Render the seasons nflverse has published, rather than failing on one it has not

[stxkxs]

Why

The app breaks on 1 September 2026, about a week before the opener.

current_season() rolls to 2026 on 1 September, but nflverse publishes a season's assets only once
its games are played. Verified against the live API with the opener two weeks out:

2026 asset	Status
schedules	✅ 272 rows
seasonal_rosters	✅ 2,930 rows
snap_counts / weekly_data / injuries	❌ 404
pbp	❌ NameError

And a mixed request fails wholesale — import_snap_counts([2025, 2026]) raises 404 rather than
returning the published 2025 season. So one unpublished season poisoned a request that a full prior
season would have satisfied, and five of six tools showed only an error toast for ~10 days.

This is pre-existing, not introduced by the recent refactor — the old hardcoded [2024, 2025] had
the same cliff, deferred until someone edited it by hand. The hermetic test suite couldn't catch it
because it never touches the network.

What changed

Tolerance, not prediction. A season list resolves one season at a time and drops what the source
hasn't published, so [2025, 2026] renders 2025 throughout the gap. Only when nothing resolves
does it raise, naming the dataset and season. Returning an empty frame instead would hand the
engines a schema-less frame — the ColumnNotFoundError class that start_sit and trade_value
carry declared output schemas to prevent.

Both failure shapes at one narrowed seam. A missing asset is HTTPError (an OSError).
import_pbp_data instead raises NameError, because its handler reads except Error as e: and
Error is bound nowhere in that module — the handler raises in place of the failure that reached
it. TimeoutError is re-raised ahead of both, since it also derives from OSError and a stalled
transfer is not a missing season.

Cache keys on one season, not a season list. The request is already split per season, so a
partial result becomes structurally unrepresentable rather than defended against. Seasons are shared
across windows, and a season published between loads is picked up because its key is simply absent.

current_season → season_for(day). The old name promised "the season we are in, which
therefore has data" — the reading that aimed six tools at unpublished assets. The new one names only
what it does. The September boundary is confirmed against every schedule since 1999.

A bug found while simulating September: start_sit and trade_value derived their season
filter from the schedule frame, which publishes months early — so both offered a season with no
statistics and defaulted to it, rendering empty against older data they were holding. Both now
derive options from the frame they render, matching the other four tools.

How tested

Live probe, real network, this repo's own loaders, today:

load_snap_counts([2025, 2026])  -> OK rows=26612 seasons=[2025]
load_snap_counts([2026])        -> RuntimeError: nflverse has not published snap counts for season 2026
load_weekly_stats([2025, 2026]) -> OK rows=5641  seasons=[2025]
load_pbp([2025, 2026])          -> OK rows=48771 seasons=[2025]
load_schedules([2025, 2026])    -> OK rows=557   seasons=[2025, 2026]

Plus all six tools driven headlessly with the date resolved to 2026-09-02: none shows an error
toast, every season filter offers only seasons with data, and every table renders.

Gate	Result
ruff format --check / ruff check	0 / 0
mypy src (strict, no carve-out)	0
pytest --cov	0 — 466 tests, 87.14% vs floor 87
uv build	0

Reviewer notes

• An unpublished asset costs one failed round trip per load — nothing negative is cached, which
  is what lets a season appear the moment nflverse publishes it. Measured at 14 failed requests
  across one pass through all six tools on 2026-09-02.
• _team_bye_weeks was checked and cleared — schedules publish complete before any game, and it
  returned 32 teams with one bye each against the real 2026 schedule.
• Start/Sit and Trade Value stay quiet early in a season by design: both need three weeks of
  history and say so in their empty state rather than rendering a projection they cannot support.
• Injury Impact runs one season shallower during the gap, since its window is three seasons and
  only two have data.
• Stale composite-key cache entries (snap_counts_2024_2025.parquet) are never read again and
  survive orphan reaping because their meta records exist. No correctness impact; cache.invalidate()
  clears them if you want the disk back.