[yujiaxu-ops]

Step 2 of the plan for the live passes' layout, after #68. Three passes bound more than a browser's eight storage buffers per stage, so the browser needed its own merged layout: a rewrite of the live source and slabs indexed through an offsets uniform. This PR rearranges those three so every pass binds at most eight. The browser now uses the native layout, one per pass. Tier R1/R2: pictures change only by rounding (see the checks below).

The change

• post (13 storage buffers): the view modes become passes of their own before it (view_raw, view_hdr, view_sample). A frame showing the lit picture does not run them. The per-pass profile times an empty pass for each, so its timestamps keep PASSES' order.
  • One behaviour change: an HDR snapshot taken in a view mode is now that view's signal. Before, it was the lit image. Only a scratch tool switches view modes.
• compose (11): the upscaler's per-sample state (Flicker, Lock) moves to a pass of its own after it, track.
• dtemporal (10): the surfaces' compact copy (sb) is now written by refl, which reads each sample whole anyway. A separate pass for it cost 0.18 ms, because it read the samples a second time.
• The sample's parts are read separately (sample_diffuse, sample_specular, sample_emission), so a pass binds the resolved reflection only if it takes the specular part.
• merged.rs's live half is gone: the slabs, the offsets uniform and the rewrite of the live source.
  • The kernel's outputs that the live passes read now start at multiples of 256 bytes in its merged buffer (@align(256), with the struct's size rounded to match), so a page can bind those ranges directly.
• Two guards: a unit test (every_pass_fits_a_browser) holds every pass to eight, and a view in the merged layout checks the same when it starts.

Merge gate

check	result
cargo test --release	125 passed
render-identity vs #68's build, native	168 byte-identical, 19 within run-to-run spread, 3 differ by rounding: particles-motion-repose (277 pixels, at most ~3e-6 nits, all on dark pixels), wind-pivot-sway-live and voxel-cut-body-push (~3e-5 relative)
render-identity, THREERT_MERGED=1	170 / 17 / 3 by rounding: skinning-shared-pose-split-live, volume-grid-uniform-live, wind-pivot-sway-live (at most 7.5e-5 nits, on values of 1 to 4)
view modes 1–8 (material tests, THREERT_VIEW_MODE)	every PNG byte-identical in every mode
tools/web-hello.ts in Chrome on this GPU	pass (the adapter reports 10 storage buffers per stage; the device is held to 8)
cost	+0.18 ms, almost all from the compose / track split (track 0.27 ms, compose 0.08 ms less). Bistro, 1440p at 0.375, alternated: the passes after the trace 6.00–6.08 → 6.19–6.24 ms. The sb copy in refl costs +0.02; the view passes cost nothing in the lit picture.

The rounding differences come from code that moved: the same expressions, split or merged differently, compile to arithmetic that differs in the last bits, and a temporal test carries those bits along. The largest absolute difference is 7.5e-5 nits.

ROADMAP § 1a records the step.

🤖 Generated with Claude Code

---

Note

Medium Risk
Touches the full live rendering pipeline and browser binding path; identity checks show tiny rounding drift and a small perf hit, but view modes stay byte-identical.

Overview
Rearranges the live post chain so every compute pass binds at most eight storage buffers, letting the browser use the same per-pass bind layouts as native and removing merged.rs's live shader rewrite (slabs, offsets uniform, merged bind group).

Shader / pass split: view-mode visualization moves out of post into view_raw, view_hdr, and view_sample (skipped in the normal lit frame). Upscaling flicker/lock state moves from compose to new track. Surface buffer sb is written in refl instead of dtemporal. sample_parts becomes sample_diffuse / sample_specular / sample_emission so passes only pull the buffers they need.

Rust: live/targets.rs always builds one bind group layout per pass; kernel merged buffers expose live-readable ranges at 256-byte alignment for WebGPU offset rules. layout::every_pass_fits_a_browser and runtime checks in merged mode enforce the eight-buffer cap. Docs (ROADMAP, RUNTIME) reflect the new model and measured +0.18 ms cost from the compose/track split.

^{Reviewed by Cursor Bugbot for commit 14b5f39. Bugbot is set up for automated code reviews on this repo. Configure here.}