Supermemory provider ignores options.limit and returns 30 results while every other provider returns 10 — cross-provider comparisons are not like-for-like #64

Description

runSearchPhase asks every provider for the same number of results:

// src/orchestrator/phases/search.ts:58-62
const results = await provider.search(question.question, {
  containerTag,
  limit: 10,
  threshold: 0.3,
})

Every provider honours that limit except Supermemory, which hardcodes 30:

Note the Supermemory call also accepts options.threshold on the very next line, so the hardcoded limit reads as an oversight rather than a deliberate provider capability.

Impact

The whole point of the tool is the README's "🆚 Multi-provider comparison: run the same benchmark across providers side-by-side". That comparison is currently unfair in three measurable ways:

1. Answer quality. runAnswerPhase feeds the full stored result set to the answering model as context (answer.ts:118-123). Supermemory supplies 30 candidate memories; competitors supply 10. More retrieved context is a direct accuracy advantage that has nothing to do with retrieval quality.
2. MemScore token component. contextTokens is derived from the prompt built out of those results (answer.ts:125-130). Supermemory's token count is inflated by ~3× the context, so its MemScore token figure is not comparable to the others' either.
3. Retrieval metrics. calculateRetrievalMetrics defaults to k = 10 and slices (retrieval-eval.ts:93-95), so Supermemory is scored on the best 10 of 30 candidates while others are scored on the only 10 they were allowed to return. Precision@10, MRR and NDCG are computed over different candidate pools.

This matters especially because the benchmark is published by the vendor of one of the providers under comparison — the numbers should be defensible on their face.

Suggested fix

-      limit: 30,
+      limit: options.limit || 30,

If a different limit per provider is ever intentional, it belongs in the Provider interface as a declared capability and must be surfaced in the report, not buried in a call site.