test(e2e): clear identities and pins between suites

[jamofer]

What

Wipe the hub, agent and client state directories when the bench sets up, so each suite starts from a clean identity world.

Why

Running the whole suite fails 3 of 56, always with the same hub message:

registration rejected: agent name pinned to a different fingerprint

Bench._reset() tears down namespaces, veths, the bridge and /etc/hosts, but never touches work_dir. The hub's TOFU pin database therefore outlives a suite. A later suite that reuses an agent name (truck42, edge) with a key from a different server's agent-state is then correctly rejected — the hub is doing exactly what it should, and the bench is lying to it.

The three failures are reconnect.robot (both WAN uplink cases) and socketcand_write_burst.robot. All three pass in isolation, which is what pointed at ordering rather than a product bug.

How

_clear_state() removes hub-state, agent-state, client-state and web.db per server at setup. logs/ is deliberately kept — it is the diagnostics for the run, and wiping the whole work_dir would delete the previous suite's logs.

Testing

• Full suite before: 53 passed, 3 failed
• The three failures in isolation: 4 passed, 0 failed — confirming state pollution, not a regression
• Full suite after: 56 passed, 0 failed

Notes

Not a product defect and not a release blocker: the e2e suite is not wired into CI yet (TST-1). It does matter if e2e is ever going to be a release gate, which is why it is worth fixing now rather than filing.

Related smell left alone: every agent on a server shares one agent-state, so two agents on the same host would share an identity. No current suite does that.