Restore state-backed tool call IDs

[brandonkachen]

Summary

• Restore deterministic tool call IDs using the functions.<tool>.<index> shape with a run/session-backed monotonic counter.
• Seed the counter from retained history and pending calls, including the older colon-delimited deterministic ID shape.
• Share the counter through spawned agent state while keeping the public agent template types free of internal allocator state.

Validation

• bun test packages/agent-runtime/src/util/__tests__/tool-call-id.test.ts packages/agent-runtime/src/__tests__/run-programmatic-step.test.ts packages/agent-runtime/src/__tests__/tool-validation-error.test.ts
• bun run --cwd packages/agent-runtime typecheck
• bun run --cwd common typecheck
• bun run --cwd packages/agent-runtime test
• NODE_ENV=production bun test cli/src/utils/__tests__/sdk-event-handlers.test.ts cli/src/utils/__tests__/message-block-helpers.test.ts cli/src/utils/__tests__/send-message-helpers.test.ts
• bun run --cwd cli typecheck
• bun run --cwd cli src/index.tsx --help

The core of this PR — reintroducing functions.<tool>.<index> deterministic IDs backed by a monotonic counter in AgentState.toolCallState, seeded from history (including the legacy colon-delimited shape) and pending calls — is well done. tool-call-id.ts is clean, createToolCallIdGenerator/ensureToolCallState correctly thread the counter through spawned agents, and the new test file (tool-call-id.test.ts) plus updates to run-programmatic-step.test.ts and tool-validation-error.test.ts cover the important cases (history seeding, hidden/legacy IDs, shared counter across spawned agents).

What's concerning is that packages/agent-runtime/src/tools/tool-executor.ts also contains a second, unrelated change to agent-spawn validation that has nothing to do with tool call IDs:

• agentIdToLoad is now normalized via normalizeAgentIdForLookup for base agents, and the returned agent object is rewritten with agent_type: agentIdToLoad.
• The effectiveInput = { ...effectiveInput, agents: validAgents } line was moved outside the if (errors.length > 0) block, so it now runs unconditionally instead of only when some agents failed to spawn.

Neither of these is mentioned in the PR description, and there's no test covering the new unconditional overwrite of effectiveInput.agents. That's a real behavior change (previously a successful spawn left effectiveInput untouched; now it always gets rebuilt from validAgents), and mixing it into a PR titled 'Restore state-backed tool call IDs' makes it easy for a reviewer to miss and hard to isolate for porting.

Please split the spawn-agent normalization change into its own PR with its own tests and description so each change can be evaluated (and reverted, if needed) independently. The tool-call-id work itself looks portable once separated.