Fix current authorization of preloaded generated relationship targets

[Dastari]

Generated nullable relationships returned preloaded objects before current target authorization or complete ownership-key resolution. A forged snapshot could therefore expose a denied or cross-tenant target. Resolve the authoritative target through the existing typed query/uncached batch loader, recheck current entity/row authority, and preflight selected target fields with their actual child contexts. Denial preserves an authorized parent, nulls its nullable child and attaches a sanitized relationship-path error. Missing/mismatched targets ordinarily return null without errors; SQL-complete visibility treats hidden rows as absent without an unauthorized existence probe.

Remove generated parents' pool-only eager preloads, which could query a denied target and fail the whole parent before nullable traversal. Keep explicit preload APIs source compatible. Apply current SQL visibility before relation paging/counting and include it in batch identity alongside DbAuthContext. Callback-only/prefiltered to-many policies require the existing host AuthorizedScanConfig: examine bounded database batches to exhaustion, retain only the requested visible window, and fail closed on an exhausted budget rather than returning partial counts or false end-of-data. Residual scans run per parent; SQL-complete loads remain batched. Existing SDL and cursor formats are unchanged. Policy callbacks may run during preflight and ordinary resolution; both receive the actual child context/arguments/alias/path. No cross-statement snapshot isolation under concurrent external DML is promised.

This is a separately reviewable dependent PR based on #96 at 4617b332014758387a3c271a7b30a5aff4873453 for the two-crate regression fixture. #96's narrow cross-crate fix remains unchanged. It does not depend on PR A or B–D, add joined/computed queries/group continuation/private view adapters, or change private repository roots. ORM/macros are aligned at 0.33.5. Version order: #96 (0.33.4), this fix (0.33.5), #97 (0.34.0); realign/rebase the pending patches if the owner chooses a different merge order. No release publication is authorized or performed.

Runnable SQL-free host example:

cargo run --manifest-path crates/graphql-orm-macros/fixtures/cross-crate-relations/Cargo.toml --locked -p cross-crate-source-models --no-default-features --features sqlite --example authorized_links

The example populates a forged cached snapshot, resolves the current target through a composite tenant/identity binding, then changes host target authority and demonstrates parent-preserving nullable denial with an alias/fragment. Public declarations remain the existing GraphQLEntity/GraphQLRelations and Database policy APIs. New doc-hidden runtime helpers support generated code; consumers do not call ORM-private methods or supply query SQL.

Executed verification (jobs=2, line-tables-only debug, incremental=0, disk-backed TMPDIR/target caches):

• External fixture cargo test --manifest-path crates/graphql-orm-macros/fixtures/cross-crate-relations/Cargo.toml --locked -p cross-crate-source-models --no-default-features --features sqlite: 6 tests passed. The separate --features postgres command also has 6 passed, using only labelled disposable test-owned containers with ownership-checked cleanup. One generated illustrative search doctest is ignored in each lane; no backend execution tests are ignored.
• Regressions cover entity denial before target I/O, a generated parent with an absent target table, forged/stale/manual preloads, a primed application HashMapCache, complete ownership-key mismatch/reassignment, reused schema/loader after permission changes, current/expired and two differently authorized identities, row denial, selected field denial, sanitized provider failures, aliases/fragments, skip/include selections, SQL-complete and bounded callback pages/counts, and budget exhaustion. The retained SQLite native tracing fixture measures two actual target SELECTs across multiple parents, one dispatch per source relation.
• cargo test -p graphql-orm --locked --no-default-features --features sqlite --lib --test authorization_mode --test composite_relations --test conditional_relations --test recursive_relations --test computed_order_and_complex_relations --test bidirectional_join_relations --test authorized_pagination: 54 passed, one pre-existing large release-mode pagination benchmark ignored.
• cargo test -p graphql-orm-macros --locked --no-default-features --features sqlite --lib: 6 passed.
• Warnings-denied Clippy for ORM/macros libraries and the external fixture's all-targets lanes, separately selecting sqlite, postgres, mssql: passed. MSSQL is compile-only, no live MSSQL execution claimed.
• Warnings-denied Rustdoc for ORM/macros, separately selecting those same three backends: passed. Root and external-fixture formatting, documentation checks (191 files), generated inventory, explicit backend dependency trees and workspace dependency architecture: passed.
• bash scripts/check-package-release-policy.sh 4617b332014758387a3c271a7b30a5aff4873453 and bash scripts/check-semver.sh 4617b332014758387a3c271a7b30a5aff4873453: passed. Never workspace all-features.

Known limits: callbacks requiring exact complete counts must finish within the host scan budget; SQL-complete policies are the bounded batched path. Explicit bulk snapshots are no longer reused for GraphQL authority. Other application loader cache types are not authoritative and trigger fresh-query fallback. Scope guards remain host-owned; no credentials, transport authority, policy language, sibling changes, data migration or releases.

• Documentation updated

Documentation impact: current generated relationship authority contract, migration/compatibility notes, changelog, SQL-free executable example and generated package inventory. Additional static query/group/private-view requirements remain independent workstreams in #91.

Additional explicit compatibility evidence: cargo semver-checks -p graphql-orm --manifest-path crates/graphql-orm/Cargo.toml --baseline-root <owned-4617b33-baseline>/crates/graphql-orm --default-features --release-type patch actually builds/parses 0.33.4 and 0.33.5 and passes 223 checks (30 inapplicable checks skipped). Issue #100 separately tracks no-analysis invocation behavior without explicit package selection. The reviewable implementation head remains unchanged.