Release v3.3.1 — surface uncounted (count -1) term-info queries

[Robbie1977]

Brings the post-v3.3.0 fix on feature/answer-quality-recipes into main (candidate v3.3.1).

Fix — false "VFB does not currently hold data" on large terms

get_term_info returns count: -1 for query types it did not pre-count (common for big ones — e.g. on the medulla FBbt_00003748 every query in the menu is -1). -1 means the data exists but wasn't counted, so the query must be run — it does not mean "no data". The term-info digest filter (lib/termInfoDigest.mjs) was q.count > 0 || examples, which dropped every -1 query, so the "Available VFB data" block shown to the model was empty and it answered "VFB does not currently hold data on the number of images of neurons with a part in the medulla" — when VFB actually holds 226,524.

Because the whole menu was -1, this affected essentially any question about a large region, not just images.

Changes

• buildTermInfoDigest filter now keeps count: -1 queries (q.count !== 0 || examples); only genuinely-counted-empty queries (count 0, no preview rows) are dropped.
• digestToText and the orchestrator's AVAILABLE VFB DATA block render count < 0 as an explicit instruction — run this query — call vfb_run_query with query_type X to get the count/results — instead of a bare -1, so the model knows the data exists and that it must run the query.
• Added a regression test (medulla / ImagesNeurons -1 fixture). Digest suite 9/9; 56/56 across the touched suites; next lint clean.

Conflict note

The only merge conflict was test-results/task-battery/latest.json — the CI-generated "latest run" snapshot, which the task-battery workflow rewrites on every push to both branches (main's run was on the PR #4 merge, feature's on the fix). Resolved by keeping the newer run (on the fix commit 744d15a). This file will conflict on every feature → main PR; worth a follow-up (stop tracking latest.json, or a .gitattributes/workflow rule to auto-take one side).

After merge

Tag v3.3.1 on main for the versioned image build. (Prod runs the feature-answer-quality-recipes image, which already has the fix.)