docs(opstack): record the green build and rewrite the handoff for a fresh context

OPSTACK-PROGRESS gets the measured entry: every number the run produced, what each one settles, and
the single open warning. OPSTACK-PLAN's status header goes from UNVERIFIED to BUILT AND GREEN.

OPSTACK-HANDOFF is rewritten end to end. It no longer describes a transition in progress but a
completed one, and it leads with the two things a cold context needs: the Underwater 0%-coverage
warning (with the truncation-vs-floor finding behind it, and the reminder that a green bit-identity
over solid rock is not evidence), and the one task everything else now waits on -- making
FRoomGraphSource::EffectOverBox answer spatially.

It also carries forward the two debts that must be paid BEFORE that lands rather than after: box
bounds computed from strate params can be too optimistic once a per-room op raises them, and AUDIT
C2 is confirmed but unfixed on the switch path.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
2026-07-28 14:13:59 +02:00
parent ce409e7bb1
commit b533294f5d
3 changed files with 208 additions and 248 deletions
+131 -245
View File
@@ -1,266 +1,147 @@
# Handoff — VoxelForge operator stack, 2026-07-28 (end of day 3)
# Handoff — VoxelForge operator stack, 2026-07-28 (Phase 2 complete and green)
> Paste the block below into a fresh session. Everything it refers to is on disk and in git.
>
> **State:** Phase 2 is **DONE — 8 of 8 archetypes ported, built, and green** (14 tests).
> `ClassifyTile` consumes `ClassifyBox`. One open warning, one clear next task.
---
You're picking up an agreed refactor of the VoxelForge UE5 voxel plugin, on branch `experimental`
(already checked out — do not create another). I'm Jahni. Design and progress are written down so you
don't re-derive them.
You're picking up the VoxelForge UE5 voxel plugin on branch `experimental` (already checked out —
do not create another). I'm Jahni. The design and the history are written down so you don't
re-derive them.
## Read first, in this order
1. **`CLAUDE.md`** — project rules. **Rule #1 is absolute: never build, compile, or run the editor.**
I build everything myself. When code is done, stop, say "ready to build", list the likely
compile-error spots, and wait.
2. **`OPSTACK-PROGRESS.md` — THE LAST ENTRY FIRST.** Append-only log of what is built vs merely
written. This is the resume point.
3. **`OPSTACK-PLAN.md`** — the plan. **§2.6.1 is the current acceptance bar** and supersedes §2.6.
4. **`OPSTACK-DECOMPOSITION.md`** — all 8 archetypes broken into ops. §2 TunnelNetwork (the one in
progress), §0.2 the worm bound, §7 islands, §5 SurfaceWorld.
5. **`AUDIT-2026-07.md`** — §C9's library half is the top open risk. **§C10 is SOLVED — don't
reopen.** §C1 was reopened and re-closed on 2026-07-28; §C2 has a new *suspected* item worth
verifying before acting on.
2. **`OPSTACK-PROGRESS.md` — THE LAST ENTRY FIRST.** Append-only log; the resume point. The last
entry is the green build with every measured number in it.
3. **`OPSTACK-PLAN.md`** — the plan. **§2.6.1 is the acceptance bar** and supersedes §2.6.
4. **`OPSTACK-DECOMPOSITION.md`** — per-archetype breakdown. **§0.2** (the amplitude bound) is the
live one; §2 TunnelNetwork and §8 Underwater are now history, not instructions.
5. **`AUDIT-2026-07.md`** — **§C2 has a CONFIRMED sub-item as of 2026-07-28, read it**; §C10 is
SOLVED, don't reopen; §C9's library half is the top open theoretical risk with 0 measured
exposure.
6. **`CODEMAP.md`** — navigation. Trust symbol names over line numbers.
## Where things stand — 6 of 8 ported and wired, the 7th half-ported, 13 tests green
## Where things stand — the transition is COMPLETE and VERIFIED
All 8 archetypes have an operator-stack twin, per-strate opt-in, each equivalence-tested **bit for
bit** against its original density function. The `switch` and the stack are now two complete,
interchangeable implementations.
