[Lng999]

Pasting a compressed party token could allocate hundreds of megabytes on Rose's plugin bridge event loop. Decoding used unbounded zlib inflation and accepted payloads longer than the defined token layout. A local example expanded to about 200 MiB and was accepted as a token.

Bound token text before decoding, limit decompression output to one byte beyond the largest valid payload, and require a complete stream without trailing data. Accept only the exact v1/v2 payload sizes. Existing valid v1/v2 tokens, optional ROSE: prefixes, omitted base64 padding and acceptance of older tokens remain supported. Truncated streams were already rejected by zlib; the new checks additionally enforce complete consumption and exact payload lengths.

Validation:

• python -m unittest party.tests.test_token_limits: 9 pass; 4 fail against base 146f7986.
• Tests cover valid legacy/current tokens, oversized inflation, overlong text, wrong payload sizes, trailing/truncated streams and invalid input. The bounded-inflation regression uses a small fixture rather than allocating the large demonstration payload.
• Combined audit integration verification: 367 Python tests and 5 JavaScript tests pass. Ruff and git diff --check pass.