[stellarfeline]

What this fixes

A release-quality render round on the cheerful town campaign found six engine defects that stop its showcase pictures being right. Each is fixed where its cause is.

1. A story moment holds the cast the cast ledger stations there. A camera's world after a step stood every NPC at its load position: move-npc effects never reached it, a deferred spawn-npc fired only by a flag-gated effect was unforced in the critical path's ledger, and a branch's world laid only the setup rows. Now every NPC the cast ledger declares stands where cast::station puts it at that moment of that path (the quest of the next step, the flags the path's flow journal holds), resolved as a talk-to step resolves it (plan::body_station, BodyScope::Beat) — moved, summoned from its ledger rows, or taken out while declared absent; a body a planned walk leaves on the mark faces its last step (view::moment::cast_at, bodies_of).

2. DW0721 refuses only a moment whose blocks, cast and sky are all load's. A step after which only NPCs move (or only the sky changes) is a moment; the refusal and its remedy now say "no block, no member of the cast and no sky differs". The binding line states <m> of the cast moved, sky <as at load | set by the story>.

3. A moment's world is the load world with every write fired on the way laid, in firing order, and the count is the bytes'. Two causes: (a) region writes were read under the proofs' causal ancestry, so a write at a step that is no causal ancestor of the next step vanished from later moments (candles lit, then unlit, then lit again); a picture is of one play of the path, so PathModel::at_moment reads the writes in the path's order. (b) The ledger's block rows were laid after the configuration's writes, so a setup row (a candle placed unlit at world init) overwrote the beat that lit it — the reason a written world could hash equal to load while "3 cells moved" was printed. Every write now goes through view::moment::Laid: setup rows, then per step its region writes (Configuration::laid_writes carries each write's firing step) and its ledger rows, laid by the one writer nav::lay_writes (which RegionState::blocks_over now calls too). The printed count is the count of cells the written bytes differ from load's. A branch's moment fires the ledger along the branch's own steps (placed::Ledger::on_path, firings_of taking the path).

4. A far frame loads the declared horizon. Where the horizon's base lays ground past the build (an ocean's sea), a frame whose rays meet that surface past the build's extent loads those chunks out to where its fog closes (frames::horizon_cover), and the fog stays where the build's extent puts it. The yellow rectangle in the sea of the plan frame is the sun's reflection in the water: the sun at the picture's tick (daytime 10500, altitude 34.9° on the game's eased curve, due west) reflected in a flat sea is seen 239 blocks west of the lens's nadir, which projects to x ≈ 347 of the 1600-pixel frame at the lens's own z (the frame's middle row); the rectangle stands at ≈ (345, 450), and the shape is the game's square sun texture. It is the path tracer's specular water, not a body.

5. A body the renderer draws nothing of, in frame, refuses the approved picture (DW1100), and only that. delvec cameras names every body of the frame's world whose kind the renderer's own statistics for that frame report skipped and that stands in frame, nearer than the fog and in sight of the lens; an hq frame is not written (exit 5), a preview is drawn and names it. The engine keeps no list of kinds: the refusal reads what Lumenwright reports skipped, so on Lumenwright 0.4.0, which draws the happy ghast (and the allay, boats, minecarts, sheep wool, the warden, the firework rocket), only what 0.4.0 still skips (a bat among them) is refused.

6. NPCs face the lens. In play a body that watches turns to the nearest player within its reach; a picture's lens stands where a player stands for it. This is a picture-time reading of the body's own declared watch (spec-0101), not a new rest facing: the engine already declares the rest facing (the anchor's), and the watch is what decides how the body faces a player. A watching body within its reach of the lens turns to it by the game's look-at bearing; the frame stands in a copy of its world with its bodies turned (world::copy_with_entities). A body declared not to watch, or watching a class, keeps its rest facing.

7. Lumenwright 0.4.0, and the sea plane over an ocean. The renderer moves to =0.4.0 (crates/delvec/Cargo.toml, versions.toml [render], frames::RENDERER_VERSION, Cargo.lock, the acknowledgements row). An hq frame whose world's horizon is an ocean states the renderer's sea plane at the level above the topmost generator water block (frames::sea_plane_at / sea_plane, one rule read by the showcase cameras, the review frames, the panorama and the single frame); every other horizon states none. The sea then reaches the horizon where the fog-less hq frame used to end it as a band.

DW1100 is the one code allocated (handed: DW1100, DW1101; DW1101 unused).