| Archetype | State |
|---|---|
| `Maze` | ✅ ported, bit-identical, wired |
| `FlatPlain` + `CrystalChamber` | ✅ **one op for both**, bit-identical, wired |
| `SurfaceWorld` | ✅ ported incl. **biome blending**, bit-identical, wired |
| `VerticalShafts` | ✅ ported, bit-identical, wired — 3 ops reused from Maze unchanged |
| `SurfaceWorld` | ✅ ported incl. biome blending, bit-identical, wired |
| `VerticalShafts` | ✅ ported, bit-identical, wired |
| `FloatingIslands` | ✅ ported, bit-identical, wired — the stack that runs **backwards** |
| `TunnelNetwork` | 🔶 **STAGE A of 3 done and verified.** Not wired — see below |
| `Underwater` | ❌ TunnelNetwork + a water flag; folds in at stage C |
| `TunnelNetwork` | **19 ops**, bit-identical incl. all 12 detail modifiers + per-room override |
| `Underwater` | ✅ same builder, second `case` — ⚠️ its check has a coverage gap, see below |
Everything is behind `UVoxelStrateDefinition::bUseOperatorStack`; the ported list lives **only** in
`UVoxelStrateManager::UsesOperatorStackForChunk`. Un-ported archetypes ignore the flag, so ticking it
anywhere is harmless.
Everything sits behind `UVoxelStrateDefinition::bUseOperatorStack`; the ported list lives **only** in
`UVoxelStrateManager::UsesOperatorStackForChunk` (now all 8). **No strate asset has the box ticked**
— that is my call and I haven't made it. But the flag is no longer a no-op anywhere: ticking it now
really switches that strate onto the stack, for density *and* for tile classification.
## TunnelNetwork is staged — read this before touching it
`ClassifyTile` **consumes `ClassifyBox`** for cave archetypes (SurfaceWorld and bedrock gaps keep
their hand-written exact-lattice proofs). `GetDensityAt` and `ClassifyTile` build the stack through
the **same** factory, `VF_BuildOpStackForChunk` — a second copy would be a hole, not a bug.
`GetDensityWithParams` is ~1080 lines: 13 detail modifiers, a two-region cache, worms, and a per-room
op override. Porting it whole before anything can be checked would be ~600 unverified lines on top of
~200 — the `AUDIT §P3` pattern this refactor has dodged seven times. So:
## The one open warning
- **Stage A — DONE, bit-identical over 6000 samples.** Vertical scale · base rock · cave warp · room
graph (+ pits + chimneys) · carve · worms · structural post. **7 ops.**
- **Stage B — NEXT.** The 13 detail modifiers of `STEP 4b4h`, gated on `Sdf < SDFBlendRadius·3`.
They insert between the carve and the worms, so the op-count assertion in the test must move.
- **Stage C.** The per-room op override (`§2`'s option (a)), the `Underwater` water flag, and only
then does `UsesOperatorStackForChunk` return true for either.
```
Underwater (stage C2): bit-identical across 2000 samples — 0 of them in open cave (0.0%)
```
**Why stage A is verifiable while incomplete:** every detail modifier is amplitude-gated, and
`FStrateGenerationParams` already defaults all of them to zero. Zeroing `SurfaceRoughness` (the one
exception) sends the *original* down exactly the path stage A ported. Stage B's test will do the
reverse — turn them on one group at a time.
**A green bit-identity over 2000 samples of solid rock is not evidence** — it is exactly what two
agreeing voids look like. Same failure as stage A's 1.1 % run, in a different slot, caught by a
counter written for it.
**⚠️ The decision that must not be undone:** `FRoomGraphSource` **CALLS** `BuildChunkCache` /
`EvaluateSDFCached`; it does not transcribe them. That is where `ARCHITECTURE §8.4`'s two-region
window-invariance discipline lives, and a transcription would *fork* it — with the fork "validated"
by a test that compares it to the original. Only the ~60 lines of glue are transcribed.
A real bug surfaced while diagnosing: the sampled chunk-Z range used `Z / CHUNK_SIZE`, and C++
integer division **truncates toward zero**. TunnelNetwork is at the top of the layout in positive Z
where truncation == floor, so it could not show there; Underwater is at the **bottom, in negative
Z**, where it shifts the upper chunk bound a notch high and the `Clamp` piles samples into the top
seal band. Fixed (`FloorDivChunk`), sampling widened 8 → 24 clusters, **and not trusted**: new
**check 5b** gives each of the three possible causes its own number and prints how to read them.
## Three things Phase 2 settled that were not in the original design
**⚠️ That commit (`ce409e7`) is written but NOT BUILT.** First action: build, run the `VoxelForge`
filter, read the `Underwater diagnosis` line. It names the cause; the fix follows from it.
1. **Height space** (`VoxelHeightOp.h`) — a *second operator family*. SurfaceWorld's terrain ops read
and write an **altitude**, not a density: no input Z, XY-pure per column. They do not fit
`IVoxelDensityOp`. §0.1 found density needed a second *channel*; this found terrain needs a second
**space**. Bonus: a height stack *cannot* hold Z-dependent data, because there is no Z in the
signature — `AUDIT §6.3`'s hazard became a type error instead of a convention.
