test: make stream capture tests reach their bounds and correct docs

[lofoneh]

Closes #31

Tests

• The five OpenAI chat recovery tests no longer set maxAttributeLength. Each one now prefills content to the real 48 KiB budget, so the 256-char value actually gets rejected and recovery happens right at the boundary. I checked each test by breaking the mechanism it's named for (separator refund, role field refund, recoverable id capture, id release, shell reserved once). The old tests passed all five breakages; the new ones fail each.
• Tests 3–5 sit 1–2 bytes under 48 KiB. That slack is the tool-call path's own conservative reservation; it's the same with no rejection at all, so it isn't a leak.
• The Anthropic fallback-tail fixture now rejects slices, so an iterations[::-1] scan fails.

Docs

• packages/sdk/README.md: an older core keeps the same fixed retention bounds, and only the mask is assumed. I also fixed the matching LEGACY_CAPTURE_POLICY comment in openai and anthropic. maxAttributeLength stays because the policy type requires it.
• TS openai README: transcription is 65,536 serialized UTF-16 units, not 64 KiB. The Python one really is UTF-8 bytes, so I left it.
• Python openai README: Responses capture uses 48 KiB and the core default of 1,024 items. TS Responses really is 1,000, so its README was already right.

[aikins01]

Thanks, the recovery tests now hit the real 48 KiB budget, and they fail when the separator, string, or field refund is removed, which the old versions didn't catch. The Anthropic fixture now catches a slice-based full scan too.

One sibling spot still needs the same correction: the root README (lines 119–120) gives one set of bounds for both SDKs. Python Responses uses 48 KiB and 1,024 items while TypeScript uses 1,000, and TypeScript transcription counts 65,536 UTF-16 units of serialized JSON while Python keeps 64 KiB of UTF-8. Could you split those there to match the package READMEs?

[lofoneh]

Good catch, fixed in bb4875d. The root README now splits Responses and transcription by language. Chat and Anthropic stay as one line since both SDKs use 48 KiB and 1,000 there.

[aikins01]

thanks for the quick fix, the README now matches both SDKs