Cover the typed batch payloads with offline tests and typechecking

[snoble]

Written by Claude, running in Steven's session — these are Claude's words and judgements, not Steven's.

Stacked on #45 (base typed-batch-ledger-entries). Tests and typechecking only — no behaviour changes. #45's five review items are all resolved there; this fills the coverage holes flagged alongside them.

What was uncovered

Nothing imported the generated per-entry classes. The existing tests assert on render_module's output as a string, and the sample fixture builds a model by hand from BASE_CLASS_SOURCE. A rendered module that declared the same field twice, or referenced a name the header does not import, satisfied all of them.

The optional-parameter branch had no coverage at all. Every CLI-generated parameter is non-null, so no snapshot query reaches the renderer's = None path.

The warning paths had none either. Each leaves a working but quietly worse SDK — a parameter that lost its type, a batch method that does not typecheck, models nothing accepts. A silent version of any of them passed the whole suite.

CI ran mypy -p fragment only. tests/ was never typechecked, so the mypy_path setting added in #45 was dead config, and the PR's central claim — that a caller gets real type errors — was verified by nothing.

On forks, no test ran at all. The only job that ran pytest needs secrets, and conftest.credentials fails the run without them.

What this adds

File	Covers
tests/test_typed_entries_generated.py	Renders into a throwaway package, imports it, uses the classes. Optional parameters, field collisions, escaped names, to_entry_inputs, the empty-module shape that fragment/sdk/typed_entries.py actually is, and every model in the committed snapshot.
tests/test_typed_entries_warnings.py	All five degradation paths, plus the silent counterpart of each so the assertions mean something.
tests/type_checks/batch_entries.py	mypy-only. Positive cases must pass; negative cases are # type: ignore[...], and warn_unused_ignores fails the day one starts passing.

make typecheck now covers tests/ too, make unit runs everything offline, and a new credential-free CI job runs both.

76 offline tests, 0.2s, no network.

Verification

Every new test was checked by breaking the thing it guards:

Mutation	Result
Revert the field-collision fix	5 fail
Drop = None for optional parameters	5 fail
Emit the escaped Python name as the wire key	3 fail
Drop typeVersion from to_input	1 fails
Drop the list(entries) coercion	2 fail
Loosen entries to Any	4 unused-ignore errors
Narrow Sequence back to list	3 type errors

The last two are the point of type_checks/: it fails both when the signature loosens and when it tightens too far.

One-line change to #45's files

pytestmark = pytest.mark.integration in the two integration modules. Their snapshot_client fixture is untouched, including the comment about **credentials — that stays accurate, since the fixture still annotates Dict[str, str].

🤖 Generated with Claude Code

[snoble]

Claude, in Steven's session.

Correction to the original description: I claimed make lint was unrunnable and rewrote ~1000 lines. That was wrong, and I've removed it.

The churn came from my own environment using isort 8.0.1 where poetry.lock pins 5.13.2. With the pinned versions, make lint is a no-op on clean dev and on this branch. isort does reformat imports and black does put them back, so the intermediate state churns, but the pair converges and the committed tree is already at the fixed point.