fix(webactor): walk an undelivered report back to a caller behind the hop

[AStaroverov]

Stacked on #15 — review that one first. Base will be main once #15 lands.

Why

#15 routes a messageerror to whoever was waiting, but only forward. That covers a caller waiting on a response — the failed envelope is routed, so the report inherits its route and continues.

It does not cover a caller waiting on a request. A request is still on its way out and carries no route, so the report was terminal: only the endpoint adjacent to the failing hop learned about it. A request that failed on a later relay went on retrying every 500 ms until its abortSignal fired — the exact symptom #15 set out to remove, just one topology over.

What changed

The way back is already recorded. An envelope on its way out accumulates the checkpoints of every hop it crossed, and response already turns exactly those into a route. The same arithmetic applies here: the checkpoints minus the hop that just failed are the way back to the sender.

forward   checkpoints  cid → cid/A/B → cid/A/B/B/C   ✗ post to C fails
report    route        cid/A/B                       ← checkpoints minus B→C
back      route        cid/A/B → cid                 usual routeEndsWith / reduceRoute

The report lands on the pending request with __route === cid and rejects it. No new routing machinery — it reuses the mechanism responses already travel on.

isCheckpointedEnvelope joins isRoutedEnvelope in utils/route so the two directions read the same way.

About the crossLink bug I reported earlier

While looking at this I went back to verify the devtools crossLink defect I had described, and it does not exist. Measured with a throwaway spec dumping the event stream for worker-chain, with and without the recording that supposedly triggered it: identical graphs — 38 nodes, 25 links, 1020 messages, 4 worker threads, 12 cross-thread links — and zero type-filtered drops in that scenario, so the suspect line never even runs there.

The guard in crossLink is correct: it deduplicates the two-sided attach, where both threads announce the same pair and one arrives through ingest. A collision with a locally inferred link is not constructible, because identify builds node ids as name<threadId-pointerId> and a remote id cannot be produced locally.

Verification

Unit 140/140, e2e 32/32, tsc / oxlint / oxfmt clean.

The new test drives caller — middle — broken and asserts the request rejects with the transport's error. Without the change it hangs for the full ten seconds and fails on timeout.

🤖 Generated with Claude Code