fix(stats): track ImageGen / VideoGen costs in usage insights

[KillerQueen-Z]

Summary

Before this, only chat LLM calls flowed through recordUsage(). The ImageGen and VideoGen tools wrote to the content library but never reported to the global stats file, so:

• franklin insights undercounted total 30-day spend by the exact amount spent on image + video generation.
• The "Top models" breakdown never surfaced openai/gpt-image-1, bytedance/seedance-2.0, xai/grok-imagine-video, etc. even when they were the dominant cost in a session.
• Monthly projections in insights were similarly low.

Changes

On successful generation each tool now fires a fire-and-forget recordUsage(model, 0, 0, costUsd, 0):

• Input/output tokens stay at 0 — image/video models don't bill by token.
• The cost is pulled from the live gateway catalog via findModel + estimateCostUsd, the same path the AskUser cost preview uses, so the recorded amount matches the amount the user was quoted.
• When the catalog lookup fails (offline / stale cache), the tools fall back to a hardcoded estimate (ImageGen: 0, VideoGen: the legacy $0.05/s) so stats still get an entry — under-approximate is better than missing.

Design notes

• Fire-and-forget, error-swallowing on purpose. A ~/.blockrun/stats.json write failure must not turn a paid generation into a user-visible error, and stats accuracy matters less than not losing the output.
• No schema change. The existing recordUsage signature handles inputTokens=0 + outputTokens=0 + explicit costUsd fine; nothing in the stats tracker or the insights engine needs to learn about media models specifically.
• Scope deliberately small — only the two tools, +33 lines across 2 files, zero net changes to any chat flow.

Test plan

1. Run franklin insights → note baseline Total Cost and Top Models list.
2. Generate one image with a known-cost model, e.g.

   franklin --prompt "use zai/cogview-4 to generate a tiny red dot"
   

   (cogview-4 is catalog-priced at $0.015/img.)
3. Run franklin insights again.
   • Total Cost should increase by ~$0.0158 (0.015 + 5% gateway margin).
   • Top Models should contain a new row for zai/cogview-4 with 1 request.
4. Repeat for VideoGen with any Seedance variant and confirm a per-second entry is written.

Verified locally on Base chain. The VS Code extension's Insights panel (which reads the same stats.json) also immediately reflected the new entries.

Related

• The same stats.json is consumed by the VS Code extension's Insights panel (via generateInsights), so this fix lands media costs in both the CLI and the extension with a single change.