2. **`IVoxelBiomeField`** — ops depend on a *capability*, never on `UVoxelGenerator`. The adapter
(`FGeneratorBiomeField`) lives in `VoxelGenerator.cpp`. This is what lets ops become assets in
Phase 3; an op holding a generator pointer never could.
3. **`FRAME` ops are RETIRED — porting all three candidates killed the idea.** `CaveWarp`'s scope is
exactly one operator (pits/chimneys explicitly read *unwarped* coords, the thing §2 called the
fiddliest in the decomposition — inside one op it evaporates). `VerticalScale` is `Z / Scale`, a
one-line pure function. The island warp was already local. **Zero frames from three candidates:**
not missing infrastructure, one idea seen three times from a distance. Marked retired in §1.
## Then the one task everything is waiting on
## What is left, in the order I'd do it
**Make `FRoomGraphSource::EffectOverBox` answer spatially.**
1. **Stage B** — the 13 detail modifiers. The bulk of the remaining lines.
2. **Stage C** — per-room op override + `Underwater` + flip both on in `UsesOperatorStackForChunk`.
**The override is NOT optional polish:** pits, chimneys and columns exist *only* through a
per-room `UVoxelTerrainOpDefinition` (`BuildChunkCache` opens its bake with
`if (!CR.RoomOp) continue;` and then reads a *fresh* param struct). No override ⇒ no pits, ever.
3. **The worm amplitude cap (`DECOMPOSITION §0.2`)** — the largest single perf item in the plan.
TunnelNetwork proves **0 of 40** tiles today because a fielded-noise carve has no spatial bound.
But its amplitude is bounded and trivial (`t ∈ [0,1]`, `Mask ∈ [0,1]` ⇒ at most `WormStrength`
toward air), and that bound already has a home in `FWormFieldSource::MaxCarveAmplitude()`. It
needs a fold that carries **numbers**, not just directions.
4. **`ClassifyTile` still uses hand-written guards.** `ClassifyBox` is brute-force verified per
archetype but **nothing consumes it in production**. That is where measured tile-skipping turns
into frames — arguably the biggest single win still on the table.
5. **PERF — parked by Jahni until the transition is complete.** The op path is measurably slower.
One cause found and fixed (the column memo discarded itself every chunk). Remaining suspects in
order: the hashed column lookup vs `GSurfColCache`'s direct-indexed box, then per-voxel virtual
dispatch. **Measure before optimising** — that is the C10 lesson.
6. **`AUDIT §C9` library half** — `sinf`/`cosf` are not IEEE-754 specified, so MSVC's CRT and glibc's
libm can differ. Currently **0 samples within 1e-6 of the isosurface**, i.e. no measured risk. The
real fix, if ever needed, is a deterministic in-house sin/cos. Also: run `CrossPlatformDigest` on
Linux, compare the SHAPE digest, pin it.
7. **`AUDIT §C2`, suspected** — `GetGenerationParams` blends params across Gradient transitions, so
two chunk Zs inside one strate may hold different params with the same XY box, strate index and
seed ⇒ the SDF cache never rebuilds ⇒ the lower chunk gets the upper chunk's rooms. **Verify the
premise before acting** (does Gradient blending actually vary within a strate?). The op stack does
not inherit it — its key folds in a params CRC.
8. **VerticalShafts proves 0 of 60 tiles.** Pessimistic, not wrong: `EffectOverBox` returns
TunnelNetwork proves **0 of 40** tiles today, and the test asserts that. The chain dies at the room
source, which returns `Both` with unknown amplitude before anything downstream is reached. Its room
and tunnel bounds (`FCachedRoom::CullRadiusSq`, `FCachedTunnel::BoundRadiusSq`) are **already in the
SDF cache**; what it costs is building that cache for the *queried box*, on the querying thread.
That cost is now clearly worth paying, and every other piece is already built to receive it:
- `ClassifyTile` consumes `ClassifyBox` in production, so a proved tile skips `GenerateMesh`
30 000+ density evaluations saved against one `BuildChunkCache`;
- the fold carries **numbers** (`MaxCarveOverBox` / `MaxFillOverBox` / `ForcedMarginOverBox`), so a
bounded worm no longer kills `AllSolid` on rock that is solid by more than it can carve;
- the twelve detail modifiers already **inherit** the room source's verdict via `VF_NoCaveOverBox`
the day the source says `Identity` for a box, all twelve follow, in one place rather than thirteen.
**Keep the brute-force check.** `VoxelForge.OpStack.ClassifyTileSoundness` verifies verdicts against
`GetDensityAt` on a world where every strate opted in. A false verdict is an invisible hole: no
geometry, **no collision**, until a player falls through it.
## ⚠️ Debts that must be paid BEFORE that lands, not after
Both were introduced knowingly and are written at the exact site a reader would land on.
