feat(executor): every order is an OrderSpec, and the second Coinbase renderer is gone (#524)

[eaitbrahim]

The flip. The executor placed orders by handing hand-built Coinbase dicts to a pre-port
client; it now builds OrderSpec values and reads Preview/PlaceResult.

The bytes on the wire are unchanged, and that is verified rather than asserted. Ten order
configurations rendered by the new path were compared against the pre-flip renderers loaded out
of git — the bracket across four size/price/increment combinations, plus every market leg — and
are byte-identical.

One renderer, not two

CoinbaseClient.preview_order/place_order take a spec and render it through
keel_broker_coinbase.translate.to_order_configuration — the adapter's own function.
executor._bracket_order_configuration is deleted, and with it the test #502 stage 1 shipped to
pin the two byte-identical while both existed. That test's own words: "The test imports both;
production code does not." There is one now, so there is nothing left to hold in agreement.

_order_configuration becomes _order_spec: BUY is MarketIOCByQuote, SELL is
MarketIOCByBase, the bracket is BracketGTC. #516's quantization is untouched and still runs
before the spec is built, including its deliberate BUY/SELL asymmetry.

The trap I nearly walked into

Every OrderSpec carries an initial_status ClassVar, and using it for the order row's status is
the obvious move and wrong. The port's vocabulary is the venue's (filled_or_rejected,
open); this column is keel's (filled, pending).

reconcile sweeps for pending. Writing open would leave every resting order invisible to the
sweep that exists to observe its fill — a bracket recorded as open is a protective order keel
would never look at again.

So _initial_status stays, mapping spec.kind to keel's words. What went is the dict inspection
(next(iter(order_configuration), "")), not the vocabulary. Caught by a test asserting 'filled',
which is exactly what that test was for.

Smaller things the types made obvious

• raw_response stored the entire placement response as JSON so _native_order_id could dig the
  id back out to cancel with. It stores {"order_id": ...} now, from PlaceResult.broker_order_id.
  No migration needed — both shapes answer the same data.get("order_id"), so pre-existing rows
  read unchanged.
• _preview_book already accepted Preview | dict and the Coinbase adapter already carried the
  book in detail, so A routing-time maximum-spread gate — refuse entries when the live book is too wide #350's spread gate and fix(executor): warn loudly when a conditional entry price is overridden by market routing (#260) #332's override warning came through untouched. The
  executor was written anticipating this flip.

Tests

~13 fakes across five files moved to the port's signatures. The dict-shaped preview payloads are
deliberately kept as dicts at the call sites and converted by one helper: dozens of tests build
a bespoke preview to exercise one degraded field (a missing best_bid, a non-numeric best_ask),
and rewriting each into a constructor would have been a larger diff than the change it accompanies,
with more chances to alter a case by accident.

Verification

4301 passed / 3 skipped, ruff clean, mypy clean across 347 files, one paper cycle run against real
venue data, and the byte-identity check above.

What's left of #524

_build_broker still constructs CoinbaseClient rather than resolving through load_broker. That
is now a small change — the client and every adapter speak the same interface — but it is gated
on two consumers that are not port methods: assets discover's list_products, and assets holdings' get_accounts. Both want a decision (a port catalogue read? holdings onto
get_balances?) rather than a mechanical move, so they belong in their own change.

🤖 Generated with Claude Code

https://claude.ai/code/session_01T6yA5khYnJ2qzheArRToQ2