fix: bound streamed tool argument memory

[Ingwannu]

Summary

• cap shared SSE records at 4 MiB, including unterminated and multiline records
• cap retained tool arguments at 8 MiB per call and 32 MiB per turn across OpenAI Chat, the Responses bridge, and Chat outbound conversion
• fail oversized calls as typed upstream errors without emitting a completed tool item or clean Chat DONE marker
• document the limits and decision tradeoffs

Why

Compatible upstreams can omit SSE delimiters or stream unbounded tool argument fragments. Those buffers were protocol-required, but they had no byte ownership limit and could grow for the lifetime of a turn. Truncating and completing a call would produce invalid or unsafe executable input, so overflow now fails truthfully while preserving the existing sequential tool event contract.

Validation

• bun test on 15 related bridge, Chat, OpenAI Chat, and SSE suites: 206 pass
• bun run typecheck
• bun run privacy:scan

Related to #820. Focused alternative to the overlapping portions of #829.

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: d177d411-a8aa-4f2c-a084-12fd4e99617c

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

_{Comment @coderabbitai help to get the list of available commands.}

[Ingwannu]

CI follow-up: the first Windows run passed 6,514 tests and timed out only the unrelated existing image-normalization P2 case at 5.39 seconds against its 5-second test limit. Re-running the failed Windows job passed. Ubuntu, macOS, Windows, package-install, target-enforcement, label, and React checks are now green.

[lidge-jun]

Thanks for this — it is a valuable part of the #820 series, especially because it bounds streamed tool-argument memory across the major translation paths. It currently conflicts with dev, so it will need a rebase, followed by an independent review from someone other than the author, before it can land.\n\nFor the broader draft series (#840, #841, #843, #844, #845, and #847), is there a specific landing order we should follow, or would it be better to consolidate any of the overlapping work? We are tracking #820 and would like to get these into the next cycle.