fix(kpi): show a dash, not zero, when the platform keeps no totals

[htcom-code]

Closes #81

Why

On a platform with per-module metrics off (protean.trace.metrics.enabled=false, the quickstart default), the KPI row showed Requests · total 0 and Error rate 0.00% while traffic was flowing. I saw this live on 2026-10-12 against quickstart during a 3,000-request burst. The platform reports the situation correctly: its ready ack says metricsEnabled: false, and its metrics frames are []. The two tiles summed that empty array and turned "not aggregated" into a confident zero. The module table already reads the ack (#63); the tiles didn't.

What

• The two tiles now follow the module table's ack-driven reading, checked in the same order:

  Situation	Tiles
  metrics has rows	the numbers (rows are data, even if the live setting changed after the ack)
  tracesEnabled: false	— and "recording off"
  metricsEnabled: false	— and "metrics off"
  metricsEnabled: true	the numbers, so 0 here is a true zero
  no ack and no rows	— and "not reported"

  tracesEnabled is checked first. metricsEnabled is the effective value, so turning recording off also reports metrics as off, and "metrics off" would point the operator at the wrong switch.

• The tiles do not fall back to summary.count. That count is capped by the ring capacity (measured: 340 requests reported as 200), so it would understate traffic more the busier the platform is. The summary trend lines (vs prev, err) stay as they are.

• The reason sits where the unit would (— metrics off), so each tile stays one line tall and its sub-line is unchanged.

• The helper is private to kpi-row.tsx, so the file adds no lint warning.

Verification

• 7 new scenarios (245 tests in total). They replay the captured protean, Spring and Go connect sequences and check every row of the table.
• Revert check: with main's kpi-row.tsx put back, 4 of the 7 fail. The other 3 cover behaviour that has to stay the same: a true zero, rows shown despite an "off" ack, and Go sending rows.
• Gates: the full suite passes run serially (245/245). Lint is clean (the same 7 warnings as main), and the build passes.
• Live, against quickstart on :8080 with metrics off (ready: tracesEnabled: true, metricsEnabled: false, metrics: []) and about 650 /items/add requests:
  • Both tiles read — metrics off, matching the module table's "Module metrics are off".
  • The summary sub-lines still show (↓ 67% vs prev, — 0.00pp · 0 err).
  • The latency tile and chart are unchanged.
• A layout fix came out of the live check (second commit). The reason was first appended to the sub-line, which wrapped the requests tile onto two lines and made it taller than the rest of the row. It now sits where the unit would, after a non-breaking space.