desktop: trim old screenshot images from pi-mono floating-bar session history

[beastoin]

Problem

The floating-bar chat attaches a fresh ~500 KB WebP screenshot to every user message. Pi-mono stores the whole conversation as a message list and re-POSTs every prior message (including images) to /v2/chat/completions on every turn. After enough turns, the accumulated images bloat the request body.

Symptom (before raising the axum limit in #6965):

[app] ScreenCaptureManager: Screenshot captured 5120x2880, WebP 502 KB
[agent] Reusing pi-mono session: pi-session-1 (key=floating)
[agent] Pi-mono: including screenshot image in prompt (image/webp)
[pi-mono] turn_end ERROR: 413 Failed to buffer the request body: length limit exceeded

Math per turn (N prior images):

• per-image: 502 KB × 4/3 ≈ 669 KB base64
• body at turn 3: ~2.06 MB → exceeds axum's 2 MB default → 413

Interim mitigation (shipped)

PR #6965 raises the backend request body limit on /v2/chat/completions from 2 MB to 16 MB. That buys roughly 20 accumulated screenshots of headroom — enough for any realistic floating-bar session — but doesn't fix unbounded growth.

Proposed fix

In desktop/agent/src/adapters/pi-mono.ts, before serializing messages to the provider, walk the conversation and replace older image content blocks with a short text placeholder. Keep only the most recent image (that's the current screen state — the only one the model actually needs).

Pseudocode:

function stripOldImages(messages) {
  // Find the index of the LAST message that contains an image block
  const lastImageIdx = messages.findLastIndex(msg =>
    Array.isArray(msg.content) && msg.content.some(b => b.type === "image")
  );
  return messages.map((msg, i) => {
    if (i === lastImageIdx) return msg; // keep the most recent
    if (!Array.isArray(msg.content)) return msg;
    return {
      ...msg,
      content: msg.content.map(b =>
        b.type === "image"
          ? { type: "text", text: "[earlier screenshot omitted]" }
          : b
      ),
    };
  });
}

This:

• Drops redundant visual data that the model doesn't need (the model is looking at the current screen)
• Preserves full conversational text context
• Keeps steady-state body size roughly constant regardless of turn count
• Doesn't touch Swift — image quality stays full

Acceptance

• After N ≥ 10 floating-bar turns in a single session, body size remains well under 2 MB
• Model still responds coherently to follow-ups ("you said earlier that...") — text history intact
• No regression on main-chat or ACP flows (only applies to pi-mono)
• Unit test covering the replacement logic