fix(kiro): retry malformed tool_call wrappers once

[SemonCat]

Summary

• add a guarded, one-shot repair path for malformed Kiro tool_call wrappers
• run the repair gate unconditionally for every successful streaming Kiro response — there is no feature toggle or provider option
• preserve normal streaming TTFT by opening the gate on the first meaningful valid SSE chunk
• retry only when a malformed wrapper is detected before visible output
• suppress the failed first attempt, append a corrective instruction, and retry at most once
• propagate retry HTTP errors such as 429 to existing account rotation/fallback logic
• add Node 18-compatible abort-signal handling, separate TTFT/stall timeouts, buffer limits, and cancellation cleanup

Root cause confirmed in production

The first implementation made the repair path opt-in. AWS live Kiro credentials did not set the option, so the real path was:

chatCore → getExecutor("kiro") → KiroExecutor.execute() → transformEventStreamToSSE()

The malformed wrapper was validated and returned as invalid_kiro_tool_call without entering the one-shot repair gate. This matched the live failure:

Invalid Kiro tool_call payload: missing nested MCP tool name at input.name

This PR now removes that option entirely. Successful streaming Kiro responses always enter the bounded one-shot repair gate. Non-streaming and unsuccessful HTTP responses remain unchanged.

Dependency

Depends on #2681. This is intentionally a stacked PR; until #2681 merges, the GitHub diff also contains its commits. The repair-only diff is:

git diff fa842ae..4bbba26

After #2681 merges, this branch can be rebased onto master so only the repair commits remain.

Behavior

The repair gate:

• releases normal text/valid wrapper streams on the first meaningful valid SSE chunk
• never treats keepalives or [DONE] as meaningful output
• classifies malformed output before opening the stream
• never leaks a failed first-attempt tool call
• performs exactly one retry
• returns kiro_tool_call_repair_retry_failed if the retry is also malformed
• does not invent or silently drop ambiguous nested tool names

Operational safety limits remain configurable; these are bounds, not enable/disable switches:

• KIRO_TOOL_CALL_REPAIR_BUFFER_MAX_BYTES
• KIRO_TOOL_CALL_REPAIR_TTFT_TIMEOUT_MS
• KIRO_TOOL_CALL_REPAIR_STALL_TIMEOUT_MS

Tests

• exported live path regression via getExecutor("kiro").execute() with empty provider-specific data
• malformed → repaired valid wrapper, with no first-attempt leakage
• malformed → malformed retry returns the repair failure code
• valid streaming opens before upstream completion and does not retry
• retry 429 remains a non-OK response for existing account rotation
• buffer cap, TTFT/stall timeout, abort, client cancellation, and ordering
• related Kiro/base suites: 9 files / 83 tests passed
• changed-file ESLint
• git diff --check
• npm run build

The branch was deployed to AWS internal 9Router before the toggle removal and passed direct Kiro text and real terminal tool-call smoke tests; post-deploy logs showed zero invalid_kiro_tool_call, kiro_tool_call_repair_retry_failed, or response.failed markers during the smoke window. The final toggle-removal commit does not change the already-active live behavior when no option is set; it removes the ability to bypass it.