refactor: use concrete T5 types and run ty at full strictness

Annotate _load_tokenizer -> T5TokenizerFast and _load_model ->
T5ForConditionalGeneration (the concrete classes AutoTokenizer/
AutoModelForSeq2SeqLM return for this T5-derived model). Unlike the
PreTrainedModel base class, T5ForConditionalGeneration exposes
generate(), resolving the last call-non-callable warning. With zero
diagnostics remaining, drop the ty.toml [rules] warn-downgrades so ty
runs at default strictness and real type errors fail instead of being
masked.

Verified: ty "All checks passed!", ruff clean, 40 fast tests pass.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

app.py

@@ -10,7 +10,7 @@ from functools import lru_cache
 import gradio as gr
 import spaces
 import torch
-from transformers import AutoModelForSeq2SeqLM, AutoTokenizer, PreTrainedModel, PreTrainedTokenizerBase
+from transformers import AutoModelForSeq2SeqLM, AutoTokenizer, T5ForConditionalGeneration, T5TokenizerFast
 
 from langmap.langid_mapping import langid_to_language
 
@@ -25,7 +25,7 @@ def _get_device() -> torch.device:
 
 
 @lru_cache(maxsize=1)
-def _load_tokenizer() -> PreTrainedTokenizerBase:
+def _load_tokenizer() -> T5TokenizerFast:
     tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, use_fast=True)
     if tokenizer is None:
         raise RuntimeError(f"Failed to load tokenizer for {MODEL_NAME}")
@@ -33,7 +33,7 @@ def _load_tokenizer() -> PreTrainedTokenizerBase:
 
 
 @lru_cache(maxsize=1)
-def _load_model() -> PreTrainedModel:
+def _load_model() -> T5ForConditionalGeneration:
     device = _get_device()
     dtype = torch.float16 if device.type == "cuda" else torch.float32
     return AutoModelForSeq2SeqLM.from_pretrained(MODEL_NAME, dtype=dtype).to(device)

ty.toml

@@ -1,9 +1,2 @@
 [environment]
 python-version = "3.12"
-
-[rules]
-# Third-party libraries with missing or incomplete type stubs
-unresolved-import = "warn"
-unresolved-attribute = "warn"
-call-non-callable = "warn"
-invalid-argument-type = "warn"