#3284 Remote worker keeps running stale code for an edited repo-local extension model

Problem

A connected remote worker keeps running the old code of a repo-local extension model after the model's source is edited. The edit only reaches the worker after swamp serve and the worker are restarted with a fresh --cache-dir.

Repro

Seen on swamp 20261008.143625.0-sha.ee36b409, with one machine running both the server and the worker:

1. Run swamp init in a fresh repo. Add a model under extensions/models/ (e.g. @stagetest/planner) whose method writes a resource, and create an instance of it.
2. Start swamp serve --port 19090, mint a worker token, and connect a worker from another directory with swamp worker connect ws://127.0.0.1:19090 --token <t> --label role=worker --cache-dir <dir>.
3. Run a workflow whose steps are placed on the worker (labels: { role: worker }) with swamp workflow run <wf> --server ws://127.0.0.1:19090. The step succeeds on the worker.
4. Edit the model's source so the method writes an extra field, then run the same workflow again.

Expected: the second run executes the edited code, as a local swamp model method run does.

Actual: the second run's records have the old shape. The edit took effect only after both processes were stopped and the worker reconnected with a new --cache-dir.

Notes

• This may be intended for pulled extensions (trusted-pulled bundle reuse). For a model being developed in the repo it is surprising, and nothing says the worker is running stale code.
• I have not checked whether restarting only the server, or only the worker, is enough.
• Found while testing a design for stagecraft (swamp-club #3274).