Fix: make the per-turn token metric readable in every theme

[DevMando]

Summary

The token count shown after each turn was effectively invisible in several themes. It now sits in a small pill with its own background rather than being faint text at the edge of the pane, and the corner shading on the themes that imitate a screen has been eased.

What was happening

The metric was drawn as pale text, right-aligned. That combination put the smallest and palest element in the application in the exact spot where the themes that imitate a display are darkest — the shading around the edge of a simulated screen.

Measured against the contrast the text needed, it was reaching well under half of it. On the CRT and amber terminal themes it was worse than a 2:1 ratio, which in practice means it was not readable at all. On the vacuum-fluorescent, vector scope, microfiche and LCD themes it was better but still short.

Worth recording for anyone who revisits this: raising the text's colour would not have fixed it. Even at full brightness the metric still failed on two themes, because the corner shading was strong enough to defeat any colour the palette could offer.

That points at a real distinction. An earlier change added a contrast floor for themes viewed through an overlay, and that works for the CRT's scanlines, which dim the whole pane evenly. Corner shading is not even — it varies by position, so the very same colour is comfortably readable in the middle of the window and unreadable near the edge. No palette-level rule can answer that.

What changed

• The metric is now a pill. It carries its own background and border, so it reads as a deliberate piece of the interface rather than text trailing off the edge, and it stays readable over a busy chat background as well. This follows the treatment the live plan status card already uses for the same reason.
• It is de-emphasised by size and containment rather than by being faint. That is the more durable choice: something small and bounded stays readable no matter what is drawn over it, whereas something faint stops being readable the moment anything is.
• The corner shading was too heavy on all four affected themes and has been eased. It had been judged by eye against ordinary body text, which has a great deal of contrast to spare, so nothing revealed that it was consuming everything fainter. Easing it also helps ordinary text near the edges of the pane, not just this metric.

Scope and risk

Low, and limited to the transcript's appearance. No change to what the metric reports, when it appears, or how it is calculated.

One thing a reviewer should be aware of: the pill is deliberately not wrapped in the container class used by the existing status pills. That class is filtered out when a saved session is reopened, on the grounds that status pills describe the live state of a process that has since ended. A token count is the opposite — it is what that turn actually cost, and it belongs to the conversation's history. Reusing that class would have silently removed these from every restored session. There is a comment at the call site recording this, since the two are otherwise near-identical and the trap is easy to walk into again.

Verification

Builds clean, 0 warnings and 0 errors. Full Desktop suite passes, 423 of 423 — unchanged, as this touches presentation only.

Contrast was measured numerically before and after for each affected theme, including compositing the corner shading over both the text and its background, which is what established that a colour change alone was insufficient.

Not covered by tests: the transcript renders in a WebView and is not reachable from the test project. A reviewer should look at Cathode Ray and Pheteven Phosphor, the two worst cases, and confirm the metric is legible in the bottom-right corner — and ideally with a chat background image set, since that is the other case the pill is meant to survive.