fix: handle remaining stream and tool-call edge cases after #3

[yibudak]

Summary

Follow-up to #3: this addresses reproducible edge cases that remain on main at a97eb903932f2253d8ebeca2bcfabdbb47908889, which already includes that PR. It does not replace the earlier streaming/backpressure work or change authentication, model routing, deployment configuration, or supported API routes.

Changes

• Downstream cancellation: observe ServerResponse.close rather than the already-completed request upload. Cancel upstream reads for disconnected SSE and JSON clients, settle backpressured writes on close/error, and avoid retrying an already-aborted caller.
• Native stream idle timeout: ReadableStreamDefaultReader.cancel() resolves a pending read as EOF; it does not reject that read. Explicitly destroy the Node stream with the idle error so the encoder's existing error/terminal-record path runs instead of an indistinguishable successful EOF. Normal successful streams retain their existing behavior.
• Bounded HTTP error bodies: keep non-2xx body consumption within the per-attempt deadline, cap it at 16 KiB, cancel unread data, and redact the supplied upstream key from HTTP diagnostics. Do not apply the header deadline to an ongoing successful generation.
• Tool-call reconstruction: avoid appending canonical arguments after the same JSON was already streamed. Emit complete OpenAI ID/name metadata only once per index; preserve final-only calls and interleaved Anthropic tool blocks. Reconcile insignificant whitespace in partial JSON prefixes without changing emitted bytes, and reject conflicting values or whitespace that splits JSON tokens.
• Non-streaming failures: do not convert an upstream error event into an empty successful response.
• Token-count input validation: return a protocol-shaped 400 for null/primitive/malformed bodies instead of a 500 or silently accepting invalid top-level input. The token count remains the existing character-based estimate.

Reproduction examples on unmodified main

• POST /v1/messages/count_tokens with JSON null returns 500.
• Canceling a real local HTTP response client leaves the synthetic upstream signal/read uncanceled.
• A tool-call-delta with {"x":1} followed by the matching canonical tool-call produces concatenated arguments {"x":1}{"x":1} in both protocols.
• A native upstream stream that stalls is treated as clean EOF after cancellation.
• A 400 response whose body never finishes outlives the configured header deadline.
• A non-streaming upstream error event produces an empty success response.

The automated reproductions use synthetic upstream events/credentials, native Web streams, and real loopback HTTP connections. They demonstrate specific edge cases, not that every model/request encounters these failures. In particular, this is not a claim to fix the provider-side model availability problem discussed in #7.

Verification

• Node.js 24 TypeScript build passed.
• Full suite: 265 passed, 0 failed across 17 test files.
• Regression tests cover disconnects, normal completion, backpressure, native idle cancellation, stalled/oversized error bodies, both tool protocols, partial whitespace/escapes, and invalid input.
• Compared the current unmodified upstream build with the patched build using the same local reproductions.
• Separate authorized live smoke checks passed for OpenAI JSON, OpenAI SSE, Anthropic JSON, and a streamed tool-call/tool-result round trip using deepseek/deepseek-v4-flash.

No real credentials, host-specific paths, private network configuration, or generated build artifacts are included. Happy to split the validation change or other independent pieces if smaller follow-up PRs would be preferred.