wip(assembler): round L7 -- shared column cache across workers: whole view 16-19% faster, world unchanged

SharedColumnCache.h: a world-owned, bounded 4-way second-tier cache of exact column values
(heightfield and floating islands), keyed by recipe/plan identity plus content hash, XY and island
id. Lock-free seqlock reads; set-local spin lock on misses. The FLazy thread-local tier stays
first. 83.6% of repeated column work was cross-worker; formula evaluations -60.9%.
W12 L1 view: Luna 2.165 -> 1.754 s; reviewer interleaved 2.059 -> 1.722 s. 8 m crossing new tiles
82 -> 52 ms. +66 MiB peak RSS. All hashes (whole view, per-LOD, exact queries, islands seed 0)
identical at W1/W8/W12, Clang off and GCC default. MSVC/UBT not compiled here.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CSFYgsdW8ejZSyiMf4qRSX

Standalone/Research/WorldAssemblerM3/Linux7/REPORT.md

@@ -0,0 +1,171 @@
+# Linux7 — shared terrain columns
+
+**Decision: keep the shared second tier.** On the L1 view it cut median W12 whole-view time by
+19.0%, well above the 3% threshold, and the 8 m camera crossing was also faster. All final output
+hashes match the `9898e05` binary.
+
+Runs used the `linux6` base, Clang 19 with FP contraction off, GCC 14 defaults, CPUs 0–11, normal
+priority, a 6 GiB address-space cap, and the unchanged `powersave` governor. Static timings below
+are three alternating base/candidate W12 runs of 2,553 tiles.
+
+## 1. What was recomputed
+
+Three complete W12 captures per mode on the L1 view. Counts below are medians of actual column
+formula executions after worker-local cache misses, grouped by LOD. “Same worker” means the worker
+had computed that XY before and later missed its local entry. Halo/normal repeats are separated
+before worker attribution; the parenthesized value is the total computed halo/normal probes,
+including unique ones. Each field's median is independent, so row medians can differ slightly from
+their sums.
+
+| LOD | Base total / unique XY | Same worker | Different worker | Halo repeats / probes | Shared total / unique XY |
+|---:|---:|---:|---:|---:|---:|
+| 0 | 556,134 / 128,240 | 22,307 | 358,908 | 46,679 / 70,972 | 135,720 / 127,370 |
+| 1 | 452,626 / 104,284 | 16,393 | 287,791 | 39,245 / 59,082 | 85,658 / 85,479 |
+| 2 | 144,251 / 32,274 | 14,667 | 74,377 | 22,933 / 30,942 | 28,019 / 28,019 |
+| 3 | 91,158 / 33,256 | 0 | 46,387 | 11,564 / 20,339 | 16,213 / 16,213 |
+| 4 | 272,776 / 91,325 | 100 | 167,083 | 14,266 / 28,089 | 90,105 / 90,065 |
+| 5 | 178,511 / 71,957 | 1 | 98,687 | 7,892 / 20,042 | 70,932 / 70,930 |
+| 6 | 342,763 / 280,334 | 20 | 56,204 | 6,205 / 19,977 | 279,266 / 279,202 |
+| 7 | 79,400 / 71,468 | 1 | 4,951 | 2,980 / 14,329 | 70,492 / 70,492 |
+| 8 | 137,062 / 106,219 | 0 | 22,138 | 8,704 / 22,609 | 106,219 / 106,219 |
+
+Across the per-LOD medians, the base made 2.255 million formula evaluations: 919 thousand unique
+XYs and 1.336 million repeats. Of the repeats, 83.6% were from another worker, 4.0% were same-worker
+evictions, and 12.0% were halo/normal probes. The shared candidate made 883 thousand evaluations,
+60.9% fewer; about 8.6 thousand repeats remained. This is mostly a worker-sharing problem, not a
+local-cache-capacity problem.
+
+## 2. Design
+
+`SharedColumnCache.h` is a bounded, four-way second-tier value cache used by both `Heightfield2D`
+and FloatingIslands. Its key includes recipe and plan pointer identity plus content hash, exact XY,
+and an optional exact sub-domain identity; FloatingIslands supplies the island stable ID because
+different anchors can use the same XY. LOD is omitted because the column value is LOD-independent.
+The heightfield table has 524,288 slots and the island table has 65,536. Both are allocated with
+the `World::Impl` and destroyed with it. The Standalone streamer stops and joins workers before
+world teardown; engine shutdown must keep the same drain-before-destroy ordering as the existing
+generator state.
