fix: correct column index for corrupt embedding BLOB errors in search

[randomm]

Automated by pi-ensemble /work driver (mechanized commit-pr).
Fixes #186

Assumptions made while resolving this issue

The issue did not fully specify the following. Each was resolved with a defensible default rather than blocking — check them:

• D3 (constant) should be done — it's 4 one-token edits and directly addresses the "cannot drift again" goal; the issue marks it optional, but it is trivial and low-risk. If the developer judges it complicates things, skip it and note in the PR.

[randomm]

Six-pass review — findings still open at the round cap

The lens review ran its 3-round fix loop and stopped with these
outstanding. None is CRITICAL and the adversarial gate passed the diff, so the
cycle carried on to CI instead of parking — but nothing listed here has been
fixed. Decide them before merging.

Review findings — 5 MEDIUM, 8 LOW (verdict: ISSUES_FOUND)

• MEDIUM TYPE_SAFETY — src/sqlite/list.rs:245 — get_many inline row mapper drops MismatchedDimensions check (type-shape asymmetry)
• MEDIUM TYPE_SAFETY — src/sqlite/mod.rs:130 — Fromrusqlite::Error downcasts any domain Error variant, broader than documented
• MEDIUM SECURITY — src/sqlite/mod.rs:131 — Error message embeds unsanitized memory id
• MEDIUM ARCHITECTURE — src/sqlite/mod.rs:135 — From<rusqlite::Error> downcasts any domain error, broader than documented
• MEDIUM ARCHITECTURE — src/sqlite/query_mod.rs:20 — Type erasure of MismatchedDimensions into opaque reason string
• LOW SIMPLICITY — src/sqlite/fts.rs:227 — id read from column 0 three times in FTS row closure
• LOW SIMPLICITY — src/sqlite/list.rs:255 — tuple element and struct field both re-read row.get(0) in get_many
• LOW TYPE_SAFETY — src/sqlite/list.rs:257 — id read from column 0 three times; local not reused
• LOW ERROR_HANDLING — src/sqlite/mod.rs:131 — Overbroad downcast: any sqlite::Error source is surfaced, not just CorruptEmbedding
• LOW SIMPLICITY — src/sqlite/mod.rs:393 — Unnecessary id.clone() when id is not consumed after map_err
• LOW ARCHITECTURE — src/sqlite/mod.rs:394 — EMBEDDING_COLUMN constant hardcodes wrong diagnostic column in 3-col projection
• LOW SECURITY — src/sqlite/query_mod.rs:38 — Untrusted BLOB length feeds heap allocation on decode path

…and 1 more. The full set is on the last lens-issues-found event in the state file.

[randomm]

Retitled: the automated PR-creation step truncated the title at 72 chars (cut 'embedding' to 'embeddin'). Corrected to a complete, <72-char subject.
This matters because we squash-merge — the PR title becomes the commit subject, changelog entry, and release-note line. See #155 for the related changelog fragility (v0.9.1's '...per-row cosine loop in Dat').