feat(executor): the increment read goes through the port, and stops fetching 900 products for one

[eaitbrahim]

#561 gave the port get_instrument and wired nothing to it. This wires it, which was the point.

Third additive step on #524, and like the two before it no live behaviour changes — the same
increments reach the same quantization through the same cache.

The read the port was built for

_base_increment_for called broker.list_products() — about 900 rows on Coinbase — to use one
field of one of them, then picked through raw dicts for product_id and base_increment. That
was the pre-port CoinbaseClient's only catalogue read.

CoinbaseClient.get_instrument() now answers the port's type too — the same bridge get_balances
is — so the executor asks one question whether it holds this client or a real adapter.
list_products stays exactly where it belongs: keel assets discover, which genuinely wants the
catalogue.

A caching argument that no longer applies

The old docstring defended writing one cache row per miss "even though the response carries every
product", because Repository.set_state commits per call and caching all ~900 would mean ~900
fsyncs inside the order-placement path. The argument was sound and the shape was not. Asking
for one product removes the temptation rather than resisting it — there is no longer a catalogue to
decline to cache.

The test that pinned it stays, with its reasoning updated rather than deleted: one venue answer must
still write one row.

And the adapter that refuses

keel-broker-alpaca raises NotImplementedError from get_instrument, deliberately (#561) —
None would tell a caller a symbol is unlisted when the truth is nobody wrote the read.

On this path that must land as unknown, not as a crash. _base_increment_for never raises by
contract: unknown means send the quantity unquantized, and an exit that refused to place because a
catalogue lookup was unimplemented would be a protective order withheld over a missing convenience.
Now pinned with a test.

Verification

4304 passed / 3 skipped, ruff clean, mypy clean across 347 files.

What's left of #524

Only order placement. Every read the executor makes — balances, increments — now goes through
port shapes and needs no further change at the flip. What remains is preview_order/place_order
taking OrderSpec instead of raw dicts, and _build_broker resolving through load_broker.

That one is not additive. It rewrites every live order path, and it's where
_bracket_order_configuration, the ConfirmFn dual arm and the capabilities() grandfather clause
finally get deleted. Worth doing as its own reviewed change rather than tacked onto this.

🤖 Generated with Claude Code

https://claude.ai/code/session_01T6yA5khYnJ2qzheArRToQ2