* Acceptance: add the load/fault tier (16/32/64 clients, burst, churn, handoff)
Extend the end-to-end harness (issue #15), harness-only — no server change.
Uses the existing ServerTuning seam to widen the local connect budget and
capacity so many clients can run over a single loopback IP.
- 16/32/64 clients across distinct cells: unique server-owned ids and no
cross-cell transform spam (generalized from the prior 16-client case)
- burst transform traffic engages the rate limiter without tearing sessions
down, and the server still relays a fresh transform afterward
- reconnect churn (12 cycles): ids stay monotonic and never reused, and each
disconnect leaves no stale active session
- authority handoff stays deterministic under churn: retiring the current
authority repeatedly hands off to the next lowest id with strictly
increasing epochs
Packet loss/reorder on snapshot traffic is inherently an unreliable-transport
(GNS) property and stays deferred with the GNS sections (blocked on #2).
26/26 pass on real sockets, deterministic across repeated runs.
* Strip verbose comments from the acceptance harness and ServerTuning
Remove the prose/narration comments across the end-to-end harness (keeping only
section dividers) and the ServerTuning header block, matching the repo's terse
comment style.
