fix(tokens): image blocks as flat ~1500 tokens; compute real contextUsagePct

[KillerQueen-Z]

Summary

Three related issues that combined to make /context and the renderer's context-window ring widely inaccurate on image-bearing sessions.

Reproduced empirically: same 4-message session with one ~100KB image,

• Before this fix: /context shows ~75k / 200k (37.8%)
• After this fix: /context shows ~1.9k / 200k (1.0%) ← matches Anthropic's true count

40x discrepancy.

Root causes

1. `estimateContentPartTokens` JSON.stringify-ed image blocks

When `tool_result.content` was a `[{text}, {image}]` array, the whole array was `JSON.stringify`-ed and the base64 `data` field counted as text. A single normalized image (~140KB base64) became ~70k phantom tokens. Anthropic actually bills `(w*h)/750` ≈ 1100-1500 per image.

Fix: walk the content array block-by-block. Text blocks count as text; image blocks count as a flat 1500 tokens (close to Anthropic's real billing — `Read` tool caps long edge so normalized images land near 1024×768 ≈ 1050 tokens). Unknown block types still stringify, but with `source.data` redacted to `` so future block kinds don't regress.

2. `getAnchoredTokenCount` returned `contextUsagePct: 0` always

Both return paths hardcoded the field. Agent loop emits this verbatim via `kind: 'usage'` events, so the desktop/extension renderer's context ring sat at 0% regardless of real fullness. `/context` was unaffected because the CLI command re-derives `pct` from `estimated` itself.

Fix: compute `(estimated / contextWindow) * 100` using the current model's window from `getContextWindow(_currentModel)`.

3. `loop.ts` rounded `contextPct` to integer

A 200-message session at 0.4% rounded to 0 and froze the renderer's ring. Match `/context`'s `.toFixed(1)` fidelity by keeping one decimal.

Test plan

• `npm run build` — passes
• Manual: ran image-bearing session, compared `/context` before/after
• Verified renderer ring updates correctly on short conversations

Impact

• Wallet deductions: NOT affected — gateway has its own (working) image handling via `imagePlaceholder` and uses its own input estimate.
• CLI `/context` display: more accurate — matches Anthropic's real `input_tokens` within ~5%.
• Desktop/extension context ring: now functional — was stuck at 0%.
• `/compact` decisions: less spurious — pre-fix, one image could push the perceived fullness past internal thresholds.