Name the second store, in the launcher and in the docs

[mspinola]

Follow-on to #26 / #27 / #28. Those made the repo depend on marketdata; this makes a
fresh clone and a fresh run say so, and records where the bars actually are.

Prompted by a real failure: ./run-local.sh died with
ModuleNotFoundError: No module named 'marketdata' on a venv that predated the
dependency. The venv was the immediate cause, but the instructions that venv was built
from are still a step behind the code, so the next clone hits the same wall.

The launcher

MARKETDATA_STORE now gets the same upfront check COTDATA_STORE has, plus a guard
against pointing both at one directory.

Without it the failure is quiet in the worst way: the import succeeds, the app binds
:5001, every COT page renders off the other store, and only the first chart wanting bars
raises. That reads as a UI regression rather than a missing environment variable.

The collision branch is there because the two stores may share a synced parent but never
one directory: each producer rewrites its own manifest.json read-modify-write, so a
shared manifest loses entries.

The docs

file	was
README.md	two siblings to clone, two editable installs, prices read via cotdata.get_prices
docs/ARCHITECTURE.md	cotmetrics depends on cotdata alone; caches busted by two counters
docs/data-path.md	cotdata.get_prices drawn as the live read path

There are three siblings, three counters, and cotmetrics/src no longer calls
get_prices anywhere. Missing ../marketdata fails at import rather than install
(cotmetrics/__init__.py re-exports signals, which imports marketdata at module
level), so README now carries that traceback verbatim.

On the counters: marketdata is the one that carries the weight. The case the guard was
written for, reconstructed volume promoted to default, is a price schema bump, so
watching cotdata alone would leave uncovered the only case it exists for.

The finding worth reading

The two ADR-0007 steps are out of step with each other, and data-path.md now records
that rather than smoothing over it.

• step 4, repoint consumers: done
• step 2, move the producer code and the bars: on ice

So the read points at $MARKETDATA_STORE/bars/futures/ while 99 price parquets sit in
$COTDATA_STORE/prices/ where the producer keeps writing them. Measured against both
live stores:

~/code/cotdata_store/prices/           99 parquet files
~/code/marketdata_store/bars/          equities/yfinance only, 13 tickers, no futures/
marketdata.get_bars("GC", "backadj")   ->  0 rows, no exception

A local checkout therefore renders positioning without prices, silently. The old diagram
is kept and marked superseded, since its vendor-abstraction argument still holds; the
measured state goes in a new section beside it with a check a reader can run.

Verification

• All three launcher branches exercised by hand: unset exits 1 naming the variable and
  the .env path, equal paths exit 1 naming the manifest reason, distinct paths fall
  through to the app.
• bash -n run-local.sh clean.
• ruff check src tests clean, 131 tests pass.
• Store contents and get_bars behaviour measured, not assumed.

Not in this PR

• .env is gitignored, so the MARKETDATA_STORE line was added to the local file only.
  Anyone else needs to add it themselves, which the check now tells them to do.
• docs/data_architecture.md is stale in two unrelated ways: it still calls Databento the
  primary price provider with a Yahoo fallback (:143) and still names
  pardo_quant_framework rather than npf (:153).
• The workspace CLAUDE.md still lists /damage as a live crowdmon consumer. That page
  left in Remove the /damage page now that crowdmon is deprecated #22 and The last crowdmon references leave the repo with the page #25.

🤖 Generated with Claude Code