+
+The S14 `FLazy` thread-local caches stay first, so their hits remain worker-private. Shared hits read
+versioned atomic entries without taking a lock; a miss takes a set-local spin lock, rechecks, then
+computes and publishes. A hit racing with a slot replacement can fall through as a miss. Fixed
+capacity bounds memory and replacement bounds reuse as the player moves. No page ring, prefetch, or
+neighbor affinity was needed. The shared table is not an immutable page store; atomic fields and
+version checks give readers a consistent value while a slot can be replaced. Workers only read
+recipe/plan identity and content hashes; they do not mutate generator state or call UObject/engine
+APIs. Epoch, shutdown, and MPSC result handling were not changed. The L3 FP-contract preamble and
+self-test remain intact, as does the `FLazy` rule for large thread-local storage.
+
+## 3. Speed and memory
+
+| W12 static whole view | Base | Shared | Change |
+|---|---:|---:|---:|
+| `whole_view_ready` median | 2.165 s | 1.754 s | −19.0% |
+| Peak RSS median (max) | 401.5 MiB (402.6) | 467.8 MiB (468.0) | +66.3 MiB median |
+
+Per-LOD worker sum is the sum of tile worker seconds, not wall time. Values are medians across the
+same three static runs.
+
+| LOD | Base worker sum | Shared worker sum | Change |
+|---:|---:|---:|---:|
+| 0 | 3.830 s | 2.667 s | −30.4% |
+| 1 | 5.301 s | 4.085 s | −22.9% |
+| 2 | 1.296 s | 0.795 s | −38.7% |
+| 3 | 1.116 s | 0.725 s | −35.1% |
+| 4 | 2.692 s | 1.674 s | −37.8% |
+| 5 | 1.911 s | 1.459 s | −23.7% |
+| 6 | 2.903 s | 2.621 s | −9.7% |
+| 7 | 1.242 s | 1.227 s | −1.2% |
+| 8 | 2.318 s | 2.271 s | −2.0% |
+
+The moving trial warmed the first view, moved the camera 8 m, then generated 144 entering tiles
+(128 LOD0, 16 LOD1) on the same world and worker pool. It ran three times per cache mode.
+
+| Moving check | Shared | Thread-local | Change |
+|---|---:|---:|---:|
+| New-tile wall median | 52.39 ms | 82.00 ms | −36.1% |
+| LOD0 worker sum median | 403.44 ms | 693.67 ms | −41.8% |
+| LOD1 worker sum median | 138.34 ms | 182.29 ms | −24.1% |
+| Peak RSS median (max) | 466.1 MiB (466.9) | 466.1 MiB (467.6) | Same cache allocation in both modes |
+
+## 4. Hashes
+
+Every W1/W12 result matched the base for Clang FP contraction off and GCC default: full detail,
+whole view, each LOD, and exact queries. The L1 base hashes are:
+
+| Scope | Density | Mesh | Payload | Topology |
+|---|---|---|---|---|
+| Full detail, 960 tiles | `0x0ee0579bd15cbbe4` | `0x32e5d4964d610fee` | `0x1932b96ee4b1d910` | `0x69743c6bad3447dc` |
+| Whole view, 2,553 tiles | `0xc10697a8f48380c7` | `0xe75c9119a1498cd5` | `0x49a9a8d1835a6c8c` | `0xedcb23288b5e8724` |
+
+| LOD | Density | Mesh | Payload | Topology |
+|---:|---|---|---|---|
+| 0 | `0x769f2a706d585a72` | `0x33eeeb39fa443816` | `0x645621a1d968983f` | `0xc2457ed905036c9d` |
+| 1 | `0x69cc5a5c0a2c60af` | `0xb48335f5410c57d5` | `0xd15c5480162ab8ba` | `0xfeca1bd025823604` |
+| 2 | `0x85deed5683a77202` | `0x53469895987aeea2` | `0xe55f8ae7e1532033` | `0xd5feb12a0e85c09b` |
+| 3 | `0x6518caefb3be7e7f` | `0xec513c90b853a614` | `0xf62fa1137ff17f82` | `0xccf1dcf681965816` |
+| 4 | `0x62d4be96021992f4` | `0xf1e4c6efcc2615a9` | `0x9dc5f9c88be02f33` | `0xfe6f55125a07c3bf` |
+| 5 | `0x5e4a8c38b8cea6fc` | `0x08ee56998aaf21e5` | `0x541eea8558f26cc8` | `0x6539374a8b3336d8` |
+| 6 | `0xe036e95b9006945b` | `0x624a33f941a99b02` | `0x81b73e14f68bdf86` | `0x4de4c63ce1f80a56` |
+| 7 | `0x8ee358af56a30d8d` | `0x46ab3c6184048b7c` | `0x0a6164b7e68e80c5` | `0x0c26b85bb03bc831` |
+| 8 | `0xc3458e1cbdb9c471` | `0x46c03d0994d75192` | `0x8f89fc5bde12d230` | `0xd16dc0cf62031898` |
+
+Exact queries (9 probes): density and unclamped `0xe4d4a8eb95b35933`, detail
+`0xefcba20c0c70c415`, collision `0xe604823a249029bf`. The FloatingIslands seed-0 fixture also
+passes at Clang/GCC W1/W12 with recipe hash `0x39a24198ad8cf88f`; its whole-view hash matches the
+base: density `0xa8fc5f8f4eaf77a3`, mesh `0xfff7d8ac8923eb9a`, payload `0xd730c51d959c1c59`,
+topology `0xbb33ddb76cd386be`.