1. **Box bounds read STRATE params, but a per-room op can raise them.** `EffectOverBox` and the new
amplitude bounds are computed from strate params, because a box spans many rooms. But `ApplyTo`
writes the op's value **even where the strate's was 0**, so a room op can enable a modifier the
strate had switched off, or give it a bigger amplitude. A box verdict on a strate with a
terrain-op pool can therefore be **too optimistic** — the dangerous direction. Harmless while the
room source answers `Both` (nothing is provable anyway); **not harmless the moment it doesn't.**
Noted at `FLayerLineMod::EffectOverBox` and `FRoomGraphSource::LocalParams()`.
2. **`AUDIT §C2` is confirmed and unfixed on the `switch` path.** `GetGenerationParams` blends params
*within* a strate (`Alpha` depends on chunk Z for `Gradient`, and on chunk XY too for
`Interleaved`), and `Gradient` + `TransitionBlendChunks = 2` are the **defaults**. The original's
SDF cache key has neither params nor chunk Z, so a worker evaluates the second chunk it builds
against the first chunk's rooms — and *which* chunk came first depends on worker order, so two
peers can diverge from the same seed. The op stack does **not** inherit it (params CRC in the
key), and `ClassifyTile`'s new path guards against it explicitly (params must be bit-identical
across every chunk coord the box touches). The fix on the `switch` path is a params CRC in its
key — a live-generation change that wants a build in front of it.
## After that, in order
1. **PERF — unparked.** The op path is measurably slower. One cause found and fixed (the column memo
discarded itself every chunk). Remaining suspects in order: the hashed column lookup vs
`GSurfColCache`'s direct-indexed box, then per-voxel virtual dispatch. Also measured and stated:
the gate is now tested twelve times per voxel instead of once (stage B5's deliberate trade).
**Measure before optimising** — that is the §C10 lesson.
2. **`VerticalShafts` proves 0 of 60 tiles.** Pessimistic, not wrong: `EffectOverBox` returns
`CarveOnly` whenever any shaft is within a `Spacing*1.6` halo instead of testing real connector
capsules. Lost CPU, never a hole.
## ⚠️ IF YOU ARE RUNNING UNATTENDED — read this section before starting
Jahni may start this session and leave, then have you shut the PC down when you're done.
**You cannot build, and nobody will build for you during this run.** Rule #1 still holds absolutely.
Everything you write stays unverified until he returns.
**That is his explicit, repeated decision, not an oversight — do not re-litigate it and do not
quietly scale the work down.** Verbatim: *"I'm fine with unverified work, if committed, I can always
manually go back and try — I want him to work towards the full op stack completeness."* So the goal
of an unattended run is **to finish the transition**, not to stop somewhere safe. Take the queue
below as far as it goes.
Two things make that reasonable rather than reckless, and they are worth knowing:
1. **Committing per group IS the safety net.** He reverts by sha. That is why the one-commit-per-group
rule is the one thing not to bend.
2. **Reaching 8 of 8 does not change his world by itself.** `UsesOperatorStackForChunk` returns false
unless a strate has `bUseOperatorStack` ticked, so adding `TunnelNetwork` and `Underwater` to the
ported list is inert until *he* ticks a box. The flag flip is safe to include; pointing a live
world at it stays his call.
### The rules that replace "build after each step"
1. **One commit per group below. Never bundle two.** Jahni reverts by group, and a failed build must
bisect to a group rather than to 700 lines. Name the group explicitly in the subject line.
2. **Every commit leaves the tree coherent.** No half-written op, no dangling brace, no factory
declared without a definition. Assume the next thing that happens is a compile.
3. **Transcribe literally. Do not improve anything.** With no test feedback, a "cleanup" is an
unfalsifiable guess. Keep the original's operation order, its constants, its `FVector`
round-trips, its odd `FMath::Max` floors. If something looks wrong, **write it down in the
progress log and port it as-is** — that is how `§C1` and the `MinDivisor` split were both found
without breaking anything.
4. **Update the test in the same commit as the code it covers**, including its coverage counters. Do
not leave "I'll add the test after" — after is a build away.
### The work queue, in order. Run it to the end — **8 of 8 is the finish line.**
- **B1 — Surface roughness** (`STEP 4b`, ~130 lines). Density-space variant: a *different op* from
`FSdfRoughnessMod` (quadratic fade, anti-fill clamp, 4 noise types). See `DECOMPOSITION §1`.
- **B2 — Terrace · LayerLines · Ribbing** (~150 lines). The three "sedimentary" remaps.
- **B3 — Overhang · Cliff · Scallop · Arch** (~210 lines).
