Residential ambient activity

Roman and Carthaginian houses look inhabited: a hearth vents smoke through an opening the model actually has, shutters stand open by day and close at dusk, washing hangs out, cloth stirs in the breeze, lamps come on at night, residents step outside, and very occasionally one carries a bowl of soup out of the door and wears it. Everything here is presentation-only. No population, selection, collision, pathfinding, vision, resource output or replay state is added, and nothing reads the live simulation while drawing.

Nothing in lockstep

The headline rule: in a town of a hundred houses, no two may look like copies and none may change state together. Every periodic thing here derives its offset, period, duration and magnitude from ambient_hash of the entity id (roll(seed, salt) in home_activity.cpp hands out an independent unit float per knob), never from raw animation_time alone. Slow global changes -- dusk arriving -- act through per-house thresholds and per-house phases, so a street responds one house at a time. tests/render/home_activity_test.cpp proves this for every schedule by sweeping the night factor and counting how many houses flip per step.

No chimneys

An ancient Mediterranean house has no chimney, so neither does this. Each nation registers a HomeSmokeAnchor from its own house renderer, pointing at an outlet that building already models:

The anchors are in house-local units and are carried through the building's own model matrix, so a plume scales and rotates with the house rather than floating at a fixed world height. The Punic anchor is deliberately off-centre: the oven is a corner structure, and a centred plume would read as a chimney.

The same anchor carries the rest of the house's dressing: the doorstep and its outward direction, the hanging cloth, the lamp, the side-wall windows for the shutters, and where a washing line can be strung.

The plume

hearth_smoke is its own effect kind, not a recoloured dust cloud or flame. CombatDust is a ground-hugging dome and BuildingFlame adds ember sparks unconditionally; neither can pass for domestic smoke. The branch (u_effect_type == 7 in combat_dust.vert/.frag) uses twenty overlapping, camera-facing wisps in one draw per hearth. Each wisp rises from the outlet, expands slowly and drifts downwind, with coherent noise breaking up its soft edges. Both ends of its lifetime fade to zero before it is recycled. There is no enclosing cone or tube, and opacity follows the fire intensity all the way to zero. The house-local plume scale is 0.68 for Rome and 0.60 for Carthage, keeping cooking smoke modest above the roof.

The shader is unlit, so the caller dims and thins smoke at night. Wisps use ordinary alpha blending with depth testing and no depth writes; they carry no sparks or glow. Their mesh is separate from combat dust and building flames.

Every hearth is a different fire

hearth_character(id) hashes size (0.8-1.25x), density (0.62-1.0x), colour temperature, rise-clock speed and a large clock offset. Tended fires breathe on a slow per-house pulse. The shader also hashes the vent position for rise speed and wind drift, and varies each new wisp while it is invisible at the lifetime boundary. Neighbouring hearths share the prevailing wind direction but never emit identical wisps together. The noise clock is the house's own (time * speed + offset), never raw animation time.

It emits no local light. Only flames and lamps feed the light buffer, and a cooking fire seen through a roof should not illuminate the street.

Scheduling

Each house is lit for a stretch and then goes out. Its offset, period, duty cycle and strength all come from a hash of its entity id, so neighbours never share a cycle and a street never animates in lockstep. Roughly a third of eligible houses are alight at midday. The intensity ramps in and out over about nine seconds at each end, so a hearth is banked up and dies down rather than switching.

Meal-times

HomeActivity::meal_bias() stretches every house's lit window by the light: 0.8x at full day, 1.35x in twilight (the evening meal cooked as the light goes, the morning one as it returns), 0.64x at full dark when most fires are banked. The bias moves slowly with the night factor and each house crosses its own window edge at its own phase, so dusk lights hearths one by one; the test sweeps the night factor in a hundred steps and allows at most three of sixty four houses to change per step.

Zero-health, ruined, neutral, previewed, pending-removal, dying and dismantling houses show nothing. Eligibility is resolved for every house in one scan of the locked render snapshot at the top of the frame; submission is then a binary search.

A house carries a UnitComponent, so the render snapshot files it with the units and it is drawn from Renderer::plan_unit_entry, not from Renderer::submit_non_unit_entry. That is where the per-frame HomeActivity is attached to the draw context.

Night

Night is not a clock. mediterranean_summer still reports a primary intensity of 0.44 at 21:00 and only the colour goes dim, so render/scene_walk.cpp measures the light the scene actually casts (

luma(primary_color) *
primary_intensity

Lamps

lamp_state(id, time) is the oil lamp inside the door. About a fifth of households are simply dark. The rest each light up at their own point in the dusk (threshold 0.10-0.52 of the night factor) and ease in over their own stretch of it, so a street comes alight window by window. Each lamp also has a long schedule of its own (240-460 s, ~70-90% lit) on which it goes out for a while — someone went to bed — and its own flicker: a quick wobble under a slow one, 1.6-3.8 Hz at a per-house phase and depth. Colour runs from amber to a deeper orange per house, always candle-warm. What is drawn is a dim pool on the threshold (alpha 0.42 at most) and a small local light (radius 3.8, intensity 0.72 at most): restraint sells it, because the effect shader is unlit and a bright glow reads as neon.

