Inspect NLP CPU execution and checkpoint replay

// Development runtime only. All input is synthetic; no network is used.
// Run in a fresh writable directory: checkpoint creation is exclusive.
import aner.nlp
import aner.parallel
import aner.nn

// One worker is valid even under a one-slot CPU allocation. Choose a larger
// admitted value to compare concurrency; worker count does not reserve cores.
let execution = parallel.local(workers: 1)
let source = nlp.corpus(List("Alice visits Paris .", "Bob visits Rome .", "Birds fly ."),
    List("execution-train-1", "execution-train-2", "execution-train-3"), role: "training")
let training = nlp.annotate(nlp.tokenize(source),
    List(List("PROPN", "VERB", "PROPN", "PUNCT"),
         List("PROPN", "VERB", "PROPN", "PUNCT"),
         List("NOUN", "VERB", "PUNCT")), task: "pos")
let held_out = nlp.corpus(List("Bob visits Paris .", "Alice visits Rome ."),
    List("execution-validation-1", "execution-validation-2"), role: "validation")
let validation_tokens = nlp.tokenize(held_out)
let gold = nlp.annotate(validation_tokens,
    List(List("PROPN", "VERB", "PROPN", "PUNCT"),
         List("PROPN", "VERB", "PROPN", "PUNCT")), task: "pos")
let model = nlp.fit_tagger(training, epochs: 5, seed: 731, execution: execution)
let prediction = model.tag(validation_tokens, execution: execution)
let report = model.evaluate_tagger(gold, execution: execution)
print("Measured synthetic POS token accuracy")
print(report.accuracy())
print("Original source tokens, labels and execution trace")
print(validation_tokens.token_texts(0))
print(prediction.tags(0))
print(prediction.execution_summary())
print(report.execution_summary())

// Learning order and averaging state continue; changing the worker count does
// not authorize a different arithmetic/backend contract at resume.
let first = nlp.fit_tagger(training, epochs: 2, seed: 731, execution: execution)
first.save("execution-pos-stage.anertagger")
let restored = nlp.load_tagger("execution-pos-stage.anertagger")
let continued = restored.resume_tagger(training, epochs: 3, execution: execution)
print("Same-profile checkpoint continuation reproduces complete training")
print(continued.fingerprint() == model.fingerprint())
print(continued.tag(validation_tokens, execution: execution).fingerprint() == prediction.fingerprint())

// Fitting lexical policy remains a host operation. Local execution selects
// only admitted numeric phases; Unicode, CSR assembly and ranking stay CPU.
let encoder = nlp.vectorizer(source)
let features = encoder.transform(held_out, execution: execution)
let matrix = features.dense(max_cells: 1000, execution: execution)
print("Sparse row IDs, OOV counts and local numeric phase trace")
print(features.ids())
print(features.oov())
print(features.execution_summary())
let memory = nlp.index(source)
let hits = memory.search("Paris visits", top_k: 2, execution: execution)
print("Ranked evidence and BM25 contribution trace")
print(hits.ids())
print(hits.scores())
print(hits.execution_summary())