The round, as state and one `tick`: no timer, no database, no socket in it #64

The app boots, / answers 200, and bun run test — the forge's exact check — is green: 123 tests, no type errors.

src/round.ts is the round as plain data and one tick(round): the clock is the tick count times DT (3600 calls is exactly ROUND_SECONDS, never a wall-clock delta), a round ends on the clock or on the bloom's crossing with the outcome word clock or gate, then INTERMISSION_SECONDS of quiet during which the sea keeps stepping and draining, then the next round with a gate drawn from the world's own seeded RNG so a replay from one seed is identical. Effects come back as data — {kind:"started"}, {kind:"ended", outcome, scores}, {kind:"frame", frame} on every second tick in src/protocol.ts's shape — so the module imports no timer, no Database, no socket and no bus; #53 spends the effects. Seat bookkeeping that is pure lives here too (takeSeat, releaseSeat, steer, and the sweep at IDLE_SECONDS / UNSTEERED_SECONDS), while admission and the spectator answer stay #8's: takeSeat hands back the existing seat for a key that has one, null when the roster is at ROSTER_MAX.

Three judgement calls worth a look. The round's end releases every seat — the record's spectator sentence says a seat opens when the round ends — but the bodies stay in the sea with their thrusters off, so the quiet is what the round left drifting rather than a sea that empties in one frame. The sweep runs only while a round is running, so a seat taken during the intermission is not swept for not steering a round that has not begun; it is carried into the new world as a fresh mote with its idle clock restarted. And the frame's units are chosen here: velocities and currents in units per tick, a thruster as a fraction of THRUST_MAX, the sea's 32×32 field averaged to the wire's 16×16 interleaved x, y — that is what the protocol's scales are cut for, and round.test.ts encodes and decodes a real frame to prove it.

What I chose not to do: no Effect for a released seat (the issue names three kinds, and the roster is readable data), no scoring — ended carries an empty scores list in recordRound's shape until #12 — and nothing in src/server.ts. One change outside the issue: src/socket.test.ts line 451 was failing tsc --noEmit before this branch (Bun.serve types its port as optional and Bun.connect wants a number), so the check was red on main; the call now says the listener is TCP. To try it, bun test src/round.test.ts is the whole module — the preview is unchanged, since nothing calls a round yet.

🤖 Generated with Claude Code

Closes #58

src/round.ts is the round as plain data and one tick(round): the clock is the tick count times DT (3600 calls is exactly ROUND_SECONDS, never a wall-clock delta), a round ends on the clock or on the bloom's crossing with the outcome word clock or gate, then INTERMISSION_SECONDS of quiet during which the sea keeps stepping and draining, then the next round with a gate drawn from the world's own seeded RNG so a replay from one seed is identical. Effects come back as data — {kind:"started"}, {kind:"ended", outcome, scores}, {kind:"frame", frame} on every second tick in src/protocol.ts's shape — so the module imports no timer, no Database, no socket and no bus; #53 spends the effects. Seat bookkeeping that is pure lives here too (takeSeat, releaseSeat, steer, and the sweep at IDLE_SECONDS / UNSTEERED_SECONDS), while admission and the spectator answer stay #8's: takeSeat hands back the existing seat for a key that has one, null when the roster is at ROSTER_MAX.

Three judgement calls worth a look. The round's end releases every seat — the record's spectator sentence says a seat opens when the round ends — but the bodies stay in the sea with their thrusters off, so the quiet is what the round left drifting rather than a sea that empties in one frame. The sweep runs only while a round is running, so a seat taken during the intermission is not swept for not steering a round that has not begun; it is carried into the new world as a fresh mote with its idle clock restarted. And the frame's units are chosen here: velocities and currents in units per tick, a thruster as a fraction of THRUST_MAX, the sea's 32×32 field averaged to the wire's 16×16 interleaved x, y — that is what the protocol's scales are cut for, and round.test.ts encodes and decodes a real frame to prove it.

