[wu-sheng]

Fixes from the release review of 0.6.0: repack, the OTLP export, invalid UTF-8 in landed data, and the container image.

Repack

• The idle clock. A merged file kept its first source file's collected time, and files were numbered slot by slot,
  so the newest file of a repacked session could carry a time weeks old. Retention reads that header, so it could
  remove or retire a session that was still active. Each new file now takes the time of the newest file its records
  came from, and files are numbered in the order their records landed; the source and the destination are equally
  idle.
• The budget. Repack counted only record lines, not the header and closing line, so a file could exceed
  max_delta_bytes with ordinary records. It now cuts with sessiondata.Cut, as collectors do.
• Order. Two files merge only when they are next to each other and their headers match, so a header that changes
  and changes back keeps its records in order.
• Cursors. Only <prefix>.cursor was carried, so a second recorder root restarted from offset zero. Every cursor
  is carried, its last_seq mapped to the file holding its own source's last record, and a relative origin, a
  recorder inside the root, gets the origin and cursor name a collector into the destination gives it.

Invalid UTF-8

A tool input holding a byte that is not valid UTF-8 landed as it was, and the OTLP export can never send such a file:
protobuf strings must be UTF-8. Every batch holding it failed on every pass, and the file after it was skipped. Now
such data is left out at landing, its size and SHA-256 kept, the part omitted (or too_large when also nested too
deep), with a dropped entry saying why; a change or execution record is written again with U+FFFD. The export
encodes each file alone, reports a file it cannot send under its own name, and never skips the next. CLAUDE.md's
"payload bytes are the source bytes" invariant names the exception.

The OTLP client

A 2xx answer whose body did not arrive whole, or was longer than 1 MiB, was taken as success and the batch recorded
sent. It is now an error, and the batch is sent again.

The container image

The image runs as uid 65532 and declared /asz/data without creating it, so a new named or anonymous volume was
root's and the first collection failed with permission denied. The image now creates it, owned by 65532.

Checked

• make check passes, tests/chain alone with a longer timeout after it timed out under load; the earlier JSON engine
  job (GOEXPERIMENT=nojsonv2) passes.
• New tests cover each fix, and each guard was proven by reverting it. The image was checked with Docker 28.4.0 on a
  new named and an anonymous volume.