v1.6.10 — Tick budget and guest accuracy

v1.6.10 — Tick budget and guest accuracy

Tagged 2026-10-12 on merge commit ed2d980 (PR #199),
from the Improve phase of the host/client benchmark project
(PR #195).

Summary

Patch on top of v1.6.9. Four changes, each chosen because the benchmark measured a cause and
a baseline for it: the sound player no longer decodes audio on the main thread, remote tanks are
smoothed on a guest, a guest no longer gets a carried mine refunded, and a guest refreshes its
fog memory when the host reveals a tile it has already seen. No wire change: the five files that
fix the byte layout of every message are identical to v1.6.9. A game with one v1.6.10 and
one v1.6.9 player was not played live for this release.

The numbers come from 90 pair runs and 20 sweep runs, none invalid, compared with the frozen
v1.6.9 baseline. The proof session ran on macOS 27.0.1 and the baseline on 27.0; see
"Found, not fixed".

Changed since v1.6.9

• Sound player (Improve 2D). SoundPlayer.play read and decoded each sound from disk on the
  main thread, 19 to 26 ms at the 95th percentile on ticks that play a sound, and the host's
  50 Hz tick waits for that thread. Sounds are now preloaded into memory and played through an
  AVAudioEngine with a pool of player nodes. A play now costs about 0.1 ms.
  • Host tick, 95th percentile: 14 to 8 ms in the soak, 17 to 9 ms at 2 players, 19 to 10 ms at 4,
    28 to 10 ms at 16.
  • Ticks later than 25 ms: 2.3% to 0.3% in the soak, 3.5% to 0.3% at 2 players, 3.9% to 0.6% at 4.
  • Jerod playtested the sound on 2026-10-12 and found it on target.
• Remote tank smoothing (Improve 1A). The smoother was keyed on state.ticks, which never
  advances on a guest, so it always snapped. It is now driven by a counter the view owns, and it
  still snaps on a respawn or a jump of more than 4 tiles. The host tank's drawn step on the
  guest fell from a median of 100 ms to 25 to 32 ms in moving scenarios.
• Mine refund (Improve 3B). A guest spends a mine when its builder launches, but the host
  spends its own only when the builder arrives. A status message sent during the walk wrote the
  host's pre-launch count over the guest's. recvSrTankStatus now subtracts the mine the
  builder is carrying. Mine-count differences between host and guest in the soaks: 10 episodes to 1.
• Fog memory on a reveal (Improve 4B). When the host reveals a mined tile the guest has
  already seen, the guest now refreshes its memory of it. Before, a tile that fell back into fog
  was drawn from the stale plain-ground snapshot. TCPSession has a new
  onTerrainRevealed callback and GameSession applies it.

Testing

• BoloKitTests: 666 pass in 15 suites. The two known-issue tests recorded for 3B and 4B in the
  Analyze phase are now plain passing tests.
• Bolo 2026Tests: a new SoundPlayerPlayTimingTests check measures the main-thread cost of a
  play (20.9 ms before, 0.1 ms after).
• Two new RemotePositionSmoother tests; ten pass.
• The Release app builds. The Improve wire check passes: the five byte-layout files are identical
  to v1.6.9.
• Benchmark: Bench/data/improve/candidate1-2d-1a-4b-3b/ (frozen, tag
  improve/candidate1-2d-1a-4b-3b). Reading: Bench/4-improve/RESULTS.md.

Found, not fixed

• The proof session ran on a newer macOS than the baseline (27.0.1 against 27.0). Simulation
  code none of these changes touch is 36 to 48% slower in medians, so medians moved with the
  machine. The tail gains above are much larger than that shift. A same-OS control run of the
  old build is #197.
• 4B is not shown by the soak. Its signature is nearly absent in both sessions. The fog
  memory faults that remain are mostly tiles one side has seen and the other has not. Open as
  part of #198.
• 1A reaches 25 to 32 ms, not the 20 ms target, which is the floor at 50 frames a second.
  Also in #198.
• Three shell-count episodes (guest 3 above the host for 0.5 to 0.7 s) appear in the new
  soaks and none in the baseline. Nothing in this release touches shells. Also in
  #198.
• The two flaky timing test classes named in docs/STATUS.md are unchanged.

Benchmark tooling in this release

Everything below is development tooling and is off in play. The recorder is compiled in and runs
only when BOLO_BENCH is set.

• Host/client measurement system and the frozen v1.6.9 baseline
  (PR #185), the Analyze phase
  (PR #186,
  #192,
  #193) and the Improve phase
  (PR #194,
  #195).
• run-baseline.sh moves a finished session's raw logs to the archive disk.
• check-no-wire-change.sh has an improve mode that fails only on a change to one of the five
  byte-layout files.