Fr0zkaandClaude Opus 5.5
a6cd7cf339
Refine x2 at LOD4-8 would save 13.6% of whole view but smooths far crests and opens holes (gate and captures agree); connector shortcuts open gaps. Default far output is unchanged; the code is on side branch wip/l15-far-budget. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01CSFYgsdW8ejZSyiMf4qRSX
196 lines
12 KiB
Markdown
196 lines
12 KiB
Markdown
# Linux15 report
|
||
|
||
## Result
|
||
|
||
Null result. No far-mesh shortcut passed the full 1 px, shading, swap, and topology gates, so no
|
||
recipe enables one by default. Evaluator version 6→7 adds the optional `far_display_mesh` declaration;
|
||
omitting it keeps the L14 output path. The controls are in the shared Core layer and do not touch
|
||
LOD0–2, exact queries, collision, edits, or server state.
|
||
|
||
Candidates measured:
|
||
|
||
- Refinement grid scale ×2 at LOD4, 5, 6, 7, and 8. The all-tile LOD4 and LOD5 runs reject; the
|
||
remaining per-LOD probes exceed 1 px or change shading, and the LOD6–8 candidates worsen at
|
||
least one adjacent swap.
|
||
- Reusing connector station radii leaves geometry p99 at 1.07 px but reaches 15.23 px maximum;
|
||
shading delta reaches 0.667 and the normal changes by 73.3°.
|
||
- Coarsening refinement at LOD4–8 is the largest measured saving, but exceeds the display budget.
|
||
Halving connector station density also saves time but opens gaps. Removing
|
||
any one ray direction also opens gaps, and the removed-panel control finds 99 new holes.
|
||
|
||
## Screen-space gate
|
||
|
||
The gate samples 65×65 surface grids across all 1,145 displayed LOD3–8 tiles (5,208,134 samples;
|
||
skirts excluded). Geometry is projected normal displacement against today's same-LOD mesh at its
|
||
closest display distance, using 90° horizontal FOV and 2,560 px (0.000613592 rad/px). Shading delta
|
||
is the maximum of fixed-light normalized Lambert change and material selector/blend-channel change;
|
||
the limit is 1/255. Each number below is p50/p99/max. Exact-path baseline values are all zero.
|
||
The gate rejects geometry max above 1 px, shading max above 1/255, any worse per-sample swap,
|
||
missing surface, or new hole/crack/floater. Its own validation fails if S13, stride-2, missing
|
||
surface, floating surface, missing fine-ring, or removed-panel controls are not rejected; the
|
||
panel control must expose at least one new connector hole.
|
||
|
||
| LOD | Tiles | Geometry px: baseline → refine×2 all | Shading delta: baseline → refine×2 all | Refine×2 holes / cracks / floating |
|
||
|---:|---:|---:|---:|---:|
|
||
| 3 | 384 | 0/0/0 → 0/0/0 | 0/0/0 → 0/0/0 | 0 / 0 / 0 |
|
||
| 4 | 320 | 0/0/0 → 0/2.370/606.562 | 0/0/0 → 0/0.023529/1 | 254 / 400 / 254 |
|
||
| 5 | 144 | 0/0/0 → 0/12.728/95.453 | 0/0/0 → 0/1/1 | 1,183 / 76 / 138 |
|
||
| 6 | 96 | 0/0/0 → 0.023/13.408/799.596 | 0/0/0 → 0/1/1 | 476 / 0 / 201 |
|
||
| 7 | 48 | 0/0/0 → 0/13.242/248.905 | 0/0/0 → 0/1/1 | 30 / 0 / 1 |
|
||
| 8 | 153 | 0/0/0 → 0.158/19.250/187.465 | 0/0/0 → 0.003922/1/1 | 45 / 0 / 946 |
|
||
|
||
The selected-patch swap report compares each baseline with refine×2 all. Values are p50/p99/max px.
