Fix what the review of the MCP and format work found

[wu-sheng]

A review of the latest changes (#44 to #48) against the OAP and Horizon found these. Each one holds on the code.

Changes

• MCP names. A Claude Code MCP tool name splits into a server and a tool only when exactly one __ follows mcp__, counting the ones that overlap. mcp__foo___bar can be foo and _bar or foo_ and bar, so it is left whole; it was split as the first. The split feeds the landed parts, the step's attrs, the metric labels and, through them, the OAP's endpoint names.
• Order of nodes on one record. A node on the whole record comes before the nodes on its parts, which follow their block, then the id. The whole record and a part were decided by id, which made a cycle (block 1 before the whole record, the whole record before block 0, block 0 before block 1), so a sort could return any of them. The OAP already orders them this way, and asz-view.md now says so.
• Where record times stop. The page read each record's time from its raw line, so it kept reading past a line that does not decode, although Session Data says a reader stops there, as the OAP does. It now decodes each record. On the largest real session, 142 MB, a conversation read takes about 0.25 s more (2.3 s to 2.55 s).
• Times at the edges. A file's and a node's first and last time keep a time before 1970; the last one used to stay at 0 and be dropped, where the OAP keeps it. A record at exactly 1970-01-01T00:00:00Z counts as none, as 0 already reads, so a file's first time no longer depends on the order Go gives a map.
• Docs. asz-view.md said the document carries no RFC 3339 strings, but a workspace change and a tool execution keep their record's time as one. session-data.md and otlp.md list execution among the file kinds. docs/en/changes/changes.md records each change.

Checked

• New tests: the MCP split for ten names, the order of three nodes on one record in all six input orders, and the page's times over a landed file with two records before 1970, a line whose off is a string, and a record after it. Each fails with its fix reverted.
• The documents of the 94 real conversations measured (65 Claude Code, 8 with MCP calls, 21 LangChain) are byte-identical before and after.
• make check passes.