[Uneadlight]

Fix #773 — expose prompt_eval_cached_count on generate/chat responses

Problem

GenerateResponse / ChatResponse silently drop the server's prompt_eval_cached_count
field, so callers cannot tell how many prompt tokens were served from the prompt cache.

Reported in #773 (and duplicated in #779).

Root cause

BaseGenerateResponse (in ollama/_types.py) declares every numeric field explicitly and
SubscriptableBaseModel does not set model_config = ConfigDict(extra='allow'), so pydantic v2
uses its default extra='ignore' and discards the unknown key on validation.

Reproduction (local, ollama 0.40.13, qwen2.5:3b)

Two identical generate calls through the SDK after this change:

--- call 1 (cold) ---
prompt_eval_count       = 36
prompt_eval_cached_count= 0
--- call 2 (same prompt) ---
prompt_eval_count       = 36
prompt_eval_cached_count= 35

Before the change both calls raised AttributeError / returned None for that attribute,
even though the raw HTTP response already carried "prompt_eval_cached_count": 0.

Change

ollama/_types.py — one field on BaseGenerateResponse (inherited by both response types):

prompt_eval_cached_count: Optional[int] = None
'Number of prompt tokens that were served from the prompt cache instead of being evaluated.'

Optional keeps compatibility with older servers that do not send the field.

Tests

Added to tests/test_type_serialization.py:

• test_generate_response_keeps_prompt_eval_cached_count — attribute, [] access, in, and model_dump()
• test_chat_response_keeps_prompt_eval_cached_count
• test_prompt_eval_cached_count_is_optional — a payload without the field still parses, and
  exclude_none does not emit it

Results: tests/test_type_serialization.py → 21 passed (2 pre-existing failures from this
machine's environment, see below); full suite → 112 passed, 8 failed (same 8 as before the change).

Why not extra='allow'

extra='allow' would accept any unknown key and would surface it through model_dump(),
which changes the public shape of every response object and could leak unrelated server fields.
Declaring the one documented field keeps the model explicit and type-checked.

Environment note (not a repository bug)

The 8 failures in this run are specific to this machine and unrelated to the change:

• 6 × ValueError: Invalid image data, expected base64 string or path to image file — the local
  temp directory cannot create files from a Python process (tempfile.mkstemp hangs), so the
  test fixtures produce empty image files.
• 2 × PermissionError in test_client_create_blob* — same temp-directory cause.

They fail identically on a clean checkout with no local modifications.