-- Build one explicitly point-in-time row per episode. All predictors carrying
-- history use a window ending at one episode preceding the target.

CREATE OR REPLACE TABLE episode_history AS
WITH history_windows AS (
    SELECT
        e.*,
        count(*) OVER prior_all AS prior_episode_count,
        coalesce(sum(attempt_count) OVER prior_all, 0) AS prior_total_attempt_count,
        count(first_outcome) OVER prior_all AS prior_labeled_episode_count,
        lag(attempt_count) OVER by_vehicle AS prior_attempt_count,
        lag(first_outcome) OVER by_vehicle AS prior_first_outcome,
        lag(final_outcome) OVER by_vehicle AS prior_final_outcome,
        lag(episode_start) OVER by_vehicle AS prior_episode_start,
        max(CASE
                WHEN first_outcome IN ('fail', 'reject', 'abort')
                THEN episode_start
            END) OVER prior_all AS prior_adverse_episode_start,
        count(*) FILTER (WHERE first_outcome = 'pass') OVER prior_all
            AS prior_pass_count,
        count(*) FILTER (WHERE first_outcome = 'fail') OVER prior_all
            AS prior_fail_count,
        count(*) FILTER (WHERE first_outcome = 'reject') OVER prior_all
            AS prior_reject_count,
        count(*) FILTER (WHERE first_outcome = 'abort') OVER prior_all
            AS prior_abort_count,
        count(*) FILTER (
            WHERE first_outcome IN ('fail', 'reject', 'abort')
        ) OVER prior_all AS prior_nonpass_count,
        count(first_outcome) OVER prior_three AS last3_labeled_episode_count,
        count(*) FILTER (
            WHERE first_outcome IN ('fail', 'reject', 'abort')
        ) OVER prior_three AS last3_nonpass_count,
        max(max_events_in_day) OVER prior_all AS prior_max_events_in_day,
        arg_max(episode_last_observed_make, episode_number)
            FILTER (WHERE episode_last_observed_make IS NOT NULL)
            OVER prior_all AS last_observed_make,
        arg_max(episode_last_observed_model, episode_number)
            FILTER (WHERE episode_last_observed_model IS NOT NULL)
            OVER prior_all AS last_observed_model,
        arg_max(episode_last_observed_model_year, episode_number)
            FILTER (WHERE episode_last_observed_model_year IS NOT NULL)
            OVER prior_all AS last_observed_model_year
    FROM inspection_episodes AS e
    WINDOW
        by_vehicle AS (
            PARTITION BY vehicle_token
            ORDER BY episode_number
        ),
        prior_all AS (
            PARTITION BY vehicle_token
            ORDER BY episode_number
            ROWS BETWEEN UNBOUNDED PRECEDING AND 1 PRECEDING
        ),
        prior_three AS (
            PARTITION BY vehicle_token
            ORDER BY episode_number
            ROWS BETWEEN 3 PRECEDING AND 1 PRECEDING
        )
)
SELECT
    *,
    CASE
        WHEN prior_episode_start IS NULL THEN NULL
        ELSE extract(epoch FROM (episode_start - prior_episode_start)) / 86400.0
    END AS days_since_prior_episode,
    CASE
        WHEN prior_adverse_episode_start IS NULL THEN NULL
        ELSE extract(epoch FROM (episode_start - prior_adverse_episode_start))
             / 86400.0
    END AS days_since_prior_adverse,
    prior_nonpass_count::DOUBLE / nullif(prior_labeled_episode_count, 0)
        AS prior_nonpass_rate,
    prior_pass_count::DOUBLE / nullif(prior_labeled_episode_count, 0)
        AS prior_pass_rate,
    last3_nonpass_count::DOUBLE / nullif(last3_labeled_episode_count, 0)
        AS last3_nonpass_rate
FROM history_windows;

