Ping, Which Had Read Zero

Moving geometry, lag-compensated hit registration, a menu that became the front door, player accounts opened by key, and the first matches that count.

Ping, Measured

LagCompensation had existed since v0.9.0 — and was wired only into a CLI harness. Every shot on the live server resolved against the present.

It now rewinds to the view the shooter actually had, using a tick the client was already reporting and the server was already throwing away. 100ms maxCompensationMs.

Two instruments came with it:

text

npm run test:lifts    walks every lift on every installed map and fails if one
                      cannot be called  -> found four dead lifts on wctf6

npm run net:viewlag   holds frames in both directions over a real socket
                      -> found the interpolation assumption double-counting

And ping, which had read 0 for every player since the field was introduced, is measured.

Movable Geometry

func_plat and vertical func_door are simulated, collided, predicted, heard, and crush whoever blocks them — one fixed-step system shared by server, client prediction and single-player, so "a rider is carried by the same code everywhere".

A Networking Audit, Not A Feature

Snapshots never said which state they were deltaed against, so a dropped datagram — or merely an ACK still in flight — could leave a component silently wrong until it happened to change again. Fixed at the protocol level. Snapshots are now message 0x0B; 0x04 is the old envelope, kept only for reading old data.

One shape behind a family of replay bugs:

something that should have been a function of which simulation step this is was
instead counted per call, and reconciliation counts the same steps more than once.

Knockback drag never ran on a live client at all.

Drawing Less

World culled by compiled PVS, players culled and quieted with it, autosprites stop rewriting vertices nothing moved.

Spatial batching was cut. It removes a third to a half of the triangles and costs about 20% more frame time, so r_bsp_cellsize stays off — with 36 captures behind the decision.

One thing was added rather than trimmed: tiered impact sparks, so a machinegun round stays quiet and the flourish is kept for weapons that hit once and hard.

The Menu Became The Front Door

A gold rail, one stage beside it, and rooms you can actually get into.

Player accounts opened by a key rather than a password, so a name belongs to somebody. Rooms can be private, with an invite only a link carries. Rooms have their own chat.

Amphi arrives as the mode the lobby leads with: 1v1, three rounds, a kit you pick and live with.

Matches that count. A seat is a signed ticket, the child server signs the result it reports, the lobby stores it once, and the leaderboard counts only matches most of whose room was people.

The Seven-Bot Bug Hunt

docs/research/v0.9.23-bughunt.md — mostly about other players. They were the only bodies in the map exempt from culling, dead-reckoned through walls when a packet arrived late, shared a single skeleton between them, and the trail on the rocket you fired at them collapsed to a flat sliver "in the one case that matters — shooting along the axis you are looking down."

Known Unverified Items

Long, and mostly unflattering, so here is the shape of it:

The lobby has never been used by two people at once, and no ranked result

between two humans has ever appeared on the board in front of anyone.

The menu has had no accessibility or real-display pass.
The placement bug is "believed fixed, not proven fixed" — no controlled

reproduction was ever built. cl_remotehitbox 1 is the cheap discriminating test still worth running.

Two fixes failed to hold: the lower-floor rocket launcher and the tab-return

camera were both reported still broken. "Each fix closed a real defect … but neither was the cause of what was reported."

A player froze mid-match while the server ran on and "What stalls the connection is

not known."

Only hitscan is rewound — projectiles still spawn against the present.
Bots do not use ladders; ladder climbing is unreachable on 35 of 39 installed

maps.

Client presentation is frame-rate dependent; uncapped frame rates are not a

supported configuration.

Babylon's per-frame CPU counters cannot be trusted as totals. 9.17

PerfCounter.endMonitoring(false) assigns its current duration rather than accumulating; cpuRenderMs reads 0-0.1ms and activeMeshes reads zero.

One Deliberate Choice

Under-compensating a jittery connection is intentional, "since erring high punishes the player who took cover."