fix(subagent): persist + reconcile tool-result turns in the gateway loop

[brettdavies]

Closes #2273

Summary

The bug. Subagent jobs that take the gateway-loop path (agent.use_gateway_loop = true, used by every non-Anthropic recipe) die deterministically after any mid-conversation interruption with [chat(litellm:gpt-X.Y)] Tool results are missing for tool calls call_<id>, .... The failure scaled with parallel-tool-call count, which read as a provider-bridge bug but was not. The error came from the next chat() dispatch operating on a stale in-memory message array, not from the wire response of the prior call.

The fix. gateway.toolLoop gains an onToolResultTurn callback fired when the tool-result user turn is built. runSubagentViaGateway wires the callback to persist the turn before the next dispatch AND reconciles on replay: when prior messages end with unanswered tool-calls, it re-synthesizes the tool-result turn from the settled subagent_tool_executions rows. Bails if any tool is still unsettled, so a non-idempotent tool is never re-run.

This mirrors the legacy Anthropic-direct path's existing reconciliation behavior; the gateway path was the only dispatch shape that lacked it.

Diagnosis

• gateway.toolLoop (pre-patch) built the tool-result user turn in memory after the tool executions settled, then discarded the index. runSubagentViaGateway had no persistence hook, so the turn never reached subagent_messages.
• On any interruption (worker recycle, rate-lease renewal, RSS watchdog, abort, clean shutdown), replay rebuilt the conversation from subagent_messages without the tool-result row. The shape was unbalanced: assistant tool_call turns with no matching tool user turns.
• The next chat() dispatch hit the provider's validator with the unbalanced shape; provider returned the "Tool results are missing for tool calls ..." error; worker treated as hard failure, retried the same unbalanced shape, dead-lettered the job.
• The legacy Anthropic-direct path in src/core/minions/handlers/subagent.ts already persists the user-turn reconciliation; the gateway path was simply never extended.

Reproduction

1. Brain configured for any non-Anthropic chat provider (litellm-proxy, codex-proxy, ollama, openrouter, openai). Set agent.use_gateway_loop = true.
2. Submit a subagent job whose dispatch fires parallel tool calls (gbrain dream --phase patterns is the deterministic local trigger).
3. Interrupt mid-conversation: kill -TERM the worker, or let the RSS watchdog recycle the process.
4. The worker re-claims the job and replays. Replay rebuilds the conversation from persisted rows.
5. Next chat() throws Tool results are missing for tool calls call_<id>, .... Job dead-letters.

Observed on gbrain 0.42.43.0 through 0.42.51.0.

Tests

• New test/e2e/subagent-gateway-toolresult-replay.test.ts: captures the real transport input shape and asserts the resumed conversation is balanced. Fails on master without the patch; passes with it. ~5s runtime.
• Verified end-to-end on a production-shaped brain (litellm + codex-proxy + gpt-5.5). Pre-patch: every interruption dead-lettered the active jobs; post-patch: every interruption resumed cleanly across worker recycles, RSS-watchdog kills, and clean restarts.
• bun run typecheck clean. bun run verify green.

Adjacent observation (not in this PR)

agent.use_gateway_loop is missing from KNOWN_CONFIG_KEYS (operators have to discover the --force escape hatch from the error message). And with the gateway path's recipe coverage now broad (every non-Anthropic provider plus Anthropic itself via native-anthropic), the default could flip from false to true. Both worth follow-up discussion; not in scope here.