- **B4 — Columns · Domes · Pinch · FloorBias** (~190 lines). Columns read `SDFCache.Columns`, so this
group needs a Column op in the test's room pool — and the stage-A guard that *errors* on
`TotalColumns > 0` must move to this commit, not before it.
- **B5 — the gate itself.** All of B1B4 live inside `if (bNearCaveSurface)`, i.e.
`Sdf < SDFBlendRadius·3 && Sdf < FLT_MAX`. Decide deliberately whether that is one scoping wrapper
or a repeated early-out in each op, write down which and why, and make the test prove a voxel
*outside* the gate is untouched.
After each group, extend the test: turn that group's amplitudes ON in `EnableTunnelFeatures`, add a
coverage counter proving the group actually fired (**not** just that its param is non-zero — see the
pit/chimney lesson below), and move the op-count assertion.
**⚠️ The subtlety that lets B precede C.** Inside `if (bNearCaveSurface)` the original *shadows*
`Params` with `LocalTerrainParams` — a copy with the nearest room's terrain op applied — and **all 13
modifiers read the shadowed copy**. Porting them against strate-level params is therefore equivalent
*only while no room carries a detail-type op*. The test's pool is Pit/Chimney only and `ApplyTo(Pit)`
writes just the four pit fields, so the two are identical today. **Keep the pool that way through
B1B5**; C1 is what makes a `Terrace` op in the pool legal. (B4's Column op is safe: `STEP 4d` reads
`SDFCache.Columns` from the bake, never the shadowed params.)
- **C1 — the per-room op override.** `§2` calls this the piece with "no clean home" and offers two
options. **Take option (a), and note that the mechanism is already established in this codebase —
do not invent a scoping predicate.** Three ops already read state owned by an upstream op via a
non-owning pointer handed over at build time: `FOverhangShelfMod``FSurfaceColumnSource`,
`FShaftLedgeMod``FShaftFieldSource`. Do the same here:
`FRoomGraphSource` gains `const FStrateGenerationParams& LocalParamsAt(x, y, z) const`, which
reproduces the original exactly — take the voxel's `NearestRoomIdx` (it already computes it), apply
that room's `RoomOp` onto a `thread_local` copy of the params, return it. Every detail op holds a
`const FRoomGraphSource*` and reads its fields from that copy instead of from its own.
This is the pit/chimney resolution a second time: the difficulty in `§2` came from assuming each
modifier must own its params. Let one op own the shared state and the rest read it, and the problem
disappears. **Transcribe the per-voxel copy as-is** — yes, the original copies a ~74-field struct
per voxel inside the gate; note it in the log as a perf item and change nothing.
Acceptance: put a `Terrace` op in the test's room pool alongside Pit/Chimney and require
bit-identity. That is the check B could not make.
- **C2 — `Underwater`.** Nearly free: `§8` establishes there is **no density difference at all**
`GetDensityAt` routes it to `GetDensityWithParams` and `WaterLevelRelative` is consumed by the
render-side water system, not by density. So it is the TunnelNetwork stack, reached from a second
`case`. Add the wiring case and an equivalence test that samples the `Underwater` slot. If you find
a real density difference, **stop and write it down** — that would contradict `§8` and is worth
more as a finding than as code.
- **C3 — 8 of 8.** Add `TunnelNetwork` and `Underwater` to `UsesOperatorStackForChunk`'s ported list.
Update `CODEMAP §3.2d` and its `UsesOperatorStackForChunk` row, tick `OPSTACK-PLAN` Phase 2, and
write the closing `OPSTACK-PROGRESS` entry: the archetype `switch` now has a complete operator-stack
twin, per-strate opt-in, every one equivalence-tested.
### After 8 of 8 — keep going, in this order
1. **Settle `AUDIT §C2`'s suspected item.** Does `GetGenerationParams` actually vary within one strate
across a Gradient transition? Answerable by *reading* `GetGenerationParams` and
`FStrateGenerationParams::Lerp` — no build. Turn the suspicion into a yes or a no with the
evidence, and correct the audit either way. Cheap, and it closes an open question honestly.
2. **The worm amplitude cap (`DECOMPOSITION §0.2`).** The largest single perf item in the plan, and
the reason TunnelNetwork proves 0 of 40 tiles. It needs the fold to carry a **number**, not just a
direction: add a numeric carve/fill amplitude alongside `EVoxelOpEffect`, have
`FWormFieldSource::MaxCarveAmplitude()` (already written, already unused) feed it, and let
`AllSolid` survive when the rock is solid by more than the sum of every remaining carve.
**⚠️ This changes the fold contract that all 13 tests rest on — its own commit, nothing else in
it,** and `VoxelForge.OpStack.BoxVerdictFold` must be extended in the same commit.
