Add delegate observability + metering (#39 slice 6)

[Jason-Vaughan]

What

The deferred non-local delegate metering, landed as orchestration-tree observability. Delegated sub-calls are now metered and broken down by backend — without corrupting the existing "spend avoided" headline.

• measurement — records gain a kind field (task default / delegate; older records read as task). rollup keeps delegate records out of the headline (no double-count — the parent task already credits the whole job) and aggregates them separately into summary["delegates"] = {count, by_backend: {model: {count, in/out tokens}}, in/out tokens, cloud_equiv_usd}. format_rollup adds a "Delegated sub-tasks" section (omitted when there are none). A process-level lock serializes the log append (delegate_many fans out across threads).
• delegate — run_delegate meters each sub-call as kind="delegate" at a single seam, so delegate_many inherits it for free. A metering error never breaks the delegation.
• GUI — the stats card renders the delegate breakdown.

Why

Slice 6 of the scatter-gather roadmap (#39). Until now, delegated work was invisible to measurement — you could see the top-level task but not which backends did the offloaded sub-work. This makes the fan-out observable (which backend → how much) while keeping the spend-avoided number honest (delegations are informational, never double-credited). It's also the data source that would tell us whether a delegate_reduce tool (slice 4) ever earns code.

Scope

The per-parent-task tree (linking each delegation to its specific top-level task across processes) is deferred — it needs a task-id propagated through the orchestrator CLIs to the MCP child, which can't be verified hermetically. This slice delivers the robust by-backend aggregate.

Test plan

• +13 hermetic tests: kind written (default/explicit); rollup excludes delegates from the headline (tasks/by_tier/tokens/spend_avoided unaffected) and into delegates; kindless record → task; by-backend aggregation; format_rollup shows/omits the section; 20-thread concurrent append → 20 well-formed lines (thread-safety); run_delegate meters kind="delegate"; a raising record_task doesn't break delegation; GUI view_stats carries delegates.
• +1 gated live metering check (temp TANGLEBRAIN_STATE_DIR).
• Test isolation verified: ran the full suite with no state-dir override — the real ~/.cache/tanglebrain/usage.jsonl was untouched (3 lines → 3 lines).
• Full suite green under clean-roster conditions: 380 pass / 14 skipped; make lint OK.
• Independent Critic review: SHIP (3 doc-parity fixes applied — stale "not metered" notes in docstrings + roster comment).

Part of #39 (slice 6; does not close the epic). Also promotes the pending slice-4 ### Internal entry into this release.

🤖 Generated with Claude Code