Select strict NLP GPU numerical phases

// Development runtime and an available native Metal/MPS Float32 backend required.
// Select CUDA/Float64 or CUDA/Float32 explicitly for a compatible CUDA build.
// No CPU fallback occurs if this strict backend is unavailable.
// Run in a fresh writable directory: checkpoint creation is exclusive.
import aner.nlp
import aner.nn
import aner.token

let execution = nn.execution_context(device: "mps", dtype: "Float32")
let source = nlp.corpus(List("Alice visits Paris .", "Bob visits Rome ."),
    List("gpu-train-1", "gpu-train-2"), role: "training")
let training = nlp.annotate(nlp.tokenize(source),
    List(List("PROPN", "VERB", "PROPN", "PUNCT"),
         List("PROPN", "VERB", "PROPN", "PUNCT")), task: "pos")
let model = nlp.fit_tagger(training, epochs: 3, seed: 731, execution: execution)
let held_out = nlp.corpus(List("Bob visits Paris ."), List("gpu-validation-1"), role: "validation")
let tokens = nlp.tokenize(held_out)
let prediction = model.tag(tokens, execution: execution)
print("Source-aligned labels; emission arithmetic on the strict GPU profile")
print(tokens.token_texts(0))
print(prediction.tags(0))
print(model.execution_summary())
print(prediction.execution_summary())

let first = nlp.fit_tagger(training, epochs: 1, seed: 731, execution: execution)
first.save("execution-gpu-pos-stage.anertagger")
let restored = nlp.load_tagger("execution-gpu-pos-stage.anertagger")
let continued = restored.resume_tagger(training, epochs: 2, execution: execution)
print("Same-profile GPU checkpoint continuation")
print(continued.fingerprint() == model.fingerprint())
print(continued.tag(tokens, execution: execution).fingerprint() == prediction.fingerprint())

let encoder = nlp.vectorizer(source)
let features = encoder.transform(held_out, execution: execution)
let dense = features.dense(max_cells: 1000, execution: execution)
let batch = nlp.batch(held_out, token.byte_tokenizer(), max_length: 32)
let mask = batch.mask(execution: execution)
print("Local Tensor placement matches the explicit profile")
print(nn.device(dense))
print(nn.dtype(dense))
print(nn.device(mask))
print(nn.dtype(mask))
print(features.execution_summary())
// Source token IDs never pass through floating point. Host Viterbi and ranking
// remain CPU work. Seeds do not promise Float32/Float64 or CPU/GPU equivalence.