Fix crash on millisecond-scale started_at in hermes agent traces

[DaoyuanLi2816]

Summary

Parsing a hermes-format agent trace whose started_at is expressed in milliseconds crashes with OSError: [Errno 22] Invalid argument.

The bug

src/datasets/packaged_modules/json/json.py normalizes started_at before feeding it to datetime.fromtimestamp (which expects seconds), using a magnitude check to detect whether the input is already in milliseconds:

timestamp = ujson_loads(lines[i])["started_at"]
milliseconds = timestamp if timestamp <= 10_000_000_000 else timestamp * 1_000
sent_at = datetime.fromtimestamp(milliseconds, tz=timezone.utc)...

The large-magnitude branch (timestamp > 10_000_000_000, i.e. already milliseconds) should convert down to seconds by dividing by 1000 — instead it multiplies by 1000 again. For a real millisecond-scale started_at (a common convention, e.g. Date.now()-style timestamps), this produces a value representing a date roughly 56 million years in the future, and datetime.fromtimestamp raises OSError: [Errno 22] Invalid argument on Windows (OverflowError/ValueError on other platforms for the same reason), aborting the whole trace-loading path.

>>> from datetime import datetime, timezone
>>> ts = 1_780_665_768_307  # same instant as the existing seconds-form test fixture, in ms
>>> ms = ts if ts <= 10_000_000_000 else ts * 1_000
>>> datetime.fromtimestamp(ms, tz=timezone.utc)
OSError: [Errno 22] Invalid argument

The fix

Divide instead of multiply, and rename the variable to seconds (what it actually holds, and what datetime.fromtimestamp expects):

timestamp = ujson_loads(lines[i])["started_at"]
seconds = timestamp if timestamp <= 10_000_000_000 else timestamp / 1_000
sent_at = datetime.fromtimestamp(seconds, tz=timezone.utc)...

Tests

Adds HERMES_SESSION_MS — the existing HERMES_SESSION fixture with started_at re-expressed in milliseconds for the same real-world instant — and a new hermes-ms parametrized case in test_json_generate_tables_with_agent_trace_metadata, asserting the exact same sent_at output as the existing seconds-form hermes case.

Verified fail-before / pass-after against the real library (via PYTHONPATH pointed at this checkout, with teich installed so the @require_teich-gated tests actually run): the new hermes-ms case raises OSError before the fix and passes after; the two pre-existing hermes cases are unaffected either way. Full file: 36 passed.

pytest tests/packaged_modules/test_json.py -k hermes

ruff format --check and ruff check are clean.