Production blocker: prove complete cache input closure before ungating

Problem

The shared cache can restore zig-out and skip compilation, so a false cache hit is a correctness failure. The current key is useful as a prototype but does not yet prove that every artifact-affecting input participates in the key.

PR #43 therefore gates cache_build.sh behind AZAZEL_EXPERIMENTAL_CACHE=1 and removes the stale-hit impossibility claim.

Required coverage

The production cache key needs an explicit declared-input closure covering at least local Zig/C/C++ sources and headers, object files, generated tool inputs/outputs, local path dependencies, package identities and relevant package build arguments, target/cpu/ABI, optimization/build options, Azazel executor/generator version, toolchain identity, and external/system dependencies or a deliberate policy that makes them ineligible for shared caching.

Arbitrary pre/post commands and environment-dependent generators need either declared inputs/environment or a hard refusal from the production cache path.

Acceptance criteria

• Mutation tests show that changing every modeled artifact input changes the key.
• Target and build-option changes produce different keys.
• Local path dependency content changes produce different keys.
• Native source/header changes produce different keys.
• Builds with undeclared side effects/inputs are rejected or explicitly marked non-cacheable.
• Cache-hit artifacts are compared against clean-build artifacts across representative real-project parity slices.
• Only after those tests pass is the AZAZEL_EXPERIMENTAL_CACHE=1 gate removed.

Start with cache_build.sh and the changes described for PR #43, including the AZAZEL_EXPERIMENTAL_CACHE=1 gate. Map the declared cache inputs against the required coverage, then add mutation and clean-build parity tests. Done means modeled input changes alter the key, undeclared side effects are rejected or non-cacheable, and the gate is removed only after the acceptance criteria pass.