Make count-question auto-run authoritative and deterministic

[Robbie1977]

Fixes the medulla image-count question end to end. After v3.3.1 it regressed to "you can run a query" (no number); before that it had answered "28 images" (PartsOf, a class query). Both were the same two root causes.

Root causes

1. Guard defeated the injector. maybeInjectCountQueryStep skipped whenever the planner had already planned any run_query — so when the planner picked the wrong query (PartsOf), the injector stepped aside and the wrong query ran; and when the planner planned nothing concrete, the injector still didn't get to correct it.
2. Relied on the weak extractor. Even when the right query ran, the count had to be pulled out of the result table by the ELM extractor, which sometimes failed → the answer fell back to the digest's "run this query" text (the "no number" regression).

Fix

• Authoritative routing — the injector no longer skips on a generic run_query. If the planner planned a run_query with the wrong query_type, it retargets that step to the semantically-matched query (PartsOf → ImagesNeurons); otherwise it injects one. It still defers only to the specialised connectivity / scRNA-seq / region-neuron-count macros, and only acts on an unambiguous match.
• Deterministic count — the step is tagged count_query + count_noun + count_term; runStep reads count straight from the run_query result and adds a deterministic evidence row (VFB holds 226,524 images …), so the number and its correct unit reach the answer regardless of the extractor.

Tests

New offline end-to-end test (mocked ELM/MCP) with the user's exact question including the deliberate medualla typo: it runs ImagesNeurons (never PartsOf) and the evidence reads 226,524 images. Suites: orchestrator 23, orchestrator_run 23; whole offline unit suite green; next lint clean.

After merge

Tag v3.3.2. (Small code delta: lib/orchestrator.mjs + two tests.)

Separately — the task-battery CI gate failed PR #7 on a single transient T3.3 timeout (errors > 0 → exit 1); worth making that non-blocking / retry-once so live-server flakes don't red-flag good PRs. Happy to do that too.