Fix review comments: backtest-only fold coverage, null_converter safety, CHRONOS_MAX_RETRIES clarity, retry tests

[Copilot]

Addresses four reviewer concerns from the previous PR: a type-safety bug in null_converter, ambiguous retry documentation, fold coverage logic incorrectly including future forecast IDs, and missing unit tests for the retry path.

null_converter() — type-safe NULL string check

x == "NULL" coerces non-character scalars and can propagate NA into an if condition. Changed to identical(x, "NULL"):

# before — unsafe for numeric/logical/NA inputs
if (is.na(x) || x == "NULL")

# after — exact, no coercion
if (is.na(x) || identical(x, "NULL"))

CHRONOS_MAX_RETRIES — clarify retry vs attempt semantics

CHRONOS_MAX_RETRIES = 3 means 3 retries, 4 total attempts. Updated the constant comment and send_chronos_request docstring to state this explicitly. Loop behavior is unchanged.

final_models.R — exclude future forecast ID from fold coverage

train_test_id_list includes Train_Test_ID == 1 (future forecast). Using it as the denominator for back test fold coverage unfairly excluded models that produced back tests but not future output. Changed to:

back_test_id_list <- setdiff(train_test_id_list, 1)
expected_back_test_fold_count <- length(back_test_id_list)

Also vectorized the per-model warning (list-columns via summarise/mutate) to avoid O(N×M) repeated filtering, added expected_back_test_fold_count == 0 edge-case handling, and ensured the empty partial_models_tbl carries all expected columns.

New tests

• send_chronos_request: 5 tests with mocked httr::POST and a no-op backoff_wait — connection error exhaustion (exactly 4 attempts), 429→success, 500→success, non-retryable 4xx (1 attempt), and total-attempt boundary (CHRONOS_MAX_RETRIES + 1).
• null_converter: 7 tests covering NULL, zero-length, "NULL" string, NA variants, normal values, multi-element vectors, and non-character scalars.