docs(telemetry): document the seconds-vs-ms epoch boundary assumption (#121)

[KRoperUK]

Summary

_coerce_timestamp disambiguates seconds from milliseconds purely by magnitude (> 1e12 → divide by 1000). That's safe for the data it parses — a Unix seconds epoch doesn't reach 1e12 until the year ~33658, and the cloud emits second-resolution TS — but the assumption was undocumented, so a future reader could mistake the one-sided heuristic for a bug. Closes #121 (low priority, as filed).

Per the issue's accepted resolution ("a comment documenting the boundary assumption, or a test pinning the split"), this does both:

• A comment explaining why 1e12 is safe for second-epochs, and that the converse edge (a pre-2001 millisecond epoch < 1e12, which the cloud has never produced) is intentionally not defended.
• A test pinning each side of the threshold. It also captures a non-obvious truth surfaced while writing it: a value just below 1e12, read as seconds, lands in year ~33658 which datetime cannot represent, so the existing OverflowError/ValueError guard drops it (returns None) rather than silently mis-parsing.

Type of change

• Bug fix
• New feature
• Breaking change
• Documentation / chore (comment + characterization test; no behaviour change)

Checklist

• ruff check + ruff format pass
• mypy strict passes (9 source files)
• pytest passes (201); the new test pins the boundary behaviour
• No behaviour change — this documents and guards existing behaviour

Behaviour changes worth calling out

None. Comment + test only.