v2.84.0 — Measured LLM cost (not estimated)

Measured LLM cost (not estimated)

pxpipe's discipline applied to the board's metrics: measure real token usage,
and only report measured cost when the measurement is trustworthy.

• SubagentStop records real usage. When a subagent finishes, the completion
  hook reads its transcript, sums the real token usage, and writes a measured
  cost for that agent — no agent self-report needed. Fail-open; opt out with
  GREAT_CTO_NO_MEASURED_COST=1.
• Cache tokens are now priced. cost-meter bills cache_creation at 1.25×
  and cache_read at 0.1× the model's input rate — ignoring them under-counted
  real spend badly (a cached turn is often 50k+ cache tokens).
• The board promotes measured → canonical past a trust gate. When enough of
  the window's completed tasks carry a real verdict cost (coverage ≥ 50%, min 3)
  and the spend is real (≥ 1¢), the board shows the measured cost and a real
  savings multiplier (source: 'measured'). Below the bar it honestly falls back
  to the time-based estimate — partial/synthetic verdict data never distorts the
  ratio.
• New scripts/lib/usage-from-transcript.mjs (+6 tests); cost-correctness gains
  2 trust-gate tests. Suites: root+board 342, lib 353, all green.

---