Dressing

All of this is skipped below fourteen projected pixels.

Cloth

A run of five strips, each in two links: the upper hangs from the top edge, the tail hangs from the upper's hem and lags it with a larger swing, so the free edge moves further than the hung edge and the strip bends rather than tilting. The gust arrives at each strip a little after the one before it, so it passes along the run instead of lifting every strip together; per house the breeze strength, gust rate, flutter rate and phase all differ. A Punic house hangs an indigo valance under its roof awning; a domus curtains its door, and that curtain (cloth_at_door) is pushed aside from the middle out as each resident goes out and again as they come in (curtain_parting).

Shutters

Both models cut windows in their side walls; each now has two leaves hinged at its jambs. shutter_angle(id, window) settles each of the four windows as closed (22%), ajar (30%) or thrown open flat against the wall (48%), so the same street is never four identical facades, and closes them for the night, each window at its own moment in the dusk.

Washing

Some houses (36% Roman, 48% Punic) have washing out. A domus strings a short line between two of its portico columns; a Punic house runs one across its flat roof from the awning post to a pole by the far parapet. One to three pieces in undyed linen, ochre, indigo or madder, each its own width and height, sway on the line and are brought in one at a time as dusk comes.

Residents

doorstep_plan(id, time) is the ordinary ambient, far more common than the gag. Each house keeps its own rhythm (69-136 s) and within it each turn outside is decided afresh from a hash of house and cycle: whether anyone comes out at all (58% of turns), for how long (12-34 s), what they do, and whether a second resident joins them. Activities:

Everyone walks out through the door and back in through it; the walk phase follows distance so nobody skates, and the walk path bends from the door to where they settle (their own spot across the door, 1.1-2.0 m out). Off the plinth they stand on the terrain. Residents are rigged actors and share the per-frame actor budget; a pair costs two.

Residents turn over 0.6 s instead of snapping (on arrival, before the walk back, and at each end of an errand) and fade walk and activity clips into each other over 0.35 s. The errand's walk phase counts the whole distance covered, so the return leg steps forward instead of replaying the outward stride backwards, and a resident standing idle breathes on a 7-9 s cycle, close to the baked 8 s idle.

Budget

Reduced effects means no rigged actor at all, not merely a cheaper one. Houses below nine projected pixels or beyond 190 world units are skipped (plumes thin out over the last quarter of that range rather than popping), and residents need twenty-two pixels before they will play. Earlier visible houses consume the budget. Fog is checked on current visibility, never on merely explored ground, so neither smoke nor lamplight can betray an unseen residence.

Soup spill

One renderer-wide slot permits at most one sequence, and it reuses the nation's civilian rig with a clay bowl hung off it as ordinary humanoid equipment. Beats are constants in Render::GL::Spill, in seconds from the moment the slot is claimed:

The path runs straight out from the doorway along the house's own facing for 2.4 m, so it never crosses the building and never needs the navigation grid. The walk phase advances with distance covered rather than wall-clock, so the resident does not skate. The spill is drawn as overlapping discs rather than one disc, because a single disc reads as a painted circle, and it fades out with the sequence.

HomeActivity::cooldown_seconds defaults to 600 and clamps to at least 300. Entity/cycle hashes select only about a quarter of eligible windows. Losing eligibility or visibility releases the slot on the next frame rather than stranding it.

Reproducible checks

Build render_tests and arena_app. Run:

QT_QPA_PLATFORM=offscreen LC_ALL=C build/bin/render_tests --gtest_filter='HomeActivityTest.*'
DISPLAY=:0 build/bin/arena_app --batch --scenario home_ambient_street --clean-capture --capture-interval 6 --artifact-dir /tmp/home-street
DISPLAY=:0 build/bin/arena_app --batch --scenario home_ambient_dense --clean-capture --capture-interval 8 --artifact-dir /tmp/home-dense
DISPLAY=:0 build/bin/arena_app --batch --scenario home_ambient_night --clean-capture --capture-interval 8 --artifact-dir /tmp/home-night
DISPLAY=:0 build/bin/arena_app --batch --scenario home_ambient_street --fog-of-war --clean-capture --capture-interval 6 --artifact-dir /tmp/home-fog
SOI_HOME_GAG_SECONDS=12 DISPLAY=:0 build/bin/arena_app --batch --scenario home_ambient_soup --duration 22 --scenario-distance 0.55 --clean-capture --capture-interval 1 --artifact-dir /tmp/home-soup

Add --scenario-distance 0.5 to read the shutters, cloth and residents. Two review hooks exist because both behaviours are deliberately rare:

• SOI_HOME_SMOKE_ALWAYS=1 lights every eligible hearth, which is the only practical way to inspect both nations' anchors in one capture, and the right way to measure a dense street's worst case.
• SOI_HOME_GAG_SECONDS=<n> forces the spill every n seconds.

home_ambient_street and home_ambient_dense destroy a house partway through; it must stop smoking at once. home_ambient_night moves the clock to 21:00 so the plume and the lamps can be judged against a dark sky. building_preview cannot show any of this — it replays captured primitive draws on the CPU with no effect or rigged-mesh path — so the Arena is the review hook.