Unbounded prompt construction in ReduceSynthesis and GroupSynthesis

[marklubin]

Problem

ReduceSynthesis and GroupSynthesis concatenate all input artifacts into a single prompt string with no token/size limit:

artifacts_text = "\n\n---\n\n".join(f"### {a.label}\n{a.content}" for a in sorted_inputs)

For large corpora (hundreds or thousands of artifacts), this will exceed context windows and cause hard failures with no useful error message.

Impact

• O(total_input_size) memory and token usage
• Silent failures when prompt exceeds model context window
• Affects both ReduceSynthesis (all inputs → 1 output) and GroupSynthesis (per-group concatenation)

Suggested approach

• Add an optional max_tokens or context_budget parameter (like CoreSynthesis has)
• Truncate or chunk inputs when they'd exceed the limit
• Log a warning when truncation occurs
• Consider automatic chunking with recursive reduce for very large input sets

Scope

This is a scale optimization — the current implementation works correctly for typical pipeline sizes (dozens of artifacts). It only becomes a problem with large corpora.