Skip to content

Commit b222709

Browse files
committed
docs(decisions): three r174 kills the GL-error storm — and is still not landing tonight
Kevin authorised landing the upgrade if it works ("if the upgrade works, then do it"), then went to sleep. It does not clear the bar I set before measuring, so it is reverted and this is the record instead. MEASURED, with a presence control, production bundle, identical 12s probe: three 0.172.0 + postprocessing 6.39.1 -> 6 glBlitFramebuffer messages three 0.174.0 + postprocessing 6.39.4 -> 0 The control is the point. A zero from an instrument never shown to produce a non-zero is not evidence, and this repo has shipped that mistake seventeen times. That confirms the upstream maintainer's theory on pmndrs/postprocessing#750: the GL storm we had filed as a SEPARATE production-bundle finding and the GodRays regression share a root cause in three r172's depth-texture handling (mrdoob/three.js#30585, fixed r174). A gap this exposed, worth more than the upgrade: prod-smoke.mjs CANNOT SEE this error class. It filters on m.type() === 'error' and these arrive as 'warning', so the repo's own production gate is structurally blind to the error storm it exists to watch. WHY IT IS NOT LANDING: r174 changes the whole look. 19 of 31 frames go over gate, many at 99%+ local density, and the sky is visibly more saturated -- a global colour/tone change, not one subsystem regressing. Landing it means rewriting all 31 baselines, which is a judgement about how the game should LOOK. That is Kevin's call, not a correctness fix the loop can self-gate, and a bundled engine-bump + re-baseline is precisely the shape baseline-trailer exists to stop. It also runs against c1d933e, which deliberately blocks 0.x MINOR bumps for the engine stack. This measurement is evidence FOR that policy. Reverted fully: package.json + package-lock.json restored from backup, npm ci re-run, 0.172.0 / 6.39.1 confirmed installed, tree clean. The upstream reply is DRAFTED, NOT POSTED -- a public upstream comment is outward-facing and stays Kevin's. It says plainly that I cannot yet confirm r174 fixes the SUN, because the global change makes that unseparable without a controlled r174 pair I have not run.
1 parent b0147d5 commit b222709

2 files changed

Lines changed: 95 additions & 0 deletions

File tree

‎docs/superpowers/DECISIONS.md‎

Lines changed: 56 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -18,6 +18,62 @@
1818
1919
---
2020

21+
## 2026-08-13 (night) — [LOOP] three r174 KILLS the GL-error storm, and is still NOT landing tonight
22+
23+
Kevin, going to sleep: *"if the upgrade works, then do it."* It does not work in the sense that clears the
24+
bar I set before measuring, so it is not landed. Both halves below are measured, with controls.
25+
26+
**Where this came from.** `vanruesc` replied to our upstream report (pmndrs/postprocessing#750, labelled
27+
`investigating`) within hours. They could NOT reproduce the missing sun on three r172 — but they DO see the
28+
`glBlitFramebuffer: Read and write depth stencil attachments cannot be the same image` error we had filed
29+
SEPARATELY as a production-bundle finding, and traced it to a three.js depth-texture-cloning bug fixed in
30+
r174 (mrdoob/three.js#30585). Their claim: postprocessing's `e19bfa3` triggers a bug that lives in older
31+
`three`. So two findings we had been treating as unrelated are one root cause.
32+
33+
### ① Does r174 fix the GL storm — **YES, measured, with a presence control.**
34+
Identical probe, identical 12s window, production bundle both times:
35+
36+
| build | `glBlitFramebuffer` messages |
37+
|---|---:|
38+
| three 0.172.0 + postprocessing 6.39.1 (committed state) | **6**, both message forms |
39+
| three 0.174.0 + postprocessing 6.39.4 | **0** |
40+
41+
The control matters and is the reason this is stated as fact: a zero from an instrument never shown to
42+
produce a non-zero is not evidence. **`prod-smoke.mjs` cannot see this class at all** — it filters on
43+
`m.type() === 'error'` and these arrive as `warning`, so the repo's own prod gate is structurally blind to
44+
the error storm it was written to watch. That is a real gap and it is now written down.
45+
46+
### ② Does r174 let us unpin postprocessing — **UNRESOLVED, and it is not the blocker.**
47+
Not answered, because the question is moot until ③ is decided: `explore-day` shows no sun disc at that
48+
camera angle in EITHER build, so the frame that would answer it has to be identified first.
49+
50+
### ③ Why it is not landing — **the upgrade changes the whole look.**
51+
`three` r174 against the committed baselines: **19 of 31 frames over gate**, many at 99%+ local density
52+
(`explore-day` 99.70% vs a 9.30% ceiling; `spell-iceball` 99.90%; `ocean-depth` 99.19%). Comparing the two
53+
`explore-day` frames by eye, the r174 sky is materially more saturated — this reads as a global colour or
54+
tone-mapping change, not a regression in one subsystem.
55+
56+
So landing it requires rewriting all 31 baselines. **That is a judgement about how the game should LOOK,
57+
which is Kevin's call, not a correctness fix the loop can self-gate** — and a bundled engine-bump +
58+
re-baseline is exactly the shape `baseline-trailer` exists to stop (ten of the last twelve baseline
59+
rewrites were bundled, which is how four monster baselines became pictures of an empty mountain).
60+
61+
It also sits against a live decision-of-record: `c1d933e` deliberately blocks 0.x MINOR bumps for the
62+
engine stack in dependabot, because for 0.x the MINOR position is the breaking one. This measurement is
63+
evidence FOR that policy, not against it.
64+
65+
**State: fully reverted.** `package.json` + `package-lock.json` restored from backup, `npm ci` re-run,
66+
three 0.172.0 / postprocessing 6.39.1 confirmed installed, git tree clean. Nothing about this experiment
67+
is committed except this entry.
68+
69+
**Owed to Kevin in the morning — two separable decisions, do not bundle them:**
70+
1. Take r174 for the GL fix alone, accepting a full reviewed re-baseline of the look? (High value: the
71+
storm currently MUTES the GL channel, so a genuine later error in that context is silenced.)
72+
2. Reply to #750 with the measurement above. Drafted, NOT posted — posting to a public upstream issue is
73+
outward-facing and stays Kevin's.
74+
75+
---
76+
2177
## 2026-08-02 — the audit batch (items ①–⑨), decided under the standing grant
2278

