fix(fts): drop and recreate code_fts to clear recurring SQLITE_CORRUPT_VTAB

migrations/0003_fts5_code_recreate.sql

@@ -0,0 +1,37 @@
+-- D1 FTS5 code_fts virtual table fresh recreate — recurring SQLITE_CORRUPT_VTAB
+--
+-- Layer = L4 Operations (sparse retrieval surface recovery)
+--
+-- Context:
+--   2026-04-24: migration 0002 applied FTS5 'rebuild' to both nat_fts and code_fts
+--   to recover from SQLITE_CORRUPT_VTAB.
+--   2026-04-28: corruption recurred on code_fts only (trigram tokenizer side).
+--   Enriched logs from PR #137 confirmed errorName=Error /
+--   D1_ERROR: database disk image is malformed: SQLITE_CORRUPT_VTAB on every diff
+--   upsert across all 5 polled repos.
+--
+-- Recovery (more aggressive than 0002):
+--   DROP the corrupted virtual table and recreate it from scratch with the
+--   same definition as 0001, then repopulate via FTS5 'rebuild' which reads
+--   from the content-owner table (search_docs).
+--
+--   Triggers from 0001 (trg_search_docs_ai/ad/au) reference search_docs_code_fts
+--   by name; they resume working as soon as the new table exists, so they do
+--   not need to be redefined.
+--
+-- Scope:
+--   Affects code_fts only. nat_fts is untouched (no recurring corruption observed there).
+--
+-- Idempotency:
+--   IF EXISTS / IF NOT EXISTS clauses keep the migration safe to re-run.
+
+DROP TABLE IF EXISTS search_docs_code_fts;
+
+CREATE VIRTUAL TABLE IF NOT EXISTS search_docs_code_fts USING fts5 (
+  content,
+  tokenize = 'trigram case_sensitive 0',
+  content = 'search_docs',
+  content_rowid = 'rowid'
+);
+
+INSERT INTO search_docs_code_fts(search_docs_code_fts) VALUES('rebuild');