Ridge baker: G-buffer cards from seams through the ray march (chain painter deleted) #170

Merged
void merged 11 commits from void/ridge-gbuffer-baker into dev 2026-11-20 03:36:49 +00:00
Member

Cut ratatoskr-backdrop:cut_spec:cut-ridge-gbuffer-baker.r2. Reports h1 to h3; Soul verdicts s1 and s2.

The baker no longer paints stage flats from peak chains. Each place authors an ordered list of split depths (card_split_km; Vienna [6, 40]). Seam searches with forward energy move each split per column into a depth gap. Each card is a layered-depth G-buffer of the terrain visible beyond its near seam: albedo, normal at the hit (across neighbouring rays at equal distance, scaled by the authored exaggeration) and log-distance. Cards are written as colour and geometry WebP with manifest v2, and the runtime lights them live.

Rulings: ratatoskr:ruling:ridge-bake-gbuffer-depth-bands, ratatoskr:ruling:ridge-cards-runtime-lit, ratatoskr-backdrop:ruling:ridge-card-splits-authored-any-count, ratatoskr-backdrop:ruling:ridge-cards-per-card-slide, ratatoskr-backdrop:ruling:ridge-cast-shadows-dropped-for-now.

  • Deleted: peak extraction, chains, slots, tapers, foot fades, the plane knock-out, sun keys, luma, sun.py and mutate_export.py.
  • Tests: 24 pass, and mutate.py kills 27 of 27 (Yggdrasil). Soul's encode round trip, receding-ridge seam fixtures and normals checks are pinned.
  • Vienna: 3 cards, 368 KB, 15.9 MB decoded. Contact sheet: https://minds.gamecult.org/void/ridge-gbuffer-bake/
  • Untouched: public/ and src/. The served backdrop does not change until the runtime cut lands.
  • Open: the any-count ruling is bounded by the decode budget at Vienna framing. Revealed-pixel masking moves to the runtime slide contract (ratatoskr-backdrop:follow_up:ridge-card-reveal-mask-in-runtime-contract). Commit 1 of the mask deletion does not build; the next commit fixes it.

🤖 Generated with Claude Code

Cut `ratatoskr-backdrop:cut_spec:cut-ridge-gbuffer-baker.r2`. Reports h1 to h3; Soul verdicts s1 and s2. The baker no longer paints stage flats from peak chains. Each place authors an ordered list of split depths (`card_split_km`; Vienna [6, 40]). Seam searches with forward energy move each split per column into a depth gap. Each card is a layered-depth G-buffer of the terrain visible beyond its near seam: albedo, normal at the hit (across neighbouring rays at equal distance, scaled by the authored exaggeration) and log-distance. Cards are written as colour and geometry WebP with manifest v2, and the runtime lights them live. Rulings: `ratatoskr:ruling:ridge-bake-gbuffer-depth-bands`, `ratatoskr:ruling:ridge-cards-runtime-lit`, `ratatoskr-backdrop:ruling:ridge-card-splits-authored-any-count`, `ratatoskr-backdrop:ruling:ridge-cards-per-card-slide`, `ratatoskr-backdrop:ruling:ridge-cast-shadows-dropped-for-now`. - Deleted: peak extraction, chains, slots, tapers, foot fades, the plane knock-out, sun keys, luma, `sun.py` and `mutate_export.py`. - Tests: 24 pass, and `mutate.py` kills 27 of 27 (Yggdrasil). Soul's encode round trip, receding-ridge seam fixtures and normals checks are pinned. - Vienna: 3 cards, 368 KB, 15.9 MB decoded. Contact sheet: https://minds.gamecult.org/void/ridge-gbuffer-bake/ - Untouched: `public/` and `src/`. The served backdrop does not change until the runtime cut lands. - Open: the any-count ruling is bounded by the decode budget at Vienna framing. Revealed-pixel masking moves to the runtime slide contract (`ratatoskr-backdrop:follow_up:ridge-card-reveal-mask-in-runtime-contract`). Commit 1 of the mask deletion does not build; the next commit fixes it. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
void added 11 commits 2026-11-20 03:36:48 +00:00
Cut 1 of ridge-gbuffer-baker (rulings ridge-bake-gbuffer-depth-bands, ridge-cards-runtime-lit,
ridge-cast-shadows-dropped-for-now). Deletes first: extract, chains, sweep, slots, tapers, the painter,
the relight family, shadows, sun.py, sun_set, luma_range, check_near_cover, the painter inputs in
places.json, and every test and mutant that exercised them. export.py is left unrunnable until the
card writer replaces write_bake; the ray caches stay valid (march, rays_key, stage0 unchanged).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ridge.py: elev (shared by march and prefilter), prefilter, normals, set_view from the march skyline,
coverage, seams (one forward-energy DP per authored split), cards (first visible hit at or beyond the
near seam, crest coverage, cross-fade over seam +- XF, nested alphas so the composite equals the opaque
coverage, cropped to the row band). load_preset requires card_split_km (strictly increasing, at most
MAX_CARDS - 1). export.py: manifest v2, two WebPs per card, check_cover, check_budget on count, total
bytes and decoded bytes, sheet. places.json: Vienna card_split_km [6, 40], prefilter_m 100.
The seam prior is 0.05 rows/nat^2 (tie-break only): at 2.0 the seam hopped across a receding ridge
between two columns, which the spec's own side-by-side test forbids.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The prior is the value where the side-by-side test passes with the swept-rows term and fails without it
(0.05 and below let a seam without the term follow the ridge, 0.15 and above hop across it with the term).
mutate.py holds 16 mutants of the card rules, the cache keys and the export checks, every one killed.
README rewritten for cards, manifest v2, the seam rule and the low-pass.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
cards() cuts from each card the pixels the nearer cards hide at every offset within MAX_SLIDE_PX
(ceil(5% of the card width), the runtime's DRIFT_FRAC), also when one card slides alone; the composite
at rest is unchanged. Manifest v2 carries max_slide_px. pack() refuses an empty card naming its splits.
coverage() no longer counts the rows below the first sample twice. SEAM_PRIOR 0.02.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Decode the written WebPs against the cards (channels, range, straight alpha); the six receding-ridge
fixtures with the 0.35 alpha-jump metric; no slide within the bound reveals a cut pixel; empty card;
SEAM_SEP, elev curvature, coverage rows, interpolation and renormalisation. mutate.py grows to 33
mutants. README describes the reveal bound.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The slide check shifted the albedo along rows, so erosion 0, single-card and frame-edge mutants
survived; it now shifts the last axis. A second test sets the seam prior to 0.1 on the 4 to 15 km
ridge, where only the swept-rows term prevents the hop (at 0.02 the mutant without it cannot tear).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Side-by-side nearer ridges, overlap at the edge of their union and a ridge across the window each
need a different term of the cut (single-card slide, slide width, frame edge). The frame-edge mutant
now treats what lies beyond the frame as held (the old one was equivalent).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The slide contract (viewport-aspect drift, simultaneous slides) lives in the runtime cut; a baker
bound of 5% of the card width was unsound and did not lower the decode total. cards() is again one
rectangle of rows with coverage; the schema is regenerated and committed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ridge-bake tests: the Vienna fixture no longer cuts for a slide
All checks were successful
Build pull request / Build pull request (pull_request) Successful in 3m23s
a9eb9acbf9
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
void merged commit 030a4a9231 into dev 2026-11-20 03:36:49 +00:00
Sign in to join this conversation.
No reviewers
No labels
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
GameCult/Ratatoskr!170
No description provided.