fix: report input_tokens in streaming responses for accurate context tracking

[andyzengmath]

Problem

Claude Code's HUD always shows 0% context usage when using this proxy because:

1. message_start was sent before extracting data.usage from the first chunk, so input_tokens was always 0
2. message_delta events only reported output_tokens, never input_tokens
3. prompt_tokens || inputTokens has a JS falsy-zero bug — if the API returns {prompt_tokens: 0}, the || operator treats it as falsy and keeps the stale value

Fix

• Reorder usage extraction before sendStartIfNeeded() so message_start includes accurate input_tokens when available on the first chunk
• Add input_tokens to all message_delta.usage events so Claude Code can update its running context window total
• Use ?? (nullish coalescing) instead of || to correctly handle 0 as a valid token count

Testing

Verified that Claude Code's HUD now shows accurate, non-zero context percentage after this fix.

[andyzengmath]

Note: this token-reporting fix is now also included in #4 (bundled with the Serper/cache/Opus-4.8 work). This PR remains available as the standalone version if you'd prefer to merge the token fix in isolation against main — whichever you prefer. If you merge #4 first, this can be closed.

[andyzengmath]

Note: this token-reporting fix is now also included in #4 (bundled with the Serper/cache/Opus-4.8 work). This PR remains available as the standalone version if you'd prefer to merge the token fix in isolation against main — whichever you prefer. If you merge #4 first, this can be closed.

[andyzengmath]

Note: this token-reporting fix is now also included in #4 (bundled with the Serper/cache/Opus-4.8 work). This PR remains available as the standalone version if you'd prefer to merge the token fix in isolation against main — whichever you prefer. If you merge #4 first, this can be closed.