fabric status is broken in production: carry the reconnect total as u64

[myobie]

Fixes a live regression I introduced in #36. fabric status fails on Silber right now with Error: u128 is not supported.

The bug

LatencySummary.total_micros was a u128. u64 microseconds overflows after roughly 584,000 years, so this is not a real downgrade — it is the type the wire can carry.

Read-only impact: only the operator command. The daemon, sessions, sync, shell, and exec are unaffected, and the telemetry file keeps recording correctly. cos independently confirmed it exits 1, not 0, so scripts fail loudly rather than reading a false green.

Why it reached production

Every check that should have caught it ran against empty state, which is the one case that works:

• unit tests asserted on the struct
• the integration test called the response builder directly and never serialized it
• my hand check used a fresh node with no peers
• rollout acceptance ran at install, before the first probe populated the map

The break is time-delayed by construction. Root's acceptance was honest and could not have caught it.

Why the obvious test does not catch it either

This is the part worth keeping. A plain serde_json round-trip of the telemetry snapshot passes while the command is broken — I wrote that test first and watched it pass. serde_json handles u128 in a normal struct.

ControlResponse is an internally tagged enum (#[serde(tag = "type")]), so serde routes it through its Content buffer, and that buffer has no u128 variant. The failure is on decode, in the client.

So the regression test:

• serializes the real ControlResponse, not the inner struct
• starts from a populated map, because empty is the case that always worked

Verified both directions: it fails on the u128 type with the exact production error — Error("u128 is not supported", line: 0, column: 0) — and passes on u64.

The plain-JSON test is kept but relabelled necessary but not sufficient, so nobody mistakes it for the guard.

Rollout note

A fixed client decodes correctly against an unfixed running daemon — the daemon serializes the value fine, only the client's decode failed. So installing the new binary repairs fabric status without requiring a daemon restart. Root's call.

Tests

197 lib tests green. Changed files pass rustfmt.