2379
Kevin's instruction, verbatim: *"what do you recommend for all the punted / awaiting decision questions?
Lines changed: 39 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,39 @@
1+
# Draft reply to pmndrs/postprocessing#750 — NOT POSTED
2+
3+
Posting to a public upstream issue is outward-facing, so it stays Kevin's call. This is the text, ready
4+
to paste. Everything in it is measured, with the control stated, because the maintainer will act on it.
5+
6+
---
7+
8+
Thanks for looking so quickly.
9+
10+
**Your depth-texture theory is confirmed on my end for the GL errors.** I measured the production bundle
11+
with an identical probe and a 12-second window on both builds, headless Chromium 151 with SwiftShader:
12+
13+
| build | `glBlitFramebuffer` console messages |
14+
|---|---:|
15+
| three 0.172.0 + postprocessing 6.39.1 | **6** — both `Read and write depth stencil attachments cannot be the same image` and `Depth/stencil buffer format combination not allowed for blit` |
16+
| three 0.174.0 + postprocessing 6.39.4 | **0** |
17+
18+
The zero is against a control that produced a non-zero on the same instrument, so it is an absence rather
19+
than a blind probe.
20+
21+
Worth flagging why this mattered to us beyond the noise: in a longer session the errors repeat until Chrome
22+
emits *"too many errors, no more errors will be reported to the console for this context"* — after which a
23+
genuine GL error later in that same context is silenced too. So the storm disables the channel our capture
24+
harness watches, which is how it went unnoticed.
25+
26+
**On the missing sun, I cannot yet confirm or deny that r174 fixes it**, and I would rather say so than
27+
guess. Upgrading to r174 changes our rendering globally — 19 of our 31 committed reference frames go over
28+
tolerance, with a visibly more saturated sky — so I cannot currently separate "the sun is back" from "every
29+
frame changed". Isolating that needs a controlled pair (r174 + 6.39.1 vs r174 + 6.39.4) which I have not
30+
run yet. I will follow up with that result rather than leave it implied.
31+
32+
One detail from the original report that may help you reproduce, since our setup differs from the demo:
33+
our `GodRaysEffect` light source is a sun disc that is **reparented into the effect's own `lightScene`**,
34+
and the effect owns a `renderTargetLight` with its own `DepthTexture`, with `clearPass.enabled = false`.
35+
If your demo's sun stays in the main scene, that difference is a plausible reason the demo renders fine on
36+
r172 while ours does not.
37+
38+
Repro environment: three 0.172.0, @react-three/fiber 9.5, @react-three/postprocessing 3.0.4, Vite 6,
39+
Chromium 151 headless (SwiftShader) and desktop macOS.

0 commit comments

Comments
 (0)