|
||
|
||
| Swap (distance) | Before | After refine×2 all |
|
||
|---|---:|---:|
|
||
| 3→2 (96 m) | 0.611/3.767/6.045 | 0.611/3.767/6.045 |
|
||
| 4→3 (192 m) | 0.752/4.630/7.513 | 0.752/4.630/7.513 |
|
||
| 5→4 (384 m) | 0/9.580/31.783 | 2.435/32.321/86.615 |
|
||
| 6→5 (768 m) | 0/9.754/70.340 | 0.343/19.142/78.077 |
|
||
| 7→6 (1,536 m) | 0/9.040/42.355 | 0.200/15.109/50.994 |
|
||
| 8→7 (3,072 m) | 0.826/18.030/47.182 | 1.528/24.547/64.449 |
|
||
|
||
The station×2 LOD6 screen scan measures 0/2/813.571 px, shading 0/0/1, with 1,349 holes and 1,690
|
||
floaters. Its connector audit counts missing ray stations by stable tile/ray/station ID: holes 8→221
|
||
(216 new), floating classifications 464→405 (111 new IDs). In the refine×2 all row above, the
|
||
screen scan reports those all-tile holes, cracks, and floaters. Station×2 leaves all six selected-patch
|
||
swap p50/p99/max rows at baseline values, including 6→5 at 0/9.754/70.340 px. All four ray directions are required:
|
||
the four one-ray-removed
|
||
controls create 1,008, 1,022, 1,171, and 1,158 new holes. The explicit panel-removal control creates
|
||
99 new holes. The exact-path audit has zero new holes, cracks, or floating pieces; its raw connector
|
||
audit counts remain 8 holes and 464 pre-existing floating classifications.
|
||
|
||
The LOD4×2 and LOD5×2 candidates also fail in isolation. Reusing station radii misses the LOD6
|
||
geometry and shading limits. The four one-ray-removed masks add 1,008, 1,022, 1,171, and 1,158
|
||
connector holes, respectively.
|
||
|
||
## Negative controls
|
||
|
||
- The S13 global-octave-drop and stride-2 column-interpolation controls are rejected at every LOD3–8.
|
||
- Missing displayed surface, a floating surface, and missing fine-ring surface controls are rejected.
|
||
- Removing connector panels is rejected with 99 new holes. The exact connector path has zero new
|
||
station IDs classified as holes, cracks, or floaters.
|
||
|
||
## Speed
|
||
|
||
Five interleaved W12 pairs per candidate, hashing off, same 2,553-tile L1 view.
|
||
`whole_view_ready` is the metric. Both candidates fail the quality gate.
|
||
|
||
Station×2 connector stations:
|
||
|
||
| Pair | L14 baseline | Station×2 | Change |
|
||
|---:|---:|---:|---:|
|
||
| 1 | 1.320 s | 1.304 s | −1.2% |
|
||
| 2 | 1.363 s | 1.330 s | −2.4% |
|
||
| 3 | 1.371 s | 1.332 s | −2.8% |
|
||
| 4 | 1.393 s | 1.341 s | −3.7% |
|
||
| 5 | 1.391 s | 1.345 s | −3.3% |
|
||
| Median | 1.371 s | 1.332 s | −2.8% (3.3% median paired gain) |
|
||
|
||
Refinement grid scale ×2 at LOD4–8:
|
||
|
||
| Pair | L14 baseline | Refine×2 all | Change |
|
||
|---:|---:|---:|---:|
|
||
| 1 | 1.305 s | 1.164 s | −10.8% |
|
||
| 2 | 1.336 s | 1.177 s | −11.9% |
|
||
| 3 | 1.362 s | 1.169 s | −14.2% |
|
||
| 4 | 1.378 s | 1.178 s | −14.5% |
|
||
| 5 | 1.384 s | 1.190 s | −14.0% |
|
||
| Median | 1.362 s | 1.177 s | −13.6% (14.0% median paired gain) |
|
||
|
||
## Hashes and invariants
|
||
|
||
The undeclared/default recipe path matches L14. LOD0–2 density/mesh/payload/topology hashes:
|
||
|
||
| LOD | Density | Mesh | Payload | Topology |
|
||
|---:|---|---|---|---|
|
||
| 0 | `769f2a706d585a72` | `33eeeb39fa443816` | `645621a1d968983f` | `c2457ed905036c9d` |
|
||
| 1 | `69cc5a5c0a2c60af` | `b48335f5410c57d5` | `d15c5480162ab8ba` | `feca1bd025823604` |
|
||
| 2 | `85deed5683a77202` | `53469895987aeea2` | `e55f8ae7e1532033` | `d5feb12a0e85c09b` |
|
||
|
||
Exact queries (density / unclamped / detail / collision) remain
|
||
`e4d4a8eb95b35933 / e4d4a8eb95b35933 / efcba20c0c70c415 / e604823a249029bf`. The 24 m crossing's
|
||
432 moved tiles match L14 at W1/W12 with Clang off and GCC default:
|
||
`337ca62131576728 / 7612f48e2ddbbaab / 69a67ed4add83b87 / f11ac1ebd0f7ddb8` (D/M/P/T).
