Preview cache hygiene: stale-signature prefetch containers, same-key put(), warm flat re-decode, inflight-latch audit

[ben-dev-au]

Pre-existing edges in the preview/prefetch caching (now fnd/tui/preview/ + fnd/tui/widgets/preview_container.py), surfaced by review on #51–#54 and deferred to keep the decomposition PRs pure relocation:

1. Stale-query containers can be cached. PrefetchEngine._mount_chunk_loop bails when the query signature changes, but its finally block still put()s a partially-mounted container if any chunk landed. Keyed by the old signature it can never be served, but it occupies the 1-slot LRU and can evict the fresh container.
2. PreviewCache.put() does not surface a same-key replacement. Overwriting an existing (parent_doc_id, query_signature) entry returns the old container in neither evicted nor anywhere else, so the caller cannot remove the orphaned widget tree from the DOM.
3. Warm flat results are re-decoded. prefetch_top_results filters candidates by preview_cache only; a flat file already warmed into flat_buffer_cache/prebuilt_cache is re-targeted and re-decoded each pass (the mount is then skipped by the cache-key check, so this is wasted decode work, not a correctness bug).
4. Audit: inflight_target release on non-finalize completion paths. The dedup latch set by fire_pending_load() is released by _finalize_via_lock, on mount cancel (fix(tui): clear in-flight preview latch on mount cancel #42), and on query change. Review claims the flat and warm/instant structural completions may leave it set, suppressing an identical re-navigation. The area has dedicated tests (test_preview_load_debounce, test_preview_stuck_fast_nav lineage); verify with a characterization test before changing anything.