Harden embedding: atomic model switch, consistent vectors, downloads

[ppXD]

Summary

A stability review of the local notes-embedding feature surfaced several ways it could silently degrade. This hardens them:

• Atomic model switch. reindex_all now clears + rebuilds chunks in a single transaction — the DELETE FROM chunk previously committed before embedding, so a mid-embed failure wiped the index for good. New set_embedder_and_reindex reverts to the previous embedder on any failure, and apply_embedding_model surfaces the error instead of swallowing it. A failed switch now leaves the previous model + index exactly as they were, rather than a new embedder pointed at stale-dimension vectors or a silently-emptied index.
• Consistent vectors. Embed passages one at a time. A quantized model's per-batch scales made a batched passage drift from the same text embedded alone (verified cos 0.965), so a stored vector wouldn't match a single-text query and two notes indexed the same sentence differently. One-at-a-time keeps every vector identical regardless of its batch.
• Download integrity. Validate the body length against Content-Length, so a truncated transfer errors + removes its .part instead of renaming a short file into place (which then fails to load on every retry, since the present short file is skipped).
• Fail loudly on a zero-length tokenizer encoding instead of emitting an all-zero "embedding".

Test plan

• wisp-library (CI-gated): new regression test a_failed_model_switch_keeps_the_old_index_and_embedder proves a failing embedder neither wipes the corpus nor activates — the previous embedder + chunks survive. Full suite 38/38.
• wisp-embed (--ignored, real models): new qwen3_passage_embedding_is_independent_of_neighbours proves a passage embeds identically alone vs. alongside others on real Qwen3 (was cos 0.965 batched, now > 0.999); encoder (BGE-M3, GTE) + decoder ranking smoke tests still pass. 20 unit tests pass.
• cargo clippy -D warnings clean on wisp-embed, wisp-library, and the app; tauri dev rebuilds + runs.