fix(sqlite): backport bundled WAL-reset fix for Crab

[forhappy]

Cellule currently bundles SQLite 3.49.1, which is affected by the upstream WAL-reset corruption race. Upgrade the ordinary published dependency to rusqlite 0.39 / SQLite 3.51.3 and preserve range-checked unsigned SQL bindings with fallible_uint.

This is the minimal backport for Crab's exact pinned Cellule revision 6ac8a03. It targets the preserved codex/node-log-expiry-rotation branch because current main has moved to the 0.2 Node Bundle representation. Crab must not import that unrelated storage/API migration as part of a SQLite safety fix. The same upgrade is forward-ported to current main in #79. No merge requested.

Contract adaptation

• Replace removed DatabaseName::Main with upstream's MAIN_DB at incremental blob I/O and read-only-root validation.
• Propagate fallible hook installation before any application SQL. rusqlite 0.39 rejects hooks on borrowed handles; ignoring that result would execute SQL without the authorizer. Preserve standard typed error sources and scoped cleanup on every exit.
• Add a borrowed-handle regression at the public SQL-query boundary. Existing protected-table, trigger/view and post-error cleanup coverage remains.
• Keep managed capture, checkpoint and verified recovery policies unchanged. No patch override, vendored source or configuration surface.
• Lockfile changes are limited to rusqlite/libsqlite3-sys, its hashlink/hashbrown/foldhash closure, and wasm-target dependencies introduced by upstream. Native builds still link the bundled fixed SQLite.

Evidence and limits

• Read the upstream release, Cargo features and downloaded SQLite source. The linked amalgamation identifies SQLite 3.51.3 with source ID 737ae4a34738ffa0c3ff7f9bb18df914dd1cad163f28fd6b6e114a344fe6d618.
• The borrowed-authorizer regression fails before the adapter correction and passes afterward.
• All seven SQL integration tests pass, including real mutation, publication and exact restore. The complete 44-test primitive integration target also passes.
• LTX replica qualification passes 240 executable tests and one doctest, with six opt-in RustFS/provider/filesystem cases ignored. The local-only feature separately passes 35 library tests, ten integration tests and its doctest. Strict LTX/runtime all-target/all-feature Clippy passes. Full workspace, cross-platform and provider qualification remain CI/live gates; this PR stays draft.
• Four separate real RustFS tests pass: native/bundled root range compaction and exact reconstruction; fragmented demand reads; origin-admission release before optional cache sync; and concurrent sparse activation. Each uses a fresh scoped test prefix. Receipts: cellule-sqlite-rustfs-compaction1.json and cellule-sqlite-rustfs-extra-receipt.json. The six opt-in cases above were skipped by the default suite; these runs explicitly qualify four. Two manual provider-cost/filesystem diagnostics remain unrun.
• Architecture boundaries, module ownership, 104 Rust documentation snippets and 1,082 local documentation links pass. A missing local PATH entry was corrected before rerunning these checks.

Review evidence map

Boundary	Evidence
Dependency owner	Workspace rusqlite shared by cellule-ltx and cellule-runtime
Callers	Managed DB capture/recovery; public sql_batch/sql_query_batch via execute_batch; incremental blob write
Callees	rusqlite MAIN_DB, checked hook registration and SQLite WAL/checkpoint implementation
Siblings	SQL read/write batches share AuthorizerGuard; existing primitive tests cover protected access and cleanup
Pinned baseline	Exact revision used by Crab; only the dependency closure and required API adaptations change
Current main	Separate 0.2 Node Bundle lineage; same dependency fix prepared independently

Is this the best fix? Use the maintained upstream SQLite fix and adapt its explicit API contracts. Global cache serialization or private SQLite patches would add ownership complexity and constrain throughput without delivering the same dependency correction.

Primary contracts: SQLite WAL-reset fix, rusqlite 0.39 release, feature authority.