docs: which machine publishes, and why only one can write a store this package reads

[mspinola]

Docs only. Two corrections, one of them to a mistake this session made and shipped.

1. The production publish runs on the producer, not on this Mac

cot-analyzer/server-side/README.md already said the panel is built on the Windows/Norgate producer. I then installed a launchd agent on the Mac and described it as the schedule. Both are true of different things and nothing said which.

The Linux server's sync originates on the Windows box, so a panel built downstream of that sync is one the sync never sees. bin/publish_damage.sh and the plist beside it are macOS/Linux, and what they feed is a local cot-analyzer — genuinely useful, and not a deploy.

The agent stays installed. Recording that it was installed against a document already saying otherwise is worth more than quietly deleting it.

Windows has no bash, so the production path is the portable driver:

set COTDATA_STORE=...\cotdata_store
set CROWDMON_STORE=...\crowdmon_store
.venv\Scripts\python bin\publish_damage.py

And the ordering constraint nothing enforces, now stated in all three places rather than left to be discovered: publish before the Windows → Linux sync, or the sync ships last week's panel beside this week's prices.

The first wording was too strong, and is corrected here

I first wrote "the panel that reaches production has to be built there", which reads as a capability limit on this package. It is not one. crowdmon runs anywhere the store is readable, macOS included, on the full price-dependent chain — a local run builds all 47 markets with beta attached and both trigger sides, and §D13's figures were measured that way.

The real rule is narrower: the panel must be built upstream of whatever ships it. Today that is the Windows box because the sync originates there. If the sync origin moves, the publish moves with it and no code changes.

Also worth separating, because it decides the fallback: the panel is the one payload with no producer-origin argument of its own. Prices are vendor data and can only come from the subscription machine; the panel is derived output that any machine with the store regenerates identically.

2. §D14 — only the Norgate producer can write a store this package can read

A tier fact, not an OS fact, and the three machines fail for three different reasons, which is why lumping them under "Windows-only" was misleading.

box	Python ≥3.10	Norgate	can produce a store this package reads
Windows	yes	yes	yes
macOS	yes, runs 3.11 under uv and builds the full panel today	no	no
Linux server	no, runs 3.9	no	no

Three facts that were each already written down, composed here because nothing had composed them:

1. notional refuses anything but unadj, riskunits anything but propadj, and propadj is derived on read from unadj + backadj — so both stored tiers are a precondition. ContractMaster.coverage() already encodes it.
2. Norgate is Windows-only by mechanism, not licence: norgatedata talks to a locally installed updater application rather than an API.
3. crucible-stack ADR-0007 scopes databento to exactly one series per symbol, backadj, and says its coverage should not broaden.

Composed: a databento-backed futures store cannot feed this package at all. The refusal is a raise and not a warning because the error survives a spot check — backadj percent vol is 201x too high for soybeans and 0.47x for gold, which never goes negative.

So the sentence to carry is not "crowdmon needs Windows". It is "crowdmon needs a full-tier store, and today exactly one machine can write one."

The trap this creates for ADR-0007 step 2

marketdata has no futures provider, so step 2 writes one. If it ships databento-first, or ships Norgate storing only backadj to match the server's needs, this package stops working — as a raise inside a weekly scheduled job.

Work order filed as cotdata/docs/handoffs/2026-08-04-adr7-step2-price-producer-split.md (mspinola/cotdata#101), which carries that requirement in §4 where its author will be reading.

Testing

Docs, a shell comment and a plist comment. 560 passed / 95 skipped from a worktree, unchanged; ruff clean; plutil -lint clean on the plist. The four pinned suite counts do not move.

🤖 Generated with Claude Code