…st stopped one word short

`.out` was in LOG_EXTS and `.output` was not. Extension matching is exact, so `.output` never
matched `out` and every build or test run redirected to one was read whole — 54 recorded reads,
9,547 tokens at or above the threshold.

The list lives in three places (lumen_core::coverage::LOG_EXTS, INTERCEPT_TEMPLATE, and the repo's
developer hook) and the word was missing from all of them, because nothing compared the shell copies
to the constant. The new drift test drives its assertion from the constant, so the next addition
cannot land in one copy only.

One trap while writing it: `out|` is a substring of `output|`, so a `contains` check would report
`.out` as routed even if only `.output` were present. The arm is split on its own delimiters instead.

Tests, per the rule that every fix carries all four:
  unit  - classify() for .output, the whole LOG_EXTS set, and .outputs staying out
  drift - both intercept copies route every extension in the constant
  e2e   - the generated hook run for real against a 400-line .output: exits 2 and names compress_logs
  bench - a new scenario, 13,842 -> 1,062 tokens (92.3% saved)

No existing benchmark scenario moved: the baseline diff is nine added lines and nothing else.