+
+## 5. Transfer estimate and Windows work
+
+The majority of the opportunity is cross-worker, so it does not disappear just because the 5800X3D
+has a larger L3. Same-worker evictions are only 4% of baseline repeats; that capacity-driven part
+may shrink. The candidate adds about 66 MiB peak RSS, and a 96 MiB L3 could keep much of the shared
+table resident, though other game data will compete for it. I expect the win to transfer and clear
+3%, likely landing in the mid-single to low-double digits; that percentage is an estimate until the
+5800X3D run.
+
+Windows to-do:
+
+1. Build the Core route with MSVC `/fp:precise`; keep the FP-contract self-test and check the
+   64-bit atomic slot path on the supported Windows target.
+2. Run the W1/W12 whole-view, full-detail, per-LOD, exact-query, and seed-0 FloatingIslands checks
+   against the base binary.
+3. Interleave three static W12 A/B runs on the 5800X3D, then repeat the 8 m camera crossing; record
+   peak RSS and the machine's unchanged governor.
+4. Exercise real streaming shutdown/profile changes and confirm workers drain before world/cache
+   destruction, with Epoch checks and MPSC result queues still intact.
+
+## 6. Failed and corrected runs
+
+- The first Clang W1 hash invocation omitted the explicit data path; it exited 2 after resolving
+  `Standalone/bin/data/owner_world.json`. The retry passed with
+  `Standalone/data/owner_world.json` supplied.
+- The first moving step was 32 m and failed the shape interval 160-million-node safety bound. The
+  benchmark was retried at 8 m, which completed three times in each cache mode.
+- The first shared FloatingIslands cache key omitted the island stable ID. The W12 fixture exposed
+  an LOD4 hash mismatch. Adding the exact sub-domain ID fixed it; final hashes match at W1/W12 for
+  both compilers.
+- An early fixture command passed `--seed=0`, which also overrode the surface-world seed. The final
+  fixture comparison omits that override; FloatingIslands itself still runs its seed-0 check.
+- No run exceeded 30 minutes or remained stuck for 15 minutes. All final Clang and GCC builds
+  succeeded with 12 jobs.
+
+## Owner note
+
+Linux measured a 19.0% W12 whole-view gain and a 36.1% gain on the 8 m crossing, at about 66 MiB
+more peak RSS. Please run the Windows hash matrix and static/moving A/B on the 5800X3D before using
+the laptop percentage as an owner-machine estimate.
+
+## Orchestrator review (Claude)
+
+- **Concurrency read:** a seqlock over all-atomic fields (seq_cst), with the version checked before
+  and after the key/value loads; readers fall back to a miss on any mismatch. Writers hold the set
+  spin lock and mark the version odd while writing. There is no non-atomic shared access, so a
+  reader can't return a torn value. The per-tile scoped `setThreadActive`/restore binds each thread
+  to its own world's table. Default on (`useSharedColumnCache = true`).
+- **Re-run by me** (W12, L1 view, interleaved): base `linux6` 2.014/2.059/2.092 s vs L7 `linux7shared`
+  1.657/1.730/1.729 s, then 1.718/1.705/1.722 s (median 2.059 → 1.722 s, -16%). The whole-view hash was
+  identical in all 8 L7 runs at W12, plus W1 (8.036 s) and W8 (1.799 s).
+- Luna's island-key bug was caught by the W12 hash fixture before review, so the determinism check
+  works as intended.