Peers: tagged nodes across a call, and a resumed caller reattaches instead of relaunching

[Evanfeenstra]

Why

Two gaps in strut/run-workflow (#123), both ruled on 2026-10-09 and folded into plans/federation.md §2.2:

1. Provenance stopped at the call. The peer's stream carries each step's nodes, but the step tailed without reading them. The caller's log had the handle and the output and nothing about which graph nodes the peer's agent read.
2. A caller that restarted launched twice. The handle only existed on step.end. A caller that crashed mid-call and auto-resumed re-executed the step and launched a second run on the peer, while the first kept running, orphaned and never cancelled. Separately, tailing a peer run that had died without a terminal event polled forever.

What

• Tagged nodes. The step folds the peer run's step.end.nodes off the tail it already reads, dedupes by ref_id, tags each with the caller's peer id (AccessedNode.peer, never a URL: a URL is a second identity and goes stale on re-register), and marks its own output. A ref only means something against the graph that recorded it, so nothing resolves a tagged one locally: the projector skips it (no edge, not counted as unresolved), and the run flyout lists it under an @peer badge and opens nothing. Once read-through lands it opens through /peers/:id/graph/nodes/:ref_id.
• Reattach on resume. Right after the launch, the handle is journaled as a synthetic step.end at <path>#launch. It has no stepType, so step counts, verify and the projector never see a step execution. A re-executed step finds it in ctx.journal and tails the same peer run from the start instead of launching another. The rule: in-process a child dies with its parent, so relaunch is right; across struts the child survives, so reattach is right.
• A stale run ends its tail. The stream route passes TailOpts.live (controllers.has). A log with no terminal event and no producer now closes at EOF with done { status: "stale" }. tailPeerRun retries a stale done with backoff on its own counter, then fails peer_run_stale:. A summary-less done (the finalize race) reattaches at once.

Verified

1376 unit tests pass. New coverage:

• the two-strut suite (real HTTP): tagged nodes, the #launch record, a hand-written stale caller log resumed through POST …/resume returning the same peer run's result with one run on the peer, and a stale peer log streamed to done stale
• tailPeerRun: the stale retry/cap and the summary race
• conformance: live on both tails
• the projector and the web fold: the tag

Also run live on 2026-10-09: a local strut calling swarm38's explore over this path.

Not here

• Hive's nodesOf still needs a one-line skip of tagged refs.
• lab:peer + resolveScope (federation.md §3, step 4) is next. Until then a peer token is the peer's whole key.