Reject catalog evidence timestamps outside DuckDB BIGINT range

[kstonekuan]

Current behavior

The catalog stores observation timestamps and interval bounds as signed BIGINT columns at src/hflow/catalog.py:82-95.

_normalized_intervals() and _normalized_observations() normalize NumPy scalars, check the integer type, and reject negative values. They don't enforce the signed 64-bit upper bound.

Measured on main at fbfd553:

_normalized_observations(... timestamp_ns=9223372036854775808) -> accepted
_normalized_intervals(... start_ns=9223372036854775808) -> accepted
DuckDB INSERT -> ConversionException: value is out of range for INT64

The catalog append reaches a backend conversion error after HFlow's evidence boundary has accepted the value. The error doesn't name the check, observation, interval, or timestamp field.

Pattern to copy

_normalized_measurements() rejects values the storage/query contract can't represent before any table write and includes the owning check and key in the error. The duplicate-observation regression also proves a boundary refusal leaves no committed episode file at tests/test_catalog_curation.py:199-222.

What to build

Require every stored nanosecond timestamp to fit the catalog's nonnegative signed BIGINT domain: 0 through 2**63 - 1.

Definition of done

1. Observation timestamp_ns, interval start_ns, and interval end_ns accept 0 and 2**63 - 1.
2. Each field refuses 2**63 and larger values with a contextual ValueError.
3. Python and NumPy integer scalars follow the same range rule after normalization.
4. Existing boolean, float, negative, and inverted-interval refusals stay unchanged.
5. An out-of-range value is rejected before any episode or dependent Parquet file is committed.
6. The maximum accepted value round-trips through observations_latest or intervals.
7. Add outcome-focused coverage to the existing catalog validation tests.
8. No catalog format change or new dependency is needed.

Non-goals

• Changing the catalog timestamp column type
• Supporting negative relative timestamps
• Adding range rules to unrelated numeric measurements

Validation

uv run ruff check --fix
uv run ruff format
uv run ty check
uv run pytest -q tests/test_catalog_curation.py
uv run pytest -q