chat using instance instead of class for downstream connectivity query

[Clare72]

[Robbie1977]

Two separate faults here, both fixed in #69.

1. Why it picked an instance. VFB's search ranks the ~500 KCg-* hemibrain individuals above the class whose symbol is KCg, so with 30 rows the resolver never saw gamma Kenyon cell at all and its token-superset fallback took the first individual carrying the token. The class lift added for #59 only applied to names like "EPG neurons"; it now also applies to a bare symbol that matched no document exactly when an exact match exists among classes. KCg is the symbol of two classes — adult gamma Kenyon cell (FBbt_00049828) and gamma Kenyon cell (FBbt_00100247) — and nothing in the question picks one, so the chat now asks: "KCg" matches more than one VFB term: adult gamma Kenyon cell or gamma Kenyon cell. Which did you mean? with a chip for each. A stage word in the question ("adult KCg") settles it without asking, and so does the model when the question itself decides (a dataset, a qualifier).

2. Why the class answer was still wrong. With the class resolved, run_query DownstreamClassConnectivity came back count: 0 and the chat said VFB holds no downstream connectivity for gamma Kenyon cells. That zero is a leftover of the pre-1.22.51 silent-zero bug, still cached at the v3-cached edge under the URL shapes the MCP uses:

run_query?id=FBbt_00100247&query_type=DownstreamClassConnectivity&offset=0&limit=25     -> count 0 (HIT)
run_query?id=FBbt_00100247&query_type=DownstreamClassConnectivity&offset=0&limit=2500   -> count 0 (HIT)
run_query?id=FBbt_00100247&query_type=DownstreamClassConnectivity                        -> count 3886 (the slot refreshed on 4 Sep)

The chat now treats a class-connectivity count: 0 like a -1: one forced recompute on the existing per-request budget, and a second zero is believed. That protects users, but the stale slots should still be purged — every (Downstream|Upstream)ClassConnectivity body with "count": 0 cached before 4 Sep is suspect, on every cache instance (a force-refresh through one instance does not update the others' disks, which is why the plain-URL refresh only half-worked).

Verified live: the question in the screenshot now asks which gamma Kenyon cell class was meant; choosing adult gamma Kenyon cell answers with 3,370 downstream classes ranked by synapses per connected pair (APL, DPM, MBON05, MBON01 leading).