Time-series query engine: micros precision, RANGE frames, FIRST/LAST/EMA/RLEID, window-join, latest-on

[whilo]

Summary

Adds the time-series feature set surveyed across DuckDB / TimescaleDB / QuestDB / kdb / Polars / datajure. Each phase is its own commit; tests cover every new operator and the legacy seconds-based code paths still behave exactly as before.

Phase A — Sub-day temporal precision (epoch-microseconds, DuckDB-compatible)

• New column metadata: :temporal-unit ∈ #{:days :seconds :millis :micros} (defaults to legacy :seconds).
• Java SIMD kernels for micros: arrayDateTrunc{Year,Month,Day,Hour,Minute,Second,Milli,Micro}Micros, arrayExtract{Hour,Minute,Second,Millisecond,Microsecond}Micros, arrayDateAddMicrosMicros, arrayDateAddMonthsMicros, arrayDateDiffMicros.
• Both vectorized (eval-expr-{to-long,vectorized}) and scalar (gb/eval-agg-expr) paths dispatch on :temporal-unit via the existing expr/*columns-meta* dynamic var (filter widened to retain temporal-unit columns alongside dict-encoded ones).
• The same metadata flows through both the planner and the legacy fallback.

Phase B — TIME_BUCKET, FIRST_VALUE, LAST_VALUE, NTH_VALUE

• [:time-bucket N unit col] — arbitrary-width bucketing on micros, days (with :weeks/:months), or seconds.
• FIRST_VALUE / LAST_VALUE / NTH_VALUE window functions; the latter reuses the :offset slot for n. SQL parser registers all three.

Phase C — RANGE BETWEEN INTERVAL frames + GENERATE_SERIES

• Two-pointer sliding window keyed off the (single) ascending ORDER BY column, O(N) per partition, for :sum/:count/:avg. Required for correct rolling aggregates over irregular time series.
• Fixed a long-standing partition-reset bug in compute-sliding-window-sum: at a partition boundary, the cumulative was overwritten, silently corrupting the last row of every non-final partition. Now uses a single monotonic prefix array, with partition-local endpoint subtraction localizing each query.
• stratum.api/generate-series produces a dense column suitable as :from, with a temporal form ((generate-series 0 (* 5 day-us) 1 :days :micros)) that tags the output with the right :temporal-unit. Combined with ASOF LEFT JOIN, this gives canonical gap-fill + LOCF without new join machinery.

Phase D — FILLS / LOCF and named moving aggregates

• :fills window op forward-fills NaN/NULL within partition (leading NaNs stay NaN).
• MAVG / MSUM / MMIN / MMAX / MCOUNT / MDEV (q-style sugar) expand at execute time to op OVER (ROWS BETWEEN N-1 PRECEDING AND CURRENT ROW). :min / :max gained sliding-frame ROWS branches; :mdev is moving population stddev.

Phase E — EMA and RLEID

• :ema op: per-partition exponential moving average. :offset >= 1.0 is treated as period N (α = 2/(N+1)); otherwise α directly. Initializes at first non-NaN; NaN inputs carry the previous EMA.
• :rleid: run-length-encoding group ID (handles long, double, string columns; restarts per partition).

Phase F — Window-join and LATEST ON / DISTINCT ON

• stratum.api/window-join (q wj semantics): for each left row at time t, aggregate the right rows whose time falls in [t+lo, t+hi]. Sorts both sides by their ts columns, two-pointer sweep with monotonic lo/hi pointers. SUM/AVG/COUNT use right-side prefix sums (O(1) per left row); MIN/MAX scan the slice.
• stratum.api/latest-on: most-recent-row-per-partition via ROW_NUMBER OVER (PARTITION BY … ORDER BY ts DESC) + HAVING rn=1.

Bonus

• validate-query now accepts SELECT references that point at :window outputs.
• :offset window-spec field accepts doubles too (needed for EMA's α parameter).

Tests

• New: temporal_micros_test.clj (44 assertions), range_frame_test.clj (16), window_value_test.clj (18), window_extra_test.clj (28), moving_agg_test.clj (27), generate_series_test.clj (16), temporal_join_test.clj (27).
• All existing tests pass; full sweep (query-test, sql-test, asof-join-test, planner tests, …) green.

OLAP-bench regression check (10M rows, T1)

Benchmark	main 1T / NT	feat 1T / NT
B1 TPC-H Q6	16.5 / 10.7	17.3 / 10.5
B2 TPC-H Q1	114.3 / 204.2	113.2 / 203.5
B3 SSB Q1.1	16.5 / 8.1	17.0 / 8.3
B5 Filtered COUNT	2.7 / 1.5	3.3 / 2.3
B6 Low-card group-by	17.5 / 8.3	17.9 / 8.5
SSB-Q1.2	15.8 / 8.0	16.3 / 8.1

No major regressions — five of six within run-to-run noise. B5 (Filtered COUNT NEQ) shows a small absolute slowdown (~0.6 ms 1T) likely attributable to the widened *columns-meta* filter; left as-is for follow-up since absolute time is sub-millisecond and the filter widening is needed for the temporal-unit dispatch.

Test plan

• cljfmt formatting applied (clj -M:ffix)
• Full test suite green (1094 assertions across the new + touched test files)
• OLAP T1 bench compared against main, no significant regressions
• CI passes