|
||
Its LOD0–2 per-LOD hashes also match the three rows above.
|
||
|
||
Determinism passed 28/28 configurations: SurfaceWorld seed 0 and FloatingIslands seed 0, Clang 19
|
||
FP-contract off and GCC 14 default, W1/W8/W12 and five W12 repeats. All per-LOD and whole-view
|
||
hashes match L14. L13 route parity passed for SurfaceWorld seeds 0, 1, and 17 and FloatingIslands
|
||
seed 0; page, native-only, higher-LOD, edit, and history mismatch counts are zero.
|
||
|
||
The L3 FP-contract preamble/self-test, L4 `FLazy`, L7 shared cache, L9/L12 proofs, L11 readiness and
|
||
crash fix, L13 route binding, and CLAUDE.md async rules remain in place. The evaluator and optional
|
||
recipe declaration are shared by all recipes. Builds and runs used CPUs 0–11, up to W12, normal
|
||
priority, 6 GiB address-space limits, and the unchanged `powersave` governor. Peak RSS was 904,400 KB
|
||
(GCC build); peak runtime RSS was 736,432 KB. No run exceeded the time or memory limits.
|
||
|
||
## Capture pairs
|
||
|
||
JPEGs are 1280×720 and each is under 74 KB. Before is L14; after is the rejected refinement×2
|
||
candidate applied at LOD4–8.
|
||
The L1 view uses `(0,-35,-487.94) m`, yaw 0°, pitch −22°. Horizon uses the same location at pitch −2°.
|
||
The mesa frame aims from the L1 location at the L6 mesa band (`(-472,-1432,-463.151) m` in the
|
||
gate) with yaw −161° and pitch 0.9°.
|
||
|
||
| View | Before | After refine×2 all |
|
||
|---|---|---|
|
||
| L1 camera | [l1-camera-before.jpg](captures/l1-camera-before.jpg) (59 KB) | [l1-camera-after-refineall.jpg](captures/l1-camera-after-refineall.jpg) (57 KB) |
|
||
| Horizon | [horizon-before.jpg](captures/horizon-before.jpg) (63 KB) | [horizon-after-refineall.jpg](captures/horizon-after-refineall.jpg) (60 KB) |
|
||
| LOD6 mesa rim | [lod6-mesa-rim-before.jpg](captures/lod6-mesa-rim-before.jpg) (74 KB) | [lod6-mesa-rim-after-refineall.jpg](captures/lod6-mesa-rim-after-refineall.jpg) (67 KB) |
|
||
|
||
## Failed and retried runs
|
||
|
||
- The first Clang build wrapper exited 127 because `/usr/bin/time` is absent; the direct Clang
|
||
`-ffp-contract=off` build succeeded. GCC 14 default also built successfully.
|
||
- A later sequential-build monitor was interrupted after Clang completed and before its GCC step
|
||
launched. Clang and GCC were rebuilt separately with the same compiler settings and 6 GiB cap.
