feat(intelligence): wire canonical runs to the SDK — hooks, recordTrace, effort compiler

[drewstone]

Close the four small wiring gaps from the unification audit so a canonical runPersonified run observes and ships through the Intelligence SDK in one line. Reuse-first: no new primitive on either side.

The four gaps (file:line)

1. runPersonified subscribes to the Supervisor stream — RunPersonifiedOptions.hooks?: RuntimeHooks (src/runtime/personify/types.ts:245) forwarded into supervisorOpts.hooks (src/runtime/personify/persona.ts:152). The Supervisor already threads opts.hooks into the root Scope (supervisor.ts:130), which already emits agent.spawn/agent.child (scope.ts:~219,345). The load-bearing one-liner — no kernel change.

2. recordTrace(events) emits loop topology, not a flat span — IntelligenceClient.recordTrace (src/intelligence/index.ts) builds a nested loop → round → iteration OTLP span tree via the existing buildLoopOtelSpans into one traceId. No second span builder.

3. compileEffort makes EffortPolicy gate at runtime — compileEffort(EffortSettings) (src/intelligence/effort.ts) is a pure compiler to run-config overrides (withAnalyst / fanout / withLoops / intelligenceBudgetUsd). No if (effort) scattered into the supervise kernel.

4. OFF/eco degrade, not throw — gated at the EffortPolicy layer: analysts:false ⇒ withAnalyst:false ⇒ the caller omits the analyst, so the dormant empty-findings path runs and the base agent still works. The product fail-closed ("don't construct the analyst") stays separate from the experiment fail-closed inside createScopeAnalyst ("hard abort"), which is untouched.

How each reused an existing primitive

• Gap 1 → SupervisorOpts.hooks + the Scope agent.spawn/agent.child stream (already shipped).
• Gap 2 → buildLoopOtelSpans (the shipped topology builder) + createOtelExporter.exportSpan.
• Gap 3/4 → the existing EffortSettings shape; the compiler is pure data, the caller reads it once.

Tests (reused harnesses)

• tests/loops/personify.test.ts: runPersonified({hooks}) forwards agent.spawn/agent.child to a mock hook (and stays silent with no hooks).
• src/intelligence/intelligence.test.ts: recordTrace produces the loop/loop.round/loop.iteration tree under one traceId (asserting buildLoopOtelSpans, not a flat per-event builder); compileEffort maps off ⇒ no-analyst/fanout-1, eco ⇒ analyst-on/no-breadth, standard ⇒ analyst-on/breadth/loops.

npx tsc --noEmit + Biome clean; full pnpm test green (886 pass, 0 fail).

Note on base: branched from feat/analyst-steer-wire per the brief; targeting main because that branch's work is already merged there (#284). Merges cleanly into origin/main.

[tangletools]

✅ Auto-approved PR — 6b5e83ef

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-14T01:03:34Z}

[tangletools]

✅ Auto-approved PR — 9d2bc175

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-14T01:13:15Z}