Multi-writer append logs are not a realistic use case. Simplify the
motivation, durability, and design rationale sections to focus solely
on the live agent trace use case.

Sealing collapses intermediate PathMeta overwrites and does not
preserve line ordering, so JSONL -> JSON -> JSONL is lossy. The
actual guarantees are:
- JSON -> JSONL -> JSON is lossless: seal(stream(P)) == P
- The sealed form is stable: seal(stream(seal(L))) == seal(L)

Updated the abstract, goals, round-trip section, signatures section,
and strict parsing rationale to reflect this.

The "sealing" and "streaming" abstractions added ceremony without
value — they're just "how to read a JSONL file" and "how to write
a JSONL file." Renamed throughout:

- Sealing Algorithm -> Reading JSONL
- Streaming Algorithm -> Writing JSONL
- seal()/stream() -> read()/write() in round-trip identities
- Updated all references in abstract, goals, signatures,
  compatibility, open questions, and design rationale

- Collapse if-blocks into match guards (collapsible_match)
- Use sort_by_key with Reverse instead of sort_by (unnecessary_sort_by)

Forward compatibility matters more than strict rejection. Older
readers should produce a correct (if incomplete) view of a path
written by a newer producer, rather than refusing to read it.

Unknown lines are skipped with a warning. Version bumps remain
available for truly incompatible structural changes.