of what was asked. §4 and §7.3 carry pointers into §8, which is what that convention asks for directly.

Explicitly not done

§7.4 repoint cotmetrics/cot-analyzer (the only repoint left standing), §7.5 delete from cotdata. MME/MFS are not ported — Norgate carries no continuous series for either. Confirmed this costs crowdmon nothing: both already fail coverage() for the same reason, pinned by its live test as the only two non-joinable, absent from both vintage panels, and heldout in the deployed params. Counts agree — 49 of 51 joinable, 49 ported.

The provider has not run against a live Norgate. It could not, on Linux, for the same reason the work order's §5.3 concludes a synced store is permanent.

One open question I did not resolve: ADR-0007 defers npf on live-book grounds, decided when npf was one of two propadj consumers. It is now the only one. That is ADR-0007's call, and crucible-stack is outside this session's scope, so §8.2 records it rather than answering it.

Fixes # (issue) — n/a, this executes crucible-stack ADR-0007 via docs/handoffs/2026-08-04-adr7-step2-price-producer-split.md.

Type of change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• This change requires a documentation update — it is the documentation

Checklist:

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have run tests (e.g. pytest) and they pass locally — 89 pass in marketdata; no cotdata code changed
• Any dependent changes have been merged and published in downstream modules — Norgate futures provider (ADR-0007 step 2) marketdata#7 is open, and no consumer is repointed yet