fix(knowledge): correct Cosmos3 PP cache failure and CI evidence - #446

Summary

This PR changes only the existing Cosmos3 knowledge Markdown pages.

• Explain the stale-cache failure correctly: a PP-missing module is mistaken for an owned layer, producing a false missing-required-weight error.
• Replace the outdated AMD status with the later, explicitly attributed maintainer report: the target shard passed in AMD build #14506 while unrelated Wan tests failed overall. Keep the limit against claiming real multi-stage PP parity.
• Update the existing owner navigation and page dates.

The ownership contract, stable rule ID, and lifecycle footer are unchanged. There are no test, code, configuration, or workflow changes in the PR diff.

Evidence

• Pinned upstream implementation
• Base-fail/head-pass reproduction and false missing-weight error
• Later current-head validation and AMD target-shard result

The AMD result is a maintainer report, not a rerun performed for this change.

Validation

Base: 8b27af27252f4133b64bd2dcab1faea3969fbce2.

• Checked the corrected explanation against the pinned upstream implementation and both evidence comments.
• Scoped repository-validator checks for the complete Cosmos3 owner directory passed: structure, relative-link destinations, index uniqueness, size limits, privacy/dangerous-command checks, and frontmatter/taxonomy.
• Existing size advisory remains: rules.md is above 16 KiB and below the 32 KiB error threshold.
• Verified the PR diff contains exactly two files, both existing knowledge .md pages.

Local scoped validation used Python 3.13.5 and PyYAML 6.0.2. No model or GPU execution was performed.

Review and remaining checks

An independent review session found no substantive defects at head 99d17a277810334d9e9999684ad4794f6a81b898 and confirmed the Markdown-only scope. Both Linux suites, both Windows checks, and the release-drift audit passed on that head. The local workspace was a sparse source snapshot, so whole-tree validators, vLLM-Omni release-drift enforcement, and the aggregate suite were not run locally. The correction does not establish runtime retrieval, GPU correctness, or deployment activation. No deployment or knowledge-maintenance feature is enabled by this change.