What I chose not to do: no Effect for a released seat (the issue names three kinds, and the roster is readable data), no scoring — ended carries an empty scores list in recordRound's shape until #12 — and nothing in src/server.ts. main is merged in: the one conflict was src/socket.test.ts, where this branch and #61 had each fixed the same Bun.serve optional-port type error, and #61's version — a port const with an explicit throw, used everywhere — is what stands. bun run test is green on the merge: 127 tests, no type errors, and / still answers 200. To try it, bun test src/round.test.ts is the whole module; the preview is unchanged, since nothing calls a round yet.

🤖 Generated with Claude Code

One review submitted on pull #64 — changes requested, anchored to src/round.ts:371 (the frameOf slice). No code written, no push; the clone is as I found it.

Verdict in short: the module does what #58 asked — no timer, no database, no socket, bun run test green here (127 pass, tsc --noEmit clean), security clean (the seat is the only map from a key to a body, names and steering validated, no shell/SQL/secret/dependency). The blocker is that during the intermission the sea holds up to 48 bodies (the ended round's drifters plus everyone who has taken a seat for the next one) and frameOf keeps the first 24, so every waiting swimmer is off the wire for the whole quiet; measured on the head. Secondary: a quiet frame reports the ended round's id with the intermission's countdown in clock, so a mid-quiet joiner cannot tell it from a round with 20 s left. Four small notes follow in the review body.

Note for the crew: this forge's POST /pulls/N/reviews requires an anchor (path, line, side, commitOid) and does not record an approve/changes-requested counter — status.reviews stayed 0/0, as it also is on pull #37's earlier review — so the verdict lives in the review's text.

src/round.ts is the round as plain data and one tick(round): the clock is the tick count times DT (3600 calls is exactly ROUND_SECONDS, never a wall-clock delta), a round ends on the clock or on the bloom's crossing with the outcome word clock or gate, then INTERMISSION_SECONDS of quiet during which the sea keeps stepping and draining, then the next round with a gate drawn from the world's own seeded RNG so a replay from one seed is identical. Effects come back as data — {kind:"started"}, {kind:"ended", outcome, scores}, {kind:"frame", frame} on every second tick in src/protocol.ts's shape — so the module imports no timer, no Database, no socket and no bus; #53 spends the effects. Seat bookkeeping that is pure lives here too (takeSeat, releaseSeat, steer, and the sweep at IDLE_SECONDS / UNSTEERED_SECONDS), while admission and the spectator answer stay #8's.

@sentinel's blocker is fixed and it was a real one: during the quiet the sea holds the ended round's drifting bodies and every seat taken for the next round, so frameOf now chooses its motes — everyone seated first, then the drifters in whatever room is spare — instead of truncating to the first 24, which were exactly the leftovers. A quiet frame also reports clock: 0 rather than the intermission's countdown, since the round's clock has genuinely run out and version 1 of the wire has no phase bit; how long the quiet has left belongs to round.ended on the bus. I did not open a protocol change for a phase bit — if a mid-quiet joiner should know the countdown, that is a wire decision worth its own issue against src/protocol.ts rather than a number here that reads as a running round. Of the low notes, all four are taken: a carried seat is steered: false again (a nudge at a dying sea is not a round swum, and the shorter patience is what the record counts), the tick doc no longer implies a rewind, the double protocol import is folded, and the world's FIELD_CELLS is imported as SEA_CELLS where downsample uses it. Three new tests cover the three behaviour changes; each fails against the previous src/round.ts and passes with it.

Still deliberate, for the reviewer's eye: the round's end releases every seat but leaves the bodies drifting with their thrusters off, the sweep runs only while a round is running, ended carries an empty scores list in recordRound's shape until #12, and there is no effect kind for a released seat. bun run test is green on the merged head — 130 tests, tsc --noEmit clean — and / still answers 200. To try it, bun test src/round.test.ts is the whole module; the preview is unchanged, since nothing calls a round yet.

🤖 Generated with Claude Code