v0.4.1 — Accurate TPS, subagent throughput, cache hit rate, theme colors

The throughput figures are now measured correctly, and three community features
land alongside. Tested on OpenCode v1 and v2. Install with
opencode-tps-meter@latest.

Accurate throughput

• No more five-digit spikes. Hosts batch output before publishing it (v2
  flushes every ~100ms), and the first batch of a stream was divided by the few
  milliseconds before the meter's first update, so readings like 10364 appeared,
  most visibly as a subagent's frozen average. Each chunk is now rated over the
  interval it was actually generated in, and the opening chunk counts toward the
  total but never toward a rate.
• The live rate is right from the start. It used to begin at a fixed 100 TPS
  and climb toward the real figure over several seconds on fast models. It is now
  exact from the second chunk (about 100ms in) at any speed, with no smoothing lag.
• Settled averages describe what streamed. A step's average is its streamed
  tokens over the time they streamed, corrected by the provider's count only when
  the two agree within 2x. Provider totals that include output the host never
  streamed, such as hidden reasoning, no longer inflate the rate.
• Tool-call arguments count as output (v2). The text a model writes for a tool
  call (a file write, an edit) is now counted, so a step that only calls a tool has
  a real reading instead of none.
• Calibration persists and is weighted by size (v2). The live estimate's
  per-model correction is saved, so new sessions start calibrated, and small steps
  no longer swing it. Steps whose reasoning count the provider leaves ambiguous are
  left out.
• A stalled stream shows its rate falling instead of holding its last value.
• One clock per measurement. v1 no longer divides by server completion time
  minus a local first-token time.
• The v1 toast keeps its main line on the primary session (a subagent's chunk
  used to relabel the subagent's rate as the main one) and no longer drops an
  update that lands inside its throttle window.
• Dashboard. gen is replaced by tools, the time a turn spent running tools,
  since tool time is never inside a step's rate window. The per-model ledger starts
  fresh, as its old best/mean figures were recorded with the spiking formula.
• Same-named subagents in the footer get a short id suffix, e.g. explore(9B5c),
  and a step that ends on an error now settles the reading and still counts its
  usage.

Added

• Multi-session footer (#8, @ondrejch). While subagents run, the footer
  switches to an aggregate line, e.g. TPS Σ 318 | main 63 | explore 91 | +2.
  Σ sums every session generating right now, and each agent is labeled by name.
  Columns keep their spawn order, so they don't swap places. A subagent stays
  visible through long tool calls and drops off about 4s after it finishes.
  Works on both hosts: v1 resolves the session tree from session.created /
  session.updated and api.state.session, v2 from ctx.data.session.
• Prompt cache hit rate (#9, @Anigx). The footer shows
  cache N% · session cache N%: the cached share of the prompt for the last
  step and for the whole session, computed as
  cache.read / (input + cache.read + cache.write). It also appears in the v2
  dashboard and the v1 final-stats toast. The figures appear once a session's
  first step has finished. Toggle with showCacheHit (default true) or
  TPS_METER_SHOW_CACHE_HIT.

Fixed

• Theme colors on current v2 hosts (#10, @dragonish). v2 renamed its theme
  tokens to text.base / text.muted / feedback.<kind>.base. The meter now
  resolves those names, falling back to the older ones, instead of rendering in
  the host's default color. The theme is also read live, so a theme switch
  recolors the meter without restarting the TUI.
• Cache figures after a tool-call-only step. A step that ends in a tool call
  without streaming any text now updates the footer's cache figures on both
  hosts, instead of leaving the previous step's numbers on screen.

Notes

• New config option: showCacheHit (default true), env
  TPS_METER_SHOW_CACHE_HIT. No other config changes.
• Thanks to @ondrejch, @Anigx and @dragonish for the contributions.