feat(gui): readable figures in Connections and Statistics, and two table fixes

[donislawdev]

Five changes to the two tables and the Statistics page, each with its own commit.

Traffic in Connections is shown in the unit that fits the number

The five traffic columns and the footer under them were fixed at bytes / 1024 with one decimal.
A 5 GB flow therefore read 5242880.0 - nine digits with no grouping - and a ninety-byte DNS
answer read 0.0, which says nothing happened on a row that exists because something did. One
fixed unit cannot serve both the user whose traffic is megabytes and the one whose traffic is
kilobytes, and the connection log holds both in the same table.

Every cell now carries whichever of B, KB, MB or GB fits its own value (utils.human_bytes,
1024-based, three significant digits), so the [KB] has left the headers and the five column
tooltips no longer promise kilobytes.

Per cell rather than per column, and that is the decision worth recording. A column-wide unit
taken from the largest row keeps the column arithmetically comparable, but at 200 000 flows it
renders everything below the first screen as 0.0. Per-cell pays for that in cross-row reading
instead, and it pays where nobody looks: the pair actually read across a row is down against
down seen, the same traffic before and after impairment, within an order of magnitude even at
99% loss, while down against up differs thousandfold and is never read as a pair. Three
significant digits cap a cell at seven characters, so the columns stay right-aligned and
scannable. avg keeps raw bytes: a packet size is 40-65535 B, where 1460 beats 1.43 KB.

Unaffected, and checked rather than assumed: sorting works on the raw numbers in views.DERIVED,
both CSV exports stay raw bytes, the search matches process, protocol, address and ports only, and
the kb column id keeps its name because it is written into the saved layout of every user who has
sorted this table. Ctrl+C copies what the cell shows, so a traffic cell now arrives as 5.24 GB;
the machine-readable path remains the CSV export.

Also in this branch

• A header tooltip describes its own column, hidden ones or not. identify_column answers with
  a display position, and the resolution indexed that into the full column list, so hiding one
  column made every header to its right explain its neighbour. Tk is asked to translate now. The
  event log gets the fix for free.
• The figures on the Statistics page can be copied - a context menu per value, and buttons for
  the whole session panel or the whole counter grid. The text is built from the same registries the
  screen renders from, so it cannot drift from what is on screen.
• The search box on the Control page sits against the page margin instead of floating in the
  middle of it.

Verification

• Full suite: 1164 passed, twice (before and after the last edits). GUI smoke and the real-Tk
  render check at 1366x768 in both languages: OK.
• Real Tk, both languages, a snapshot spanning 0 B to 5 GB: nothing clips, the widest cell being
  1023 MB and the tightest column keeping 23 px of slack. The fake tkinter cannot answer this -
  it has no font metrics and no column geometry.
• A repaint costs 2.1x what it did (0.28-0.43 ms against 0.13-0.21 ms at a 23-slot viewport,
  measured paired because the absolutes drifted by half between runs on an idle machine). That is
  once per 700 ms tick.
• Mutation registry: three new entries for this change, all caught; the full registry run is green
  with the canary reporting BROKEN as it must.

Not done

• Digit grouping for packets and dropped, which are also long at 200 000 flows.
• The Statistics session panel keeps its fixed MB, and the chart axis its fixed KB/s.
• The decimal separator stays a dot everywhere, as it is in the rest of the program.

🤖 Generated with Claude Code