Closes #1597.

Problem

Sub-agent LLM calls never recorded token usage, so every sub-agent's input/output tokens were invisible to netclaw stats.

This is a pre-existing gap, not a regression. git log -S for ISessionMetrics/UsageDetails/SessionMetrics scoped to SubAgentActor.cs returns zero commits across all history — the actor never had an ISessionMetrics dependency and silently discarded the ChatResponse.Usage it already receives from StreamingResponseReader. The recent observability PRs (#1428, #1468, #1472/#1499) only added logging and pruned dead OTel Activity objects; none ever touched token tracking.

Fix — record at the source

ISessionMetrics (the DailyStatsPublisher) is a process-wide singleton that the parent session and every sub-agent share. So the correct fix is to record each sub-agent LLM call once, in the child, exactly as the main session records its own calls:

• SubAgentActor — inject ISessionMetrics?; record response.Usage on every LlmResponseReceived (one per LLM turn: tool-call turns, retries, the forced-no-tools final turn, and repair turns); add cumulative input/output token totals to the completion summary log.
• SubAgentSpawner — inject the DI singleton and pass it through CreateProps. No Program.cs change: DI already registers the singleton and injects it into the optional parameter, matching SessionCatalogService/MemoryCurationWorkerService.

Why not also aggregate in the parent (the issue's suggestion B)?

Because the parent and child share the same singleton, re-recording propagated totals in the parent would double-count every sub-agent token in netclaw stats. Folding sub-agent tokens into the parent's UsageOutput would also corrupt the parent's context-window percentage — the sub-agent has its own context window. Recording once at the source is both correct and minimal.

Tests

Four regression tests, each verified to fail if the recording is removed (neutralizing the call made all four fail, incl. the summary log showing inputTokens=0, outputTokens=0):

• Records_token_usage_to_session_metrics_on_text_response — single-call billing.
• Records_token_usage_for_every_llm_call_across_the_turn_loop — sums usage across a tool-call turn + final-text turn (the crux: internal calls, not just the final output).
• Completion_summary_reports_cumulative_token_totals — summary log carries token totals.
• Spawned_sub_agent_bills_its_llm_calls_to_session_metrics — full SubAgentSpawner → CreateProps → actor wiring, so dropping the pass-through anywhere in that chain fails here.

Adds a UsageOverride hook to the sub-agent test FakeChatClient and extracts a shared RecordingSessionMetrics test helper.

Validation

• Full Netclaw.Actors.Tests: 2551 passed, 0 failed; Netclaw.Daemon builds clean.
• dotnet slopwatch analyze: 0 issues. Copyright headers: all present.

Follow-up

Filed #1599 to consolidate the three divergent FakeChatClient doubles into a shared Netclaw.Tests.Utilities helper (deliberately kept out of this PR to stay focused).