The worker now always dispatches through the in-process ocrs engine.
`AnalysisConfig::effective_backend()` returns Ocr regardless of the
parsed `ANALYSIS_BACKEND` (warning if `vision` was requested), and
`spawn_analysis_daemon` selects on it. The vision dispatcher, client,
and readiness probe stay compiled and intact behind the unreachable
match arm, so re-enabling is a one-line change.
The startup-reclaim regression test no longer leaned on the vision arm
to skip model loading; it now points the engine at a missing model path
and asserts the orphaned lease is still reclaimed before the (failing)
engine build — proving reclaim runs independently of engine readiness.
Also gitignore backend/models/*.rten so locally-downloaded ocrs models
for native `cargo run` aren't committed.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
de6eaf9f8b