3. **Make `ClassifyTile` consume `ClassifyBox`.** Verified per archetype, still unconsumed in
production — this is where measured tile-skipping (Maze 23/60, slabs 3640/60) becomes frames.
A false verdict is an invisible hole, so keep the brute-force check in every archetype's test.
4. **Re-read your own diffs against the original**, group by group, as a reviewer rather than an
author. On a run with no builds, a second reading is the only oracle you have. Do this rather than
starting anything new when the queue runs out.
### Still not yours to do, even unattended
- **Do NOT build, do not launch the editor, do not push.** Unchanged, not negotiable.
- **Do NOT tick `bUseOperatorStack` on any strate asset.** C3 makes the archetypes *available*;
pointing a live world at an unverified path stays Jahni's decision.
- **Do NOT start Phase 3** (ops as data assets). It is a design conversation, not a transcription.
### The commit messages ARE the ledger
**Jahni's words: "just commit with explicit name what changed, I'll revert if needed."** So there is
no separate bookkeeping — the commit history is the record, and it has one job: let him revert a
single group without touching the others.
That makes the subject line load-bearing. Name the group and the scope, not the intent:
```
feat(opstack B2): port Terrace, LayerLines, Ribbing detail modifiers
feat(opstack B4): port Columns, Domes, Pinch, FloorBias + column op in the test pool
feat(opstack C1): per-room op override — detail ops read LocalParamsAt from the room source
feat(opstack C3): TunnelNetwork + Underwater in the ported list — 8 of 8
```
Body: what was transcribed, anything that looked wrong and was ported as-is anyway, and what breaks
first if the group is wrong. One group per commit — that rule exists *because* of the revert
workflow, so it is the one thing not to bend.
**Close the run with one `OPSTACK-PROGRESS` entry** listing every commit made, in order, with a
one-line "what this touches" each, and the test filter to run (`VoxelForge`). Jahni's first action on
return is a single build; that entry is what makes it efficient. Then shut down if he asked you to.
3. **`AUDIT §C9` library half** — `sinf`/`cosf` are not IEEE-754 specified, so MSVC's CRT and glibc's
libm can differ. Currently **0 samples within 1e-6 of the isosurface**, i.e. no measured risk. Run
`CrossPlatformDigest` on Linux, compare the SHAPE digest, pin it. The real fix if ever needed is a
deterministic in-house sin/cos.
4. **Phase 3 — ops as data assets.** A design conversation, not a transcription. Don't start it
unprompted. What makes it possible is already in place: ops depend on capabilities
(`IVoxelBiomeField`), never on `UVoxelGenerator`.
## Hard rules that prevent real bugs
- **Density sign:** negative = solid at the mesher. Inside the op stack the convention is INTERNAL
(**positive = solid**), negated once by the caller. The SDF channel uses standard SDF convention.
- **Never run both density paths in one world.**
- **The acceptance bar is `§2.6.1`:** *same seed ⇒ same world on every peer*. Resemblance to the
pre-refactor world is **not** required. The equivalence tests are **port-correctness oracles**, not
fidelity checks — keep them for that reason.
- **Every cache key includes `LayoutVersion` AND the params.** See `§C2` and the overhang regression
of 2026-07-27, where omitting the params silently deleted the overhang and only 1 sample in 20 000
- **Never run both density paths in one world.** **Comparing them is legitimate** — §C10 is closed
since `FPSemantics = Precise`, and all eight equivalence tests compare bit for bit. They are
**port-correctness oracles**, not fidelity checks: §2.6.1 requires *same seed ⇒ same world on every
peer*, not resemblance to the pre-refactor world.
- **Every cache key includes `LayoutVersion` AND the params.** See §C2 and the overhang regression of
2026-07-27, where omitting the params silently deleted the overhang and only 1 sample in 20 000
crossed the isosurface.
- `ProcessQueue` stays `EQueueMode::Mpsc`; `Epoch` carries through every async path; don't "optimize"
the `ARCHITECTURE §8.10` invariants.
@@ -268,7 +149,7 @@ return is a single build; that entry is what makes it efficient. Then shut down
- Update `CODEMAP §3`, `ARCHITECTURE §8`, tick `OPSTACK-PLAN`, append to `OPSTACK-PROGRESS.md`.
- **When inserting a class into `VoxelDensityOpStack.cpp` / `VoxelHeightOpStack.cpp`, put it ABOVE
the labelled end of the anonymous namespace.** Anchoring on the FACTORIES banner puts it outside,
and the brace added with it closes nothing. Made that mistake twice; both files now say so at the
and the brace added with it closes nothing. Made that mistake twice; both files say so at the
exact line.
## Method lessons this refactor actually paid for
@@ -278,28 +159,33 @@ Ordered by how much they cost.
