perf(context): cap CONTEXT_MAX_CANDIDATES 5 → 3

Each extra candidate costs ~35 µs of searchContent + findAllSymbols
work for diminishing return. The by_file ranking already heavily
favors the first 1–2 high-quality identifiers extracted from the
task. Drop the cap to 3 — keeps the lead identifiers, sheds the
tail.

Marginal improvement on the bench task (which only has 4 distinct
extracted identifiers) but bigger savings on tasks that mention
many symbols.