feat: derive Anthropic max_tokens from context budget instead of hardcoding 4096

[AndrewDongminYoo]

What

getProviderConfig() now derives the default Anthropic max_tokens from the configured context window instead of letting the wire request fall back to 4096.

This completes a follow-up that context-budget.ts already documents as pending:

// The response reserve is a "passive" reservation: we don't pass
// `max_tokens: responseReserve / 3` to the LLM (yet — that's a
// follow-up). We just refuse to fill above (maxCtx - responseReserve)

Why

computeContextBudget() sizes a 15% response reserve so the model has room to answer, but that reserve was passive — it only capped how full we packed the context. The actual Anthropic request still defaulted to max_tokens: 4096 regardless of the configured context window, so a large context could be assembled but the response was silently clipped to 4096 tokens.

How

• Compute anthropicBudgetTokens from the same responseReserve (chars → tokens at ~3 chars/token), capped at 16 384 to stay within typical per-request limits.
• Pass it as the default max_tokens for all three Anthropic-wire providers (anthropic, minimax, custom/anthropic_messages).
• Explicit caller overrides (e.g. max_tokens: 300 for routing decisions) are spread after the default, so they still take precedence.
• Update the now-outdated context-budget.ts doc comment to reflect that the follow-up is done.

Effect on a 200K-char context: default max_tokens 4096 → 10 240.

Testing

• src/lib/__tests__/llm-providers.test.ts — 50 pass / 0 fail (updated to assert the derived default and that overrides win).