Phase 2: detection engine v2 — archive/text/audio/video similarity + near-dupe correctness

[zackkitzmiller]

Why

Detection was shallow: near-duplicates covered images only via an O(n²) sweep, hamming_distance silently computed wrong distances for mismatched hash lengths in release builds, the required zip↔extracted-folder matching didn't exist, and text/audio/video similarity was absent. Archive introspection also flattened entries by filename, so a/index.html and b/index.html clobbered each other.

Stacked on #15.

Changes

• Archive ≡ folder (fingerprint/archive.rs): every zip entry is streamed through blake3 (no extraction) and matched against scanned files at the same relative paths under a common root. A full match emits DuplicateType::ArchiveMatch with the archive as the deletable side — the extracted tree is never the default deletion.
• Text similarity (fingerprint/text.rs): 64-bit simhash over word 3-shingles, threshold 6 — catches lightly-edited copies (contract v1 vs v2). Byte-identical pairs are left to exact dedupe; tiny/huge files skipped.
• Audio (fingerprint/audio.rs): chromaprint via fpcalc when on PATH — same recording across mp3/flac/bitrates at ≥90% bit-similarity. One log line + no-op otherwise.
• Video (fingerprint/video_frame.rs): keyframe phash via ffmpeg when on PATH feeds videos into the existing perceptual pool — video↔video and video↔image matches for free. No-op otherwise.
• Correctness: hamming_distance returns None on length mismatch (was debug_assert + wrong result in release).
• Scale: pigeonhole banding (threshold+1 bands) generates near-dupe candidates from shared band buckets — provably identical results to the pairwise sweep, near-linear time.
• extract_zip preserves entry relative paths under the temp dir.
• TUI: dupe mode renders MATCH TYPE / WHY panels for the three new kinds; --dupes-only runs all detectors.

Test plan

• cargo test — 366 passed (new: zip↔folder positive/negative/missing-entry fixtures, simhash edited-copy + unrelated-doc + identical-skip, banding-vs-bruteforce, mismatched-length regression, fpcalc similarity scoring, garbage-video none)
• cargo clippy --all-targets — clean
• Degradation: audio/video detectors are runtime-gated on fpcalc/ffmpeg presence with a single informational log