- **Instrument before hypothesising.** §C10 cost six builds and five refuted hypotheses, then was
solved for free by a build setting changed for an unrelated reason. Park a question whose
consequences are measured and benign.
- **Verify the premise before reasoning from it.** Four times now, a confident chain rested on an
- **Verify the premise before reasoning from it.** Five times now a confident chain rested on an
unchecked assumption and the check reversed it: C1's *documented* fix was wrong; "C9's risk is gone
after FPSemantics" was wrong; "C1 is closed, 0 sites left behind" was wrong (the sweep matched a
*spelling*, `SeedF * K`, and the survivor spelled it `(float)S * K`); "PitDensity enables pits" was
wrong. **A grep over a spelling is evidence about the spelling.**
*spelling*); "PitDensity enables pits" was wrong; "there are 13 detail modifiers" was wrong (twelve,
and only eleven read the per-room copy). **A grep over a spelling is evidence about the spelling.**
- **Read the code, not the comment.** The cliff modifier's comment promises a sampled Z±1 gradient;
the code samples nothing and uses a Z-stretched Perlin it *calls* `VertGrad`. Ported as written —
and written down, so nobody "fixes" it from the comment.
- **A perf change can be a correctness change.** The column-memo optimisation silently deleted the
overhang; the tests caught it the same day. The failure was invisible to inspection and produced
plausible terrain.
- **Coverage is a number, not a boolean.** Three related traps, each of which produced a green run
overhang; the tests caught it the same day. Invisible to inspection, and it produced plausible
terrain.
- **Coverage is a number, not a boolean.** Four related traps, each of which produced a green run
that proved almost nothing:
- *A test that prints nothing on success is indistinguishable from one that never ran.* Report
counts, not just failures.
- *A guard that only trips at zero does not measure coverage, it notices absence.* A run with 65 of
6000 samples in open cave (1.1 %) passed a `== 0` guard silently. Use fractions.
- *A test that prints nothing on success is indistinguishable from one that never ran.*
- *A guard that only trips at zero notices absence, it does not measure coverage.* Use fractions.
- *A success message that **asserts** coverage instead of reporting it reads as evidence while
measuring nothing.* One said "pits and chimneys exercised" through a whole run in which zero pits
existed.
measuring nothing.*
- *A check can be vacuous as well as a counter.* "Nothing leaked" is worthless unless something
happened — so the gate check also reports how many samples move when the modifiers are zeroed.
- **Enabling a feature is not evidence it fired — ask the structure, not the output.** Setting
`PitDensity` did nothing (wrong struct). Diffing two stacks with/without the op pool would have
*lied* (the pool is not in the SDF cache key, so both share the `thread_local` cache). What worked:
call `BuildChunkCache` and look at `Pits.Num()`. Prefer the check that can fail for exactly one
reason.
- **An oracle that shares the defect under test proves nothing.** The stale-cache check was going to
compare two interleaved param sets against the original — which keys its SDF cache without the
params and would have failed. Compare against the thing itself evaluated alone instead.
call `BuildChunkCache` and look at `Pits.Num()`. **Prefer the check that can fail for exactly one
reason** — and when a zero has three possible causes, give each one its own number.
- **An oracle that shares the defect under test proves nothing.** The stale-cache check compares each
stack against *itself evaluated alone*, never against the original — which keys its SDF cache
without the params and would fail it.
- **One definition, not two kept in sync.** `VF_BuildOpStackForChunk` exists because a tile skipped on
the verdict of a stack that is not the one producing its density is a hole. A "keep these in sync"
comment would not have been enough.
+5 -3
View File
@@ -10,9 +10,11 @@
> FloatingIslands · TunnelNetwork · Underwater.** The archetype `switch` now has a complete
> operator-stack twin, opt-in per strate.
>
> ⚠️ **The last three groups (stage B, C1, C2, C3) are UNVERIFIED — written, committed, never
> compiled.** Run the `VoxelForge` filter; `OPSTACK-PROGRESS.md`'s last entry lists every commit and
> what breaks first if one is wrong.
> **BUILT AND GREEN, 2026-07-28 — 14 tests.** TunnelNetwork A+B bit-identical over 6000 samples
> with all twelve group-coverage probes non-zero, all four noise branches covered, 0 gate leaks, and
> C1 proved by 10 Terrace-op rooms containing 1119 samples. One open warning: the `Underwater` check
> landed 0 samples in open cave, so its bit-identity proves little — diagnosed, not guessed, in
> `OPSTACK-PROGRESS.md`'s last entry.
>
> **Not done, and it is the next real prize:** `ClassifyTile` still uses hand-written guards and does
> not consume `ClassifyBox`. That is where measured tile-skipping becomes frames.
