docs(review): consolidate repository-specific review checkpoints

Problem

The repository-specific code-review guide has accumulated overlapping and incident-specific rules. It also tells the reviewer both to comment on minor style preferences and not to comment on optional improvements. This makes material SDK risks harder to distinguish from review noise.

A corpus review of automated approvals and later human feedback shows recurring misses at public/persisted compatibility boundaries, cross-layer API propagation, resource lifecycle, secret handling, and packaged runtime behavior. False-positive comments are concentrated in optional refactors, style, documentation, and ungrounded test requests.

Desired Behavior

The guide should organize stable SDK review boundaries into a concise decision procedure, and AGENTS.md should give contributors the matching implementation checks needed to pass that review.

Acceptance Criteria

• Organize the guide around a small set of stable, repository-specific blocking checkpoints.
• Preserve the eval, release, dependency-freshness, new-model preflight, and runnable-example policies.
• Require findings to identify a concrete failure on the current PR head.
• Remove contradictory advice and incident-specific implementation recipes that duplicate AGENTS.md.
• Keep links to the authoritative deprecation and package-specific policies.
• Add matching implementation guidance to AGENTS.md so contributors can satisfy the review checkpoints before opening a PR.

Start by reading the repository-specific code-review guide and AGENTS.md, then compare their overlapping checkpoints and preserved policies. Consolidate the guide around stable blocking boundaries, add matching contributor checks to AGENTS.md, and verify that authoritative links and the listed eval, release, dependency, preflight, and example policies remain intact.