The Atlas, part 1: fight where you stand

[Justrada]

Reopened against main — this was PR #2, which GitHub auto-closed when its stacked base branch was deleted on merge of #1. Same commits, same branch.

Overnight work on the map feature. This is the foundation plus a visible first slice — you can pick a ground type when staging a fight and the battlefield is generated, drawn, and enforced.

What to look at first

docs/MAPS.md is the real deliverable — the canonical design for the whole feature, with the reasoning recorded so it outlives any one implementation. §9 has five open questions that are genuinely yours to answer.

Then: host a table → Begin Combat → The Ground.

How the design was chosen

A 14-agent council: 4 recon agents mapped the existing seams, 3 architects proposed independent architectures under deliberately opposed priors, 6 adversarial judges attacked them, and a chief architect synthesised. Separately, deep research into procedural generation, hex tactical design, and the VTT map market.

The judges did something better than agree with me — they read this working tree and found that all three proposals were written against stale recon and would have replaced the edge-wall arena model with something worse. The synthesis kept what had shipped and designed the world layer around it.

The architecture, in one line

The world is a seed. The battlefield is compiled data.

Every hard constraint here is a constraint on bytes — clipboard-only sharing, ~5 MB of shared localStorage, PeerJS with no chunking, SVG-only rendering. A world is about one fifth the size of skillTree.json. The atlas doesn't point at a battlefield, it compiles one, which is why the map ships with zero new combat concepts: HexCoord never grows a Z axis and a generated fight rides the combat_start that already exists.

What's in here

Walls are edges, not cells. A wall sits on the border between two hexes, so a seven-room interior costs ~100 edge records and zero standable ground — versus ~93 of 252 hexes (37% of the board) if walls were tiles. It also gives doors the right arity: an edge door joins exactly two hexes, where a door hex would open in six directions. Foundry and Roll20 both model walls this way.

Line of sight gates targeting at any range, including Battlefield. It's deliberately permissive — the hex-line rounding isn't symmetric (8 of 31,626 hex pairs disagree about who can see whom), and players forgive "they shouldn't have seen me" more readily than "I can't shoot the thing I'm looking at."

Three arena archetypes, each designed against a specific failure:

• interior — BSP on the offset rectangle, because axial space is sheared and subdividing there yields parallelograms rather than rooms. Doors follow a spanning tree plus loops: a pure tree gives every room one approach and the fight becomes a queue.
• cave — cellular automata at B4/S4. The canonical 4-5 rule is tuned for eight neighbours and does not transfer to six; ported naively it gives mush or solid rock.
• open — obstacles at minimum separation 2, which makes the field connected by construction rather than by luck.

Tests assert across 40 seeds per archetype that no hex is stranded and both deploy zones are standable and mutually reachable.

Determinism primitives chosen against measurements, not familiarity: sfc32 over the Weyl generators (where a seed is an offset into one global sequence rather than a selector between independent ones), 53-bit seeds (32-bit collides around 65k entities), prefix-free path hashing (otherwise ['ab','c'] and ['a','bc'] become the same place), and an fBm normaliser that's fixed rather than a running sum — which is what makes zoom add detail instead of rescaling terrain that's already on screen.

purity.test.ts enforces all of this mechanically: it reads every file in src/engine/atlas/ and fails on Math.random, Date.now, console, timers, DOM, React, or a store import.

Bugs found and fixed along the way

Each of these was real and pre-existing:

• HexBoard's layout memo had an empty dependency array while reading the grid size — it would have silently kept rendering the first geometry it ever computed.
• The reachable-tile highlight ignored terrain, so the board would have offered tiles through walls the engine then refuses. The action menu had the same gap.
• Occupancy was three different rules. The engine tested currentHP > 0, so a mover walked onto an unconscious body; the store and board tested !isDead. The board's hex→combatant map is last-writer-wins, so the stacked combatant vanished. Now one OCCUPIES predicate at all four sites.
• The engine mixed Math.random() into every log id, which made "combat is deterministic given a seed" quietly untrue.
• A live multiplayer bug: the GM never sent hello, so the handshake was one-sided. A GM page-reload restarts the outbound sequence counter at 0 while a player still holds a high watermark — and every subsequent snapshot is dropped silently, forever. The GM now greets joiners and the player re-baselines.
• text-glow-mystic was used on the Marketplace heading but never defined in CSS.

Also adds CI

main deploys on merge, so there was no point at which a broken branch could be caught. .github/workflows/ci.yml now gates every PR into main with typecheck, the full suite, and the same production build the deploy runs.

Verified

• tsc --noEmit clean, npm run build clean.
• 438 tests across 42 files, up from 311. Every pre-existing test passes untouched — the arena is an optional field, so a fight without one behaves exactly as before.
• Checked in the browser: staged an interior fight, confirmed walls render on the hex borders, closed doors show as gold dashes, and both foes deploy into their nominated zone.

Not in here

Phases 2–7 from the doc: the scale ladder and zoom camera, multi-floor structures, the authoring tools, split-party sync, marketplace bundling, and the globe.

🤖 Generated with Claude Code