fix: avoid duplicate spans when a full fetch follows a typed fetch

[Nishchay Mahor (NishchayMahor)]

What

CachedSpanFetcher.get_spans() returns duplicate spans when a type-filtered fetch is followed by a full (unfiltered) fetch:

await fetcher.get_spans(["llm"])   # caches llm-1, llm-2
await fetcher.get_spans()          # -> fn-1, llm-1, llm-1, llm-2, llm-2  ❌

The no-filter branch calls _fetch_spans(None), which re-fetches every span, and _fetch_spans always appends into self._span_cache[type]. Since llm was already cached from the first call, its spans get appended a second time. Scorers that call trace.get_spans([...]) and then trace.get_spans() get doubled results.

Fix

A full fetch returns the authoritative complete set, so reset the per-type cache before it:

if not span_type or len(span_type) == 0:
    self._span_cache = {}
    await self._fetch_spans(None)
    ...

The typed-fetch path is unchanged (it only fetches types not already cached, so it never double-appends). An alternative would be to make _fetch_spans replace per-type lists instead of appending; I went with the smaller, more localized change, but happy to switch if you'd prefer that.

Testing

Added test_fetch_all_after_typed_fetch_has_no_duplicates to TestCachedSpanFetcher using the existing offline fetch_fn= injection (no API key/network). pytest -k CachedSpanFetcher → 17 passed; the new test fails on main and passes with the fix.

[Abhijeet Prasad (AbhiPrasad)]

thanks!