Proof

Defect	Test	Red on main	Green here
cast at a moment (items 1, 2), watch at the lens (item 6)	it::story_moment::a_moment_holds_the_cast_the_ledger_stations_and_a_watcher_faces_the_lens	main's delvec cameras refuses the camera after obj/look with DW0721 ("block for block"), exit 2	the keeper at [3.5, 65, 4.5] after obj/look, [5.5, 65, 4.5] facing north at load; each facing copy's keeper at the bearing to the lens
firing order and the count (item 3)	it::story_moment::a_moment_lays_every_write_its_path_fired_and_counts_what_it_wrote; it::post_beat_camera::a_beat_s_write_lands_over_the_setup_s; view::moment::tests	main writes no gold at 1,65,1 after obj/talk and prints 6 cells moved	the gold is there, 7 cells moved, equal to the written worlds' difference
horizon (item 4)	view::frames::tests::a_frame_looking_past_the_build_loads_the_horizon_to_its_fog	— (new function)	the cover past the build, bounded by the fog, none for a frame that sees nothing past it
renderer (item 5)	it::story_moment::a_frame_of_a_body_the_renderer_cannot_draw_is_refused (a bat keeper); it::story_moment::a_happy_ghast_is_drawn_in_hq_and_a_bat_is_still_refused	on 0.3.0 the happy ghast is skipped and refused	exit 5, DW1100 naming minecraft:bat dw_npc_keeper at [5.5, 65.0, 4.5], no picture; a preview is drawn and names it; in a world holding both, the renderer reports the ghast drawn and the bat skipped
sea plane (item 7)	view::frames::tests::the_sea_plane_is_stated_for_an_ocean_horizon_only; it::story_moment::the_sea_plane_changes_an_hq_far_frame_over_a_sea	— (new)	the rule states sea_level + 1 for an ocean and nothing otherwise; the same hq far frame over a sea differs in pixels when the plane is stated

The town, redrawn (preview, fixed engine; content campaign/town-rpg @ 3655d61f, no content commit)

• The moments: after Lady Agathe's talk 1 of the cast moved (Agathe to the square); after the loaf 4 (Agathe, Sanna, Oskar, Runa on the square); the ending on chorus-sings+oven-loaf+nest-moved 8; on …+dumpling-tail the deferred npc/dumpling-square stands over the square and npc/dumpling is gone. The candle steps' worlds hold the lit candles and agree byte for byte (after-obj-find-the-wings and after-obj-climb-to-the-tail share one sha256).
• Runa (forge-near), Fenna (tavern-near) and four more frames' watchers face the lens.
• sea, plan, from-the-castle and castle-far load the sea to their fog; in hq the sea plane carries it to the horizon: the sea draft shows the sea to the horizon with no band under it.
• castle-far frames the happy ghast at the castle gate: on 0.4.0 its hq picture is drawn (the ghast is a white box-shaped body at the gate at draft size), no longer refused.

Checks (head 3128edb99; the town renders on the local Mac, the gallery baseline on the remote build machine)

• fmt (tools/ci/fmt-workspaces.sh): 2 of 2 workspaces clean; clippy --workspace --all-targets -D warnings: clean.
• cargo test -p delvec --lib: 1417 passed, 0 failed.
• cargo test -p delvec --test it over the render and frame tests (frames, story_moment, post_beat_camera, view_cli, camera_sky, review_frame_sky, codes, panorama, cameras): 115 passed, 0 failed, of 3078 in the binary; the full it runs in CI.
• Gallery baseline (tools/ci/gallery-baseline.py, built from source, compared with the merge base): REPORT — 10 builds each side, 10711 emitted paths, 0 added, 0 removed, 0 changed; 10 recorded values moved (the dw_codes lists and the renderer version); warning rows 283 against the ledger's 283. No row to regenerate.
• Docs-job checks of ci.yml (17 listed, the fetch step aside) and validation/check-versions.sh: all exit 0.

Fixes #1326 #1327 #1328 #1329 #1330

Later commits

• Lumenwright's pin registry entry moves to 0.4.0 with its sites.
• The gallery probe a-beat-that-changes-nothing is removed. Under the corrected moment rule every step the gallery's critical path carries once leaves load in a block, the cast or the sky, so no single camera edit of the gallery reaches DW0721's "equal to load" cause any more. That is a loosening of the gallery's demonstration of that one cause; the refusal itself is unchanged and the cause stays asserted by the compiler's own test (it::post_beat_camera), which is where a refusal of the compiler's output is proven. DW0721's other three gallery probes are unchanged.

🤖 Generated with Claude Code