feat(#24): cost-guard pre-check at outbound dispatch

[suzuke]

Closes #24.

Summary

When a target instance hits its daily cost limit, the cost-guard's limit handler pauses it. Before this PR, the rest of the fleet kept dispatching send_to_instance / delegate_task to the paused target — the sender got no signal until the no-reply timed them out, and the target accumulated a backlog it could not work on.

This PR surfaces the limit to the sender immediately, at dispatch time.

Changes

• OutboundContext gains costGuard: CostGuard | null. FleetManager already owns this field; the interface change just exposes it to handler code.
• sendToInstance short-circuits with an error when ctx.costGuard?.isLimited(target) is true. Error names the target and its daily limit so the caller can decide whether to back off or escalate.
• delegate_task inherits the gate via wrapAsSend (it routes through sendToInstance with request_kind: "task").
• report_result (request_kind: "report") bypasses the gate — terminal status updates must reach the orchestrator even when the target is paused, otherwise the merge gate stalls and impl can't escalate.
• Null-safe: fleet without cost guard → ctx.costGuard is null, optional chain skips the check.

LOC: src/outbound-handlers.ts +23 / -0; tests/outbound-cost-guard-precheck.test.ts +184.

Test plan (operator-mandated three states)

• npx tsc --noEmit — clean
• npx vitest run — 63 files / 524 tests pass (+1 file, +7 tests over baseline; no regressions)
• tests/outbound-cost-guard-precheck.test.ts:
  • (a) limited → handler responds with error containing cost-guard, target name, and limit dollar amount; target IPC receives 0 messages (sendToInstance + delegateTask both verified)
  • (b) not limited → normal dispatch; target IPC receives 1 message; respond called with success object
  • (c) ctx.costGuard === null → graceful dispatch identical to (b), no error path triggered
  • report-kind bypass → message with request_kind: "report" dispatches normally even when target is limited

Risks considered

• Existing cost-guard pause logic unchanged: this PR adds a pre-check at dispatch time. The existing costGuard.on("limit", …) pause flow (fleet-manager.ts:370) still fires when the target itself reports cost over the limit. Two complementary signals, neither replaces the other.
• request_information (kind: "query"): also routes through sendToInstance and IS gated by this change. Defensible — an over-budget instance shouldn't be answering questions either, and the sender gets a clear signal. If review disagrees, narrowing to only "task" + direct sends is a 1-line change.
• Bypass list: request_kind === "report" is the only bypass. update / query / task / unspecified all gate. Report bypass is justified because reports are terminal status (no follow-up cost).
• Race: a target can transition from "not limited" to "limited" between gate check and dispatch, but that's pre-existing race surface (any pause signal arrives mid-dispatch). The pre-check shrinks the window, doesn't eliminate it. No new race introduced.
• No fleet config migration: costGuard is already optional in fleet config, and ctx.costGuard is null when disabled.

Files

• src/outbound-handlers.ts — interface field + gate at top of sendToInstance
• tests/outbound-cost-guard-precheck.test.ts — new

🤖 Generated with Claude Code