fix(catalog): give DeepSeek V4 its own window and price

[raymondginger2018-sudo]

TL;DR

deepseek-v4-flash / deepseek-v4-pro were missing from the model catalog's seed table, so both fell through to the deepseek family rule and inherited the deepseek-v3 row. Two numbers were wrong as a result: the context window (128K / 8K instead of 1M / 384K) and the price (V3's rate applied to both tiers, so a pro/flash split was invisible to anything that accounts by model id). This seeds the V4 ids from the vendor's published table and adds focused regression tests.

Changes

• Changed core/providers/catalog.py (+29)
  • added seed rows for deepseek-flash (the vendor's current name for V4.1-Flash), deepseek-v4-flash (a retired alias that still resolves and is still billed at the Flash price) and deepseek-v4-pro - 1M context / 384K max output;
  • added ("deepseek-flash", ...) and ("deepseek-v4", ...) family rules ahead of the deepseek rule, so an unseen point release inherits the cheapest V4 tier instead of V3's rate.
• Added tests/test_catalog.py (+65) - 4 tests / 8 parametrized cases.

Notes on the numbers

• Source: https://api-docs.deepseek.com/quick_start/pricing (read 2026-09-27).
• V4 is priced by time of day and a catalog row holds a single number, so the row carries the peak rate; off-peak is exactly half of it. The peak number is the upper bound, which is the safe direction for a budget guard.
• deepseek-flash and deepseek-v4-flash deliberately point at the same numbers: both spellings occur in real traffic, and separate rows would fragment cost accounting by spelling rather than by model.

Test Plan

• python -m pytest tests/test_catalog.py -q -> 18 passed
• ruff check and ruff format --check clean on both files (ruff 0.15.21, the version pinned in .pre-commit-config.yaml)

Dependency note

None. Stdlib-only change; the only new import is pytest in the test file.

[Zongwei9888]

Merged into main as e7dea0c. Thank you @raymondginger2018-sudo — I checked every number against https://api-docs.deepseek.com/quick_start/pricing today: 1M context and 384K output for both tiers, peak rates $0.30/$1.20 (Flash) and $1.32/$3.96 (Pro) per 1M tokens, and deepseek-v4-flash as a retired alias billed at the Flash price. Seeding both spellings to one row and putting the V4 family rules ahead of the generic deepseek rule is the right shape.