fix(strata): PRD 618 complete the forest before revealing it

[jonit-dev]

Keep the loading cover up until the initial world and render pipelines are ready. The current licensed overview comparison still shows missing distant trees. This PR is a draft while the remaining visual and performance criteria are investigated.

• Verified: 5/10 PRD boxes. Six headed NVIDIA browser runs measured median readiness of 12,558.5 ms → 8,977.5 ms (ratio 0.715). The browser and desktop reveal scenarios each passed 10/10: pipelineCensusAvailable=true, post-reveal pipelinesCreated=0.
• Open: AC-5 reveal-to-+3 s visual comparison was NEUTRAL on both pairs. AC-7’s fresh licensed overview judge returned REGRESSION on both pairs (missing trees behind the lake and on the right hillside); river was NEUTRAL. Four labelled overview images and judge context.
• Still needed: reliable terminal-distance evidence and final steady-state performance qualification. The earlier full-source gate at 7967 with 10,727 tests passed is historical evidence, not proof for the current final candidate.

The PR remains draft: CI has no checks, so no green CI result or full acceptance is claimed.

[jonit-dev]

Startup BEFORE/AFTER screenshots — two headed NVIDIA WebGPU runs per arm. Baseline source 88b988cd7; local startup candidate 3ccb616b6 (implementation push pending final local checks). These are procedural fallback frames; they do not qualify the licensed far-tree criteria.

Fresh read-only visual judge: NEUTRAL — AC-5 not passed. Both BEFORE and AFTER appear stable between reveal and +3 s; the judge found no visible improvement. AC-5 remains open. The zero late-shadow counters alone are not a visual verdict.

Run / frame	BEFORE	AFTER
1 / reveal
1 / reveal-plus-3s
2 / reveal
2 / reveal-plus-3s

Controlled timing, three alternating runs per arm on the same preview-server setup and GPU: median 10,831.4 → 11,572.7 ms (+6.844%), within the PRD’s 1.25× bound. All three startup-candidate reveal scenarios pass with zero late view publications, shadow invalidations, and pending impostor bakes. Final foliage integration must repeat these checks.

[jonit-dev]

Leaf clipping fix is pushed in aa545be (current PR head 5d3817a).

Fresh read-only visual judge: IMPROVEMENT in both runs. BEFORE has straight clipped card edges; AFTER preserves complete needle tips and internal transparency holes.

Run	BEFORE	AFTER
1
2

Scope: actual procedural crown material and atlas, same close-up pose. The generator now fits complete needle strokes inside each tile instead of clipping them at its boundary. Fine tips remain jagged and dense bands remain flat-looking. All four formal captures verified NVIDIA/Turing with fallback=false and produced byte-identical images to these judged pairs; this visual verdict does not pass startup AC-5 or licensed-foliage AC-7.

[jonit-dev]

Pushed 085d01e: native staging no longer places its manifest in the web root; stopping a game cancels startup timers instead of reporting a delayed false failure. Focused regression checks, 93 startup/game tests, full typecheck and lint pass.

The leaf BEFORE/AFTER comparison and IMPROVEMENT verdict remain the current visual proof.

Not ready to merge: the full terrain run exposed WebGPU vertex-input overflow in the new dither path, in both view and shadow pipelines. That engine regression is being fixed. Startup AC-5 still fails; licensed-foliage acceptance and final paired startup timing remain open.

[jonit-dev]

Current captures are below. They do not yet prove the promised visual improvement. A fresh read-only judge rated both matched timepoints and the overall comparison NEUTRAL; AC5/7 remain open.

• BEFORE 943a2458a: reveal
• BEFORE 943a2458a: +3 seconds
• AFTER 1ca595062: reveal
• AFTER 1ca595062: +3 seconds

These are headed Intel gen-12lp WebGPU captures at 1280×720, with procedural fallback assets and CPU profiling enabled. Original pixels are preserved below the labels. There is one run per arm; this is diagnostic evidence, not the required repeated visual or unprofiled startup qualification.

The CPU profiles show reflector-inclusive sampled time falling from 2,812 ms to 14 ms. This is a measured CPU improvement, not a visual verdict. The subsequent native forest reveal scenario now passes with no diagnostics and 10 observations over 3.1 seconds; its separate native command-buffer discard fix is local and has not yet been pushed. Licensed-foliage visual acceptance and the final startup ratio remain open.

[jonit-dev]

Native reveal proof from local source f0c4693ae: two runs, 8/8 assertions each, ready in 29.53 / 29.27 seconds. All three late-work counters are zero; both runs report zero console errors, network errors and runtime diagnostics.

Time	Native run 1	Native run 2
Reveal
+3 seconds

Fresh read-only judge: NEUTRAL in each run and overall for native temporal stability. Same scene and camera, no added or missing terrain/tree/prop masses, no major shadow or exposure change; small foliage motion remains.

These captures use procedural fallback assets. Licensed far-tree visual improvement, browser timing and pipeline census remain pending. PRD progress is 3/10; the PR stays draft. Original screenshot bytes are preserved on prd-618-screenshots at 7e0adc536222447994606cdb299e3216e33240db.

[jonit-dev]

Latest visual proof: not ready to merge. PRD progress is 5/10 verified locally.

Matched licensed-asset comparison: all 67 models, both cooks at 1024 texture size, same overview pose at 1600×900, headed NVIDIA/Turing WebGPU, two runs per arm. BEFORE source 88b988cd70d5db7266eb0281e5f74bd0e030b3e7; AFTER source 7967fe079eb4fbce1f79a40813bb989c770804b0.

Fresh read-only visual judge: REGRESSION on both overview pairs. AFTER loses tree coverage behind the lake and on the right hillside. River pairs were NEUTRAL. These screenshots demonstrate the outstanding failure; they do not establish successful visual completion. The separate first-uncovered/+3s licensed video frames were NEUTRAL, so AC5 remains open too.

Run 1: BEFORE / AFTER

Run 2: BEFORE / AFTER

Current work: isolate streaming, terminal LOD selection and impostor drawing before changing the implementation. No merge queue submission until the remaining acceptance criteria pass.

[jonit-dev]

Current proof: native passes; visual acceptance still fails. PRD remains 5/10 verified and the PR stays draft.

The corrected native project root passes the licensed reveal + overview scenario: 12/12 assertions, 1,147 frames, zero console/network/runtime diagnostics, NVIDIA/Turing RTX 2080. The earlier texture errors came from the test fixture’s project root; no native engine fix is claimed.

Matched visual comparison: two headed Intel gen-12lp runs per arm, 1600×900, all 67 licensed models with matching 1024 cooks. BEFORE is source 88b988cd; AFTER is the recorded dirty-worktree snapshot based on a3b20b0f (runtime diff 4a9aa3c8…), with the scaled spruce terminal gate and baked-alpha raymarch correction.

Fresh read-only judge: REGRESSION on overview in both runs; river NEUTRAL. AFTER has missing foreground coverage at 4s and conspicuous tree fill by 19s. BEFORE remains stable. At run 1 overview, AFTER admission backlog is 53→0 and impostor draws 1,046→2,247; loads and bake queues are already zero. Investigation is now on the CPU admission path (gpuScene: false), not GPU selection.

Run 1

Time after overview	BEFORE	AFTER
4 s
19 s

Run 2

Time after overview	BEFORE	AFTER
4 s
19 s

All 16 labeled overview/river captures. No visual completion, final performance result, or green CI claim is made.