docs: merge the #417 and #392 subagent-steering designs into one

[justrach]

Closes the design half of #417 and #392. No behaviour change; this is the reconciliation both issues were blocked on.

The two issues describe one mechanism from two directions: #417 from addressability (durable child registry, children reachable by name, messages that steer a running child), #392 from resumption (a task id from spawn, follow-ups that continue the same worker, a queue-vs-trigger split). Neither is implementable without most of the other, and built separately they would produce two id schemes, two registries and two message paths.

Three findings from grounding the merge against the source

#392's central premise is false. It argues resume is "plumbing, not new state" because the worker transcript already persists under .graff/subagents/. It does not. cards.zig:writeSubagentDetail writes a final report plus metadata as markdown: no message history, no tool-call record, no provider state. Resuming a worker "with its accumulated context" therefore requires persisting history that is currently discarded when the pool thread finishes. #392 is materially larger than its own estimate.

Identity half-exists, twice. Background spawns already carry a stable id and a mutex-guarded session-global registry (AgentJob / g_agent_jobs), but in memory only, so it dies with the process. Separately, cards.zig:subagentId produces sa-<ordinal>-<hash> for inspect links. Two schemes naming one concept; unifying them is a prerequisite, or the registry keys on one and the persisted artifacts on the other.

The arc collides with #441. That issue excludes subagents from the append-only transcript on the grounds their history is never persisted by design, while #392/#417 require exactly that history. Both defaults are right for their own case, so the resolution is opt-in retention per spawn (retained: true): #441 keeps its default and gains one documented exception, unretained workers keep today's behaviour byte for byte, and the model gets a real cost signal at the moment it knows whether a follow-up is likely.

What the merged design settles

• Identity: one scheme, sa-<ordinal>-<hash>, already stable across runs and already in persisted artifacts. The background u32 demotes to an internal index. Retained, addressable subagents with parent-child steer messaging #417's names become aliases, never identity, preserving its own harness-assigned-sender rule.
• Steering: one delivery call with a trigger flag rather than Codex's two tools, so the expensive case is visibly the exception. Delivery stays queued and never awaited, which is the invariant that stops mutual sends between two busy agents deadlocking, and which the existing admission counter's comment already documents the hard way.
• Containment (Structural worker containment: depth cap, strip orchestration tools from workers, fail-open/fail-closed role classes (OpenCode + grok-build patterns) #393): retained children still cannot spawn or steer, addressing is family-scoped, and a resumed worker draws from the same Landing reserve protects the work but not the wrap-up: tight-budget runs finish and verify, then die narrating #390 ledger since it is the same worker continuing.
• Four slices, with slices 1 and 2 worth doing even if steering never ships, since they make worker output durable and inspectable, which the trajectory and GUI agent-overview: cheap-model session recaps + NEEDS_INPUT/COMPLETED status chips #419 overview work both want independently.

Open questions are listed rather than silently decided: retention default for workflow workers, a tighter worker transcript cap given fan-out, and whether a tombstoned child's transcript is swept immediately or with the session.