Most reactive-crowd web toys take one of two easy outs: a canvas particle system with no individual agent behavior, or a single hero character on a scripted animation loop. Neither supports a crowd that notices and reacts to a subject individually while still behaving like a crowd - canvas particles are cheap exactly because they don't negotiate with each other, and DOM elements are expensive exactly because they do. Building on raw, mutually-repelling DOM elements was the harder, less obvious choice, and the one that actually matches what the source protest looked like: not one face, thousands of ordinary ones.

The security-raid mechanic exists because the real ending of a protest engagement isn't 'you win, credits roll' - it's a spectrum from fully suppressed to holding the ground. The four crowd-recovery outcomes (full, medium, low, none) are the game-design translation of that spectrum.

A crowd needs one shared lifecycle primitive, or every new behavior duplicates the last one's bugs. Crowd spawn/despawn staggering and decay-toward-floor logic, and the security units' own spawn, repel, and despawn lifecycle, ended up sharing one EntityPool/raidRules primitive instead of two parallel systems slowly drifting apart. That consolidation only became obvious after building the second system and noticing it was quietly re-solving the first one's problems.

The physics has to hold up under a mode switch, not just in isolation. Eyes track, cockroaches swarm, fingers point, placards raise - four different visual behaviors riding the same repulsion and separation core. Building four separate physics loops for four creature types would have meant four separate places for the same bug to hide.

Satire needs a mechanic, not just a skin. The idea only becomes a toy through a reversal: the crowd surrounds and watches the dropped-in subject, instead of a single figure looming over a crowd. That reversal only reads if the tracking and repulsion behavior is convincing enough to feel like attention, not decoration - which is why the physics came before the visual polish in the build order.

No canvas anywhere. Every eye, cockroach, finger, and placard is a real DOM element, individually positioned and force-repelled every frame.

Creature modes. CreatureMode ('eyes' | 'pointedFinger' | 'cockroach' | 'placard') selects behavior and rendering inside CreatureGrid and creaturePhysics.ts. Adding a new mode is additive - it doesn't require forking the grid or the physics loop, since every mode rides the same repulsion and separation core and only swaps its own tracking and render behavior.

The raid and recovery lifecycle. Shaking the dropped-in avatar summons wandering security units that repel and permanently thin the crowd. A hold-and-release Protest button charges a power meter: full power wins outright (a win panel, the sticker locking to the floor, the avatar's repel radius widening), medium and low power partially regrow the crowd and escalate the raid instead. Crowd staggering, decay-toward-floor, and the security units' own lifecycle all share one EntityPool/raidRules primitive rather than duplicated bookkeeping.

HUD and onboarding. A DOM HUD layer - gallery, filters, a menu panel, an onboarding carousel - sits above the physics scene, with a lightweight poof as the only staged animation. Everything else is the physics resolving in real time.

This is satire built on top of a real protest movement, not a neutral tech demo - the tone is intentional, not a placeholder waiting for a more corporate rewrite.

v1 through v4 are all shipped; nothing here is a stub. It has a cross-browser compatibility pass, a unit test suite, and an accessibility/performance pass on the sticker-loading path.