+72
View File
@@ -2395,3 +2395,75 @@ difference *is* the T1.d gain.
After that: make `FRoomGraphSource::EffectOverBox` answer spatially. Everything else is now waiting
on it, and the two commits above were built to receive it.
---
## 2026-07-28 — **IT ALL COMPILES AND IT IS ALL GREEN.** 14 tests. 8 of 8, verified.
Jahni built it. Everything written across the unattended run is now verified except one number.
```
TunnelNetwork STAGE A+B ... bit-identical, 6000 samples (1002 open cave, 4699 rock)
Cave coverage ............ 16.7 % (floor 10 %)
Group coverage ........... roughness 64.8 % · terrace 60.2 % · layer lines 25.7 %
ribbing 13.4 % · overhang 39.4 % · cliff 26.2 %
scallop 1.6 % · arches 3.4 % · domes 1.6 % · pinch 1.0 %
floor bias 18.2 % ← ALL TWELVE NON-ZERO
Roughness noise sweep .... 8/8 variants bit-identical over 1500 samples
Gate check (B5) .......... 979 outside the gate, 0 leaked; 4880 inside move
Params fingerprint ....... 395/400 = 98.8 % differ, 0 served wrong
Bake coverage ............ 49 rooms, 26 pits, 30 chimneys, 39 columns
Per-room override (C1) ... 51/51 rooms carry an op, 10 Terrace, 1119 samples inside them
Box verdicts ............. 0 proved, 40 Mixed (asserted)
Underwater (C2) .......... bit-identical, 2000 samples — 0 in open cave ⚠️
```
**Every claim this run made about correctness now has a measurement behind it.** The twelve
modifiers are transcribed correctly *and* each demonstrably fired; the four noise branches are
covered, not just the one; the gate suppresses everything outside it with zero leaks; and C1 is
proved by the only check that could prove it — 10 rooms drawing a `Terrace` op whose params
overwrite the strate's, with 1119 samples inside them, still bit-identical.
### The one number that is a warning, and it is the test working
`Underwater: 0 of 2000 samples in open cave.` **A green bit-identity over 2000 samples of solid rock
is not evidence** — it is exactly what two agreeing voids look like. Same failure as stage A's 1.1 %
run, in a different slot, caught by a counter written for precisely that.
**A real bug surfaced while diagnosing it.** The sampled chunk-Z range used `Z / CHUNK_SIZE`, and
C++ integer division **truncates toward zero**. TunnelNetwork sits at the top of the layout in
positive Z, where truncation and floor agree — so the bug could not show there. Underwater sits at
the **bottom, in negative Z**: `-1 / 32` is `0` truncated and `-1` floored, so the upper chunk bound
starts a notch too high and the `Clamp` piles the excess onto the strate's last voxel, inside the
top seal band, i.e. solid rock. Fixed in both point builders (`FloorDivChunk`). Same family as the
`DivideAndRoundDown` lesson already in the project notes.
That is a **candidate**, not a conclusion, so the fix does not stand alone: new **check 5b** gives
each of the three possible causes its own number — rooms baked for the Underwater strate index (the
bake), samples landing inside the seal-free interior (the Z range), and both non-zero (the XY
spread) — and prints how to read them. Sampling also widens 8 → 24 clusters. **Unverified: this
commit has not been built.**
### What the green run settles, beyond the tests
- `AUDIT §C10` stays closed: eight archetypes now compare **bit for bit** against their originals
across every detail path, including four noise types and a domain warp.
- The `FRoomGraphSource::FState` hoist, the lazy `LocalParams()` memo and the whole C1 mechanism are
compiled and exercised — the pattern "one op owns the state, the rest read it" now has three
users, all green.
- The extraction of `VF_BuildOpStackForChunk` did not disturb `GetDensityAt`: the equivalence tests
that route through it are unchanged.
- `0 proved, 40 Mixed` remains **asserted** for TunnelNetwork. That is not a defect and not stale:
the chain still dies at `FRoomGraphSource::EffectOverBox`, which is the next piece of work.
### Next single action
Re-run the filter and read **one line**: `Underwater diagnosis`. It names which of the three causes
produced the 0 %, and the fix follows from that rather than from a guess.
Then the one thing everything else is now waiting on: **make `FRoomGraphSource::EffectOverBox`
answer spatially.** Its room and tunnel bounds are already in the SDF cache; what it costs is
building that cache for the queried box, and that is now clearly worth it — a skipped tile saves
30 000+ density evaluations, `ClassifyTile` consumes `ClassifyBox` in production, the numeric
amplitude fold is in place, and the twelve modifiers already inherit the source's verdict. Every
piece is built to receive it and nothing else moves until it lands.