fix(models): flatten MiddleSide to the wire; alias confidence; add quality_tier

[Mlaz-code]

Fixes #18.

The two defects, reproduced against live payloads

MiddleOpportunity.model_validate(<live row>)
  -> 2 validation errors
     side1.odds  Field required
     side2.odds  Field required

EVOpportunity.model_validate(<live row>)
  -> parsed, but: wire confidence=80  ->  model.confidence_score=None
                  wire quality_tier='A'  ->  model has no such attribute

client.middles() could not parse any successful response. confidence_score read None on every row.

One correction to the issue

#18 says the wire sends sportsbook rather than book. It doesn't — live sides send book, which the model already accepted. The only field that broke parsing was the nested odds. Changing book on that report would have introduced a bug rather than fixed one.

Live side shape, all 12 fields present on 4/4 sampled sides:

book  selection  line  odds_american  odds_decimal  odds_probability
fair_probability  stake_percent  odds_age_seconds
external_event_id  market_id  selection_id

Choices worth flagging

confidence_score is aliased, not renamed. AliasChoices("confidence", "confidence_score") matches the file's existing idiom (fair_probability/true_probability, sharp_book/devig_book) and keeps the public attribute name, so nobody's code breaks.

MiddleSide is a hard shape change — side.odds.american becomes side.odds_american. Normally breaking; here it cannot be, because the model never parsed a response, so no working consumer exists to break.

Bonus defect found while checking the release path

pyproject.toml = 0.4.1, __init__.py = 0.4.0. #13 bumped one and missed the other, and 0.4.1 is already on PyPI — so the published package reports sharpapi.__version__ == "0.4.0". PyPI is immutable, so that stays wrong for 0.4.1 forever; the test here only stops the next one. Synced to 0.4.1 (matching what is published), which is a correction, not a release bump.

Tests — 9 new, all against captured live payloads

tests/test_wire_contract.py + tests/fixtures/{middles,ev}_live.json. Hand-written dicts could never have caught this: they'd be written from the same wrong belief as the model.

Two notes on what is deliberately not asserted, both discovered by writing the check and watching it misfire:

• "every declared field is populated by some payload" — dropped. A 2-row fixture cannot tell "the SDK declares a field the wire never sends" from "this sample lacked a conditional field." It flagged sharp_odds_american, is_suspended and the *_ref objects, all legitimately conditional.
• importlib.metadata.version() for the version check — replaced with a direct read of pyproject.toml. Dist metadata is only as fresh as the last pip install; the editable install on the dev box reports 0.2.0, so that test would have passed in CI and failed locally for a reason unrelated to the bug.

Run locally with PYTHONPATH=src — the editable install resolves to the canonical checkout, not a worktree.

Release

Not bumping the release version — this repo lands those as separate chore(release): commits, and PyPI needs a gh release create to publish. This fix should land before the docs audit's Wave 1, since the docs mirror these model shapes and the SDK is the source they follow.

Type: fix