[Ricardo-M-L]

Summary

Bound each polling HTTP request by the polling time remaining, while preserving a shorter configured request timeout.

The loop already bounds the sleep between requests, but an in-flight status request can outlive the entire remaining polling budget. For example, a 50 ms polling budget with a 1 s request timeout currently waits around 1 s when the server stalls. This is separate from the sleep fix in #283.

Changes

• Recompute the remaining budget before every polling request and pass the smaller of that budget and config.timeout.
• Convert fractional-second request budgets into integer milliseconds suitable for Node's AbortSignal.timeout, correcting floating-point noise around whole milliseconds and rounding genuine fractional milliseconds down.
• Preserve TimeoutError and AbortError while reading both successful and HTTP-error response bodies. Malformed JSON retains the existing error mapping.
• Add regression coverage for pending requests, subsequent polling iterations, shorter request timeouts, millisecond boundaries, and both body-reading paths.

This bounds the HTTP wait at the start of each polling iteration. It does not redesign OAuth credential refresh or claim a strict wall-clock deadline during credential resolution or event-loop blocking.

Validation

Base: 38aba5f26173dd1065b115b221d79c9ddd87c6bb.

• The initial pending-request and subsequent-budget regressions fail against the unmodified implementation.
• Targeted polling, HTTP, CLI video, and SDK video tests: 67 passed.
• Full suite: 636 passed, 1 failed; pristine base in the same environment: 613 passed, the same 1 failed. The unchanged failure is the macOS Zed XDG_CONFIG_HOME expectation in editor configuration tests. Browser launcher commands were stubbed in both runs to avoid opening a browser; tests were not skipped.
• bun run typecheck, bun run lint, and bun run build pass (lint retains one existing speech any warning).
• Built CLI Node smoke checks pass.
• Real Node HTTP server probes cover stalled headers, a partial 200 JSON body, and a partial 500 JSON body: the 50 ms budget aborts in roughly 52–55 ms. The original header/body cases take roughly 1 s. Body timeouts retain timeout exit code 5 rather than becoming a general JSON error.
• Two independent review rounds were performed as required by AGENTS.md; confirmed first-round findings were fixed, and both second-round reviewers reported no blocking findings. Both independently ran the 37 HTTP/polling tests and native Node timeout checks.
• git diff --check passes.

bun test test/polling/poll.test.ts test/client/http.test.ts test/commands/video/generate.test.ts test/sdk/video.test.ts
bun run typecheck
bun run lint
bun run build

Implemented and tested with OpenAI Codex assistance. Independent reviews above were agent reviews, not a claim of maintainer or human approval.

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}