[kvlonge]

Why

The planner read compiled audits, functions, seeds and sources by generic attribute lookup. That hides which native view each read depends on and blocks moving planning onto plain Rust records. Models already go through one counted native accessor. This applies the same treatment to the other resource kinds.

Changes

• audit_reads, function_reads, seed_reads and source_reads return each attribute the planner reads from that kind's native view, built exactly as its Python attribute is. Any other object is counted in the compiled_view_fallbacks stage counter and then read by attribute.
• resource_reads handles name, key, deps, destination, schema_entry and references for sites that read every resource kind alike.
• Planning code now reads through these helpers: audit entries and attachment, source deferral, relations, identity hashes, function, seed and scenario entries, warehouse checks, and command and clone plans. Reads selected by a name stored in a variable now use typed reader tables.
• The fallback integration test also plans the waffle shop fixture, so sources, seeds, functions and audits are proven to be read natively.

No behaviour, output, diagnostics or API changes.

Verification

• Compiler differential: 373/373. Planner differential: 11/11, after each commit.
• Plan output is byte-identical on two large synthetic projects. The fallback counter is 0 for compile and plan on those projects and on waffle shop.
• Runtime instruction counts (dense compile, config build, import, warm and cold sqb plan) are at parity, within ±0.2%.
• Build work: clean release +0.35%, incremental +0.22%.
• Rule budgets unchanged. Focused unit, integration and e2e tests pass, and so do clippy -D warnings and make check-ci.