[AnderEnder]

Concurrent cache writes could mix bodies with validators, malformed HTTP 200 responses could evict a usable feed, and large durations could panic. Publish metadata and body in one atomic record, preserve valid cached feeds and unrelated files, and use checked duration arithmetic.

Enforce a 16 MiB decoded-body limit during HTTP reads (including gzip and chunked responses), cache reads/writes, and direct parsing. Existing cache pairs remain readable; older binaries cannot read newly written atomic records.

Validation: formatting, clippy with warnings denied, the full test suite (292 tests), and release build. Regressions cover concurrent writers, malformed replacement, legacy cache upgrades, cache clearing, oversized bodies, and duration overflow.

Stack position: 1 of 4; targets main. Merge the stack in order. Consolidated documentation is in the final PR.

Merge order: #53 → #54 → #55 → #56.