|
||
- The first hash matrix omitted `--seed=0` and was discarded. A second FloatingIslands matrix used
|
||
the SurfaceWorld spawn camera; it was rerun with the L14 auto-surface camera. The corrected matrix
|
||
passed all 28 configurations.
|
||
- Initial SurfaceWorld route probes used no camera for seed 0 (the 2,370-tile bound), omitted
|
||
`--region-crossing` for seed 0 (shape compiler safety bound), and used the fixed seed-0 camera for
|
||
seed 1 (LOD2/exact-query parity mismatch; seed 17 was not run with that setup). They were rerun
|
||
with L14's camera selection and region-crossing option; all four final route audits passed. The
|
||
first 24 m crossing also omitted `--region-crossing` and exceeded the shape-build safety bound;
|
||
all four corrected compiler/worker combinations passed.
|
||
- The first connector audit compared aggregate counts. It was replaced with stable per-ray station
|
||
ID comparison before the final gate. The initial L4 selected-tile gate missed a changed tile; the
|
||
all-desired-tile coverage gate found and rejected it. The full-view gate is used in the table above.
|
||
- The first JPEG encoding pass did not set CPU affinity or an address-space cap. All six files were
|
||
re-encoded under CPUs 0–11 and the 6 GiB cap; sizes and pixels stayed the same.
|
||
- The first five speed pairs included screen-gate color capture on ordinary streaming tiles. That
|
||
diagnostic copy is now enabled only for gate runs; all five cost pairs were rerun and the table
|
||
above uses only the final logs. The final exact gate passed and station×2 remained rejected.
|
||
- No builds, captures, final audits, or cost runs timed out or breached their memory limits. The
|
||
intentional candidate rejections are listed above, not treated as infrastructure failures.
|
||
|
||
Evidence logs are in `Saved/Linux`: the final baseline and station×2 all-tile gates are
|
||
`l15-final-gatecolors-exact.log` and `l15-final-gatecolors-station2.log`; the best-payoff
|
||
refinement×2 all-tile gate is `l15-final-gatecolors-refineall.log`. Single-LOD refinement coverage is in
|
||
`l15-gate-alltiles-refine-l4x2.log` and `l15-gate-alltiles-refine-l5x2.log`. Single-LOD refinement
|
||
and connector-ray probes are `l15-gate-refine-l{x}x2.log`, `l15-gate-reuse.log`, and
|
||
`l15-gate-rays{14,13,11,7}.log`. Hash, route, and crossing evidence is in
|
||
`l15-final-hash-audit.txt`, `l15-final-route-*.log`, and `l15-final-crossing-24m-*.log`; final
|
||
station×2 speed pairs are `l15-final-speed-pair{1..5}-{before,after}.log`; refinement×2 pairs are
|
||
`l15-final-speed-refineall-pair{1..5}-{before,after}.log`.
|
||
|
||
## Note for the owner
|
||
|
||
The default far terrain stays the same because every measured reduction exceeded the 1 px geometry
|
||
or shading limit, worsened a swap, or opened holes. Refinement×2 across LOD4–8 saves 13.6% by the
|
||
median whole-view ratio, but creates hundreds-pixel errors and visible gaps, so the captures show it
|
||
for review only and it is not enabled.
|
||
|
||
## Orchestrator review (Claude)
|
||
|
||
- I looked at the LOD6 mesa-rim pair. After refine×2, the jagged crests are visibly smoothed, black
|
||
holes appear in the cap/ceiling, and a dark sliver shows on the left horizon. That confirms the
|
||
gate's rejection: the instrument is right this time, not the world.
|
||
- **Not merged into `assembler/world`:** the candidate knobs and the evaluator 6→7 bump add no default
|
||
output and would invalidate version stamps for nothing. The full L15 code (the extended gate with
|
||
shading, swap and connector-ID audits, and the candidate controls) is kept on side branch
|
||
`wip/l15-far-budget`. `assembler/world` stays at L14 code plus this report.
|
||
- Conclusion of L14+L15: far terrain has no exact or within-budget saving left with today's design. The
|
||
remaining far cost is the price of the current refinement and connector-repair geometry.
|