Add team pitching: new table, ingestion path, and pitches-per-game page

[Mattsface]

Summary

The first feature here that is not free. Every previous page read the hitting gameLog split already being fetched, or derived new figures from rows already stored. Pitching is a separate MLB stat group in a separate request, landing in a new table.

• New team_game_pitching_lines table + migration 27a202039134
• /pitching page: pitches per game as the chart, with ERA, WHIP, K/9 and BB/9 in the summary cards
• Ingestion fetches both stat groups and persists them in one transaction

Closes #41.

Innings are stored as outs

The single most important decision in this PR. MLB returns inningsPitched as a string in baseball notation:

inningsPitched = '10.2'   means ten and two-thirds innings, NOT 10.2 of them

Parsing that as a decimal silently corrupts every derived rate, and the result stays plausible enough to go unnoticed. The same split carries outs as an exact integer, so that is the column:

outs 32  ->  ER 9 * 27 / 32 = 7.59        MLB's own era for that game: '7.59'
reading '10.2' as a decimal:  9 * 9 / 10.2 = 7.90     wrong

innings_pitched_display reconstructs the 10.2 form for display only, and a test asserts it is never used in a calculation.

Only raw components are persisted — ERA, WHIP, K/9 and BB/9 are derived on read, so a stored rate cannot drift from its components. Balls are not stored either: MLB leaves that field empty on the team game log, and balls are number_of_pitches - strikes.

Counts and rates aggregate differently

Every previous page charts a count, where the mean of the per-game values is the right season figure. Pitches per game follows that rule.

ERA, WHIP, K/9 and BB/9 are rates, and a rate over several games is the ratio of the summed totals:

season ERA (correct)   629 ER * 27 / 4388 outs   =  3.870
mean of 162 game ERAs                            =  3.965

That 0.094 gap would match no published source. The same rule governs the rolling window (which accumulates earned runs and outs rather than smoothing game ERAs) and the league context (outs-weighted, not game-weighted).

The tests that cover this build seasons with unequal innings on purpose — with a regulation nine innings in every game the correct aggregation and a naive mean agree exactly, so a test built that way would pass against a wrong implementation. One test documents that case explicitly.

Ingestion costs one extra request, not three

get_team_game_lines shares the team lookup and the season schedule between the two game logs:

get_team 1 + get_schedule 1 + get_team_stats 2  =  4 requests, not 6

Over a 30-club league import that is 60 requests saved. Both persist in the same transaction, so a team-season can never hold batting rows without pitching ones. include_pitching=False drops back to the original three requests.

The two logs are validated against opposite sides of the score — a hitting split's runs must equal the selected team's scheduled score, a pitching split's runs are runs allowed and must equal the opponent's. That makes the schedule an independent check on which stat group a split belongs to.

Notes for review

• A separate table, not more columns. A season imported before pitching existed has no pitching rows, so nothing needs the nullable-until-backfilled treatment the strikeout and baserunner columns required. Every pitching column is NOT NULL.
• The migration was amended rather than stacked. number_of_pitches and strikes were folded into the create migration, which had not been merged or released. Any machine that already ran the earlier version needs alembic downgrade -1 && alembic upgrade head.
• Shared test fakes now serve both stat groups and filter by the requested group, so a service reading the wrong group fails loudly. Existing batting-path tests pass include_pitching=False.
• The batting/pitching upserts are one generic function, addressing duplication flagged in the review of Add team baserunners/game trend #38.
• No MLB line on the chart. A league-wide pitches-per-game figure needs every club's pitching imported. The league comparison machinery is built, tested, and wired into the route; it simply has no line yet. Its sign convention is inverted from every other page — negative ERA difference is better — which is worth a look before it surfaces.

Test plan

• poetry run pytest — 1370 tests passing, 94% coverage (59 new; team_pitching 96%, league_pitching 100%)
• poetry run ruff check . / ruff format --check . clean
• Migration upgrade/downgrade/upgrade round-trip verified; 18 check constraints present; the three definitional constraints verified to reject bad rows
• Relational constraints validated against 648 real 2025 team-games across four clubs before being encoded
• Verified against the live API and a real local database: 162 rows imported for the 2025 Mariners — 23,623 pitches over 1,462.2 IP, 145.8/game, 65% strikes, ERA 3.87, WHIP 1.22, K/9 8.77, BB/9 2.79, all matching figures computed directly from the raw payload
• Re-importing batting reported Unchanged: 162, confirming ingestion stays idempotent
• /pitching returns 200 for a season with pitching and 409 with re-import guidance for one without

🤖 Generated with Claude Code

https://claude.ai/code/session_0178J2Rs8QQfS1wTqG98FSK9