CREATE OR REPLACE TABLE feature_mart AS
SELECT
    vehicle_token,
    vehicle_bucket,
    vehicle_bucket < 10 AS is_vin_audit,
    episode_number::BIGINT AS episode_number,
    episode_start,
    first_outcome,
    first_outcome_label_source AS target_outcome_label_source,
    CASE
        WHEN first_outcome = 'pass' THEN 0
        WHEN first_outcome IN ('fail', 'reject', 'abort') THEN 1
        ELSE NULL
    END::INTEGER AS target_nonpass,
    episode_start >= TIMESTAMP '2016-01-01'
        AND first_outcome IS NOT NULL
        AND prior_episode_count >= 1
        AND prior_total_attempt_count <= 50
        AND coalesce(prior_max_events_in_day, 0) <= 4
        AS eligible_returning_target,
    CASE
        WHEN episode_start < TIMESTAMP '2016-01-01' THEN 'historical_context'
        WHEN episode_start < TIMESTAMP '2023-01-01' THEN 'train'
        WHEN episode_start < TIMESTAMP '2024-01-01' THEN 'tune'
        WHEN episode_start < TIMESTAMP '2025-01-01' THEN 'calibrate'
        WHEN episode_start < TIMESTAMP '2026-01-01' THEN 'test'
        WHEN episode_start < TIMESTAMP '2027-01-01' THEN 'shadow'
        ELSE 'out_of_scope'
    END AS temporal_partition,
    CASE
        WHEN episode_start < TIMESTAMP '2016-01-01' THEN 'before_target_period'
        WHEN first_outcome IS NULL THEN 'unlabeled_first_outcome'
        WHEN prior_episode_count < 1 THEN 'cold_start'
        WHEN prior_total_attempt_count > 50 THEN 'prior_event_count_over_50'
        WHEN coalesce(prior_max_events_in_day, 0) > 4
            THEN 'prior_daily_activity_over_4'
        ELSE NULL
    END AS eligibility_exclusion_reason,
    public_county,
    source_era,
    CASE
        WHEN month(episode_start) IN (12, 1, 2) THEN 'winter'
        WHEN month(episode_start) IN (3, 4, 5) THEN 'spring'
        WHEN month(episode_start) IN (6, 7, 8) THEN 'summer'
        ELSE 'fall'
    END AS target_season,
    CASE
        WHEN last_observed_model_year BETWEEN 1886 AND year(episode_start) + 1
        THEN greatest(year(episode_start) - last_observed_model_year, 0)
        ELSE NULL
    END::INTEGER AS vehicle_age,
    prior_episode_count,
    prior_total_attempt_count::BIGINT AS prior_total_attempt_count,
    prior_attempt_count,
    days_since_prior_episode,
    days_since_prior_adverse,
    prior_nonpass_rate,
    prior_first_outcome,
    prior_final_outcome,
    last_observed_make,
    last_observed_model,
    last_observed_model_year,
    prior_labeled_episode_count,
    prior_pass_count,
    prior_fail_count,
    prior_reject_count,
    prior_abort_count,
    prior_nonpass_count,
    prior_pass_rate,
    last3_labeled_episode_count,
    last3_nonpass_count,
    last3_nonpass_rate,
    prior_max_events_in_day,
    prior_first_outcome IS NULL AS prior_first_outcome_missing,
    last_observed_model_year IS NULL AS prior_model_year_missing,
    (SELECT source_data_kind FROM build_config) AS source_data_kind,
    (SELECT population_estimate_allowed FROM build_config)
        AS population_estimate_allowed
FROM episode_history;

CREATE OR REPLACE TABLE cohort_flow AS
SELECT
    'event' AS grain,
    'input' AS stage,
    'all_rows' AS reason,
    count(*) AS observation_count,
    count(DISTINCT try_cast(vehicle_token AS VARCHAR)) AS vehicle_count
FROM stg_events

UNION ALL

SELECT
    'event',
    'excluded',
    exclusion_reason,
    count(*),
    count(DISTINCT vehicle_token)
FROM event_exclusions
GROUP BY exclusion_reason

UNION ALL

SELECT
    'event',
    'accepted',
    'clean_events',
    count(*),
    count(DISTINCT vehicle_token)
FROM clean_events

UNION ALL

SELECT
    'episode',
    'constructed',
    'all_episodes',
    count(*),
    count(DISTINCT vehicle_token)
FROM inspection_episodes

UNION ALL

SELECT
    'episode',
    'eligible_returning_target',
    coalesce(eligibility_exclusion_reason, 'eligible'),
    count(*),
    count(DISTINCT vehicle_token)
FROM feature_mart
GROUP BY coalesce(eligibility_exclusion_reason, 'eligible');