passages: prepared per-segment data + per-column XY projection cache (L98, exact)

Per-world prepared chord/length/radius/Catmull-Rom/floor-profile data (310 segments, ~66 KB, <0.04 ms); same expression grouping, divisions and square roots, so output is bit-identical (zero density/tile mismatches, Clang vs GCC equal). No measurable readiness gain - the cost is the number of samples running passage work, not per-sample arithmetic. Note: the report's world labels are shifted by one row (the 2,850-tile row is SurfaceWorld). Fast tier PASS; Linux UE host builds. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016dhxM43dnWvUu2B5Ph1iG1
2026-10-28 12:50:19 +01:00
1 parent 9d9504ae4e
commit 36140b573f
@@ -0,0 +1,61 @@
+# L98 — exact passage building blocks
+
+## Changes and work counts
+
+Passage setup now prepares per-segment chord endpoints/deltas, squared lengths, radii, Catmull–Rom inputs, and floor-profile samples once per world. Capsule-route and mouth-support callbacks also use prepared segment endpoints and deltas. The scalar expression grouping, divisions, and square roots remain in the same order. Relief interpolation sharing was not added.
+
+Measured preparation per world: 310 passage segments, 313 floor samples, and 32 capsule-route segments; the trace reports 66,296 bytes. The 32 mouth-route segments report another 4,608 bytes. Preparation measured 0.025–0.040 ms in the whole-view hash runs (mouth preparation rounded to 0.000 ms).
+
+The streamed-view counters below are prepared-data reads/uses, not baseline arithmetic counts. They directly confirm the prepared paths were exercised. Avoided recomputations are inferred from the replaced per-sample expressions; the old arithmetic was not separately instrumented.
+
+| World | Floor-profile reads | Carve uses | Support/tread uses | Capsule uses | Mouth-support uses |
+|---|---:|---:|---:|---:|---:|
+| FI | 129,382,181 | 8,255,393 | 5,131,690 | 2,693,464 | 14,773,980 |
+| TN | 396,540,009 | 50,346,149 | 19,176,668 | 11,657,176 | 63,632,356 |
+| SW | 318,412,931 | 50,093,885 | 17,065,041 | 10,323,467 | 60,102,531 |
+| VS | 99,874,533 | 5,141,277 | 3,142,836 | 662,072 | 8,984,390 |
+
+The XY cache keeps the scalar numerator grouping `(qx-ax)*dx + (qy-ay)*dy`; clamp, winner, reach, and composition remain per lane. A direct probe exercised 310 z-runs × 8 samples in each world: 2,558 numerator computations and 31,085 cache hits, with zero density mismatches against scalar evaluation. The measured hit fraction was 92.4% of numerator requests. The ordinary streamed view does not call the z-run API, so these XY reuse counts come from the opt-in direct probe, not natural view traffic.
+
+## Exactness
+
+Compared rebuilt committed-HEAD baseline to L98 Clang, then L98 Clang to L98 GCC. Density tile hashes aggregate the density sample bit patterns. All comparisons had zero tile-hash mismatches at every LOD, zero LOD0–2 density-tile mismatches, and identical whole-view density, mesh, payload, and topology hashes.
+
+| World | Whole-view tiles | LOD0–2 density mismatches | Whole-view density | Mesh | Payload | Topology |
+|---|---:|---:|---|---|---|---|
+| FI | 2,850 | 0 | `40b76d0072a2adcd` | `8559df0850d4d7f1` | `7700965ff31ac702` | `0e7f2a16b4cc71a9` |
+| TN | 2,249 | 0 | `99ca213b21ce86e4` | `2977525405b79fff` | `3d7d2efb581710bd` | `af1916c949224871` |
+| SW | 2,233 | 0 | `8d23c26f5c092be1` | `15c3aad2e29d9ea0` | `3d48f68d511c21df` | `5b1584676884edd8` |
+| VS | 2,329 | 0 | `876b44981d106be3` | `aaef57cb08bd6845` | `b19617df9f4acbb1` | `e8eee991260adff4` |
+
+The direct XY probe also reported 2,480 sample comparisons and zero mismatches for each world. Hash runs and probe logs are under `Saved/Linux/l98/hashes/` and `Saved/Linux/l98/reuse-counts/`.
+
+## L77 passage-share census
+
+Mean near-LOD0–2 passage share, three profile runs per world and binary:
+
+| World | HEAD before | L98 after | Change |
+|---|---:|---:|---:|
+| FI | 57.82% | 56.74% | −1.08 pp |
+| TN | 55.94% | 55.70% | −0.24 pp |
+
+This is a census comparison, not evidence of a timing gain. Logs and raw run values are in `Saved/Linux/l98/census/`.
+
+## Cold readiness
+
+Median seconds from three interleaved HEAD/L98 pairs per cell. Each binary/stratum/worker-count cell had a warmup. Cold timing includes world/layout setup and passage preparation. The LOD0–2 and whole-view tile counts are included below.
+
+| World | LOD0–2 / whole tiles | W8 HEAD → L98 (LOD0–2 / whole s) | W12 HEAD → L98 (LOD0–2 / whole s) |
+|---|---:|---:|---:|
+| FI | 1,408 / 2,850 | 1.070 / 2.341 → 1.099 / 2.369 | 1.029 / 2.147 → 1.005 / 2.123 |
+| TN | 1,216 / 2,249 | 2.672 / 4.034 → 2.680 / 4.059 | 2.211 / 3.357 → 2.207 / 3.357 |
+| SW | 1,344 / 2,233 | 2.713 / 3.929 → 2.496 / 3.623 | 2.255 / 3.313 → 2.176 / 3.183 |
+| VS | 1,344 / 2,329 | 1.292 / 2.315 → 1.264 / 2.257 | 1.094 / 2.015 → 1.107 / 2.041 |
+
+All 48 measured readiness runs completed. Results are close and mixed across worlds; they do not show a broad readiness speedup. Full results and logs are in `Saved/Linux/l98/readiness/`.
+
+## Verification and failed attempts
+
+- Fast tier: PASS (269 pass checks, 5 diagnostic checks), 80 runs, 303.5 s. The first invocation stopped before checks because its output directory had no built headless executable; building it and rerunning passed. Summary: `Saved/Linux/l98/verify-fast/fast-summary.json`.
+- Linux UE build: `VFHostEditor Linux Development`, `-NoUBA -MaxParallelActions=4`, succeeded in 36.60 s after timing. The first attempt stopped before compilation when UBT could not rotate its default log in the read-only user config directory. Retry used an explicit log at `Saved/Linux/l98/unreal-build-ubt.log`; no `--as` limit was used.
+- No measured run exceeded the 30-minute limit. All experiment processes used `taskset -c 0-11`; headless runs also used the 6 GiB address-space limit.