Started as an allocation pass over `StepCache` and grew into a cleanup of the surrounding pathing engine. Four commits, each independently reviewable; net **−560 lines**.
Build clean (0 warnings). All 122 `Server.Tests.Pathfinding` tests pass.
---
## 1. `perf`: pool the strata buffer, cut a hot-path dictionary lookup
**The headline is that `TryGetMask` — the actual hot path — was already allocation-free.** `StepMask` is a readonly struct, `StaticTileEnumerable` is a `ref struct`, `ChunkMissState` is a struct in a `Dictionary`. So most of this is a bake-throughput and GC-churn win, with one exception noted below.
`BuildChunk` accumulated packed multi-Z strata into a `List<byte>` that grew by doubling (256 → 512 → 1024 → …) and then paid a final `ToArray()`. A full map bake runs it ~114k times. It now writes into a `byte[]` rented from `STArrayPool<byte>.Shared` through a span writer, and hands the chunk one exact-size copy.
**This required fixing a latent out-of-bounds guard.** The record-fit check reserved headroom for **8** strata (`StratumByteLength * 8`) while `ComputeStandableSurfaceZs` can return up to **16** — so a cell could write 305 bytes starting from a 65,383-byte offset. Against a `List` that was benign (it just grew past 64 KB, and emitted offsets stayed under the `NoStrata` sentinel). Against a fixed-size rented buffer it is an out-of-bounds write, so tightening it was a *prerequisite* for the pooling, not a drive-by. The guard is now exact, which additionally proves no emitted offset can collide with `NoStrata == ushort.MaxValue`.
**One genuine query-path win:** `ShouldPromoteAfterMiss` did *two* dictionary lookups per miss — a `TryGetValue`, then an indexer assignment that re-hashes and re-probes. It now mutates in place via `CollectionsMarshal.GetValueRefOrNullRef`. This runs on every uncached chunk touch during A* expansion. The window-expiry branch keeps its explicit early return, so `MissPromotionThreshold == 1` still resets rather than promoting.
Also dropped `StepProbe.ComputeStrataAt` / `ComputedStratum` (dead code, zero callers) and collapsed six 18-argument `new StepMask(0, 0, …, kind)` blocks into `Fallthrough(kind)`.
**Considered and rejected:** pooling the `Direction[]` that `Find` returns. It *escapes* the call — `MovementPath` holds it across ticks while `PathFollower` walks `m_Index` through it — so it cannot be rented-and-returned, and it cannot be borrowed from the shared `BitmapAStarAlgorithm.Instance` without one creature clobbering another's in-flight path. `CheckPath` rate-limits repaths to one per 2s per creature, putting this at roughly 60 KB/sec at 1,000 pathing creatures. Not worth a public API break plus a use-after-return footgun.
## 2. `docs`: rewrite the comments for publication
The comments had accumulated as development notes: internal phase jargon (`Tier 4`, `the Phase-2 synthesizer`), change narration aimed at a reviewer (`which the old ComputeStandingZ anchor missed`, `legacy behavior`), benchmark anecdotes (`benchmarked as near-optimal`, `a ~20 ns lookup`), and paragraphs restating the code.
Rewritten to keep the rationale you cannot recover by reading the code — why the source-Z guard cannot be widened, why multis fall through with a halo, why the promotion gate counts Finds rather than calls, why `ComputeFingerprint` must hash the *files* and not the live tile tables — and drop the history that got us there.
Three comments were **factually wrong**, not just wordy:
- `CacheEvictionTimer` and `CacheStats` documented a class called `StaticWalkabilityCache`. No such class exists — it is `StepCache`.
- `StepCacheFile` declared `File layout v8` while `FormatVersion` is 9, and called the current record layout "the v6 layout" in four places. The layout descriptions are now unversioned so they cannot drift again.
- `StepProbe.ComputeStandingZ` claimed `StepCache` uses it to bake `SourceZ`. It has not since the baker moved to the clearance-aware `ComputeStandableSurfaceZs`; only a parity test calls it.
## 3. `refactor`: simplify `StepCacheFile.Write`, consolidate the format tests
`SaveToFile` walked `_keysList` **twice** — once to count the map's chunks, then again through a `ChunkEnumerator` closure to emit them — because `Write` needed the count up front to size its index array. Both loops had the same root cause. Passing a **span** collapses them: the count is just `span.Length`.
That deletes the `ChunkEnumerator` delegate, the closure over the list enumerator, and **both `InvalidOperationException` throws**, which existed only to police the delegate's "yield exactly `chunkCount` chunks" contract — a contract a span makes unrepresentable.
`Write` now patches the header's `IndexOffset` by seeking back to it rather than reaching into the writer's live buffer with `BinaryPrimitives`. That also retires `IndexOffsetFieldPosition`, a hand-maintained byte offset that had to track the header layout, and sidesteps the stale-array hazard that motivated the manual patch (`BufferWriter` reallocates on growth).
**Tests:** `StepCacheFileV6/V7/V8Tests` were named for the format version that introduced each transform — and the format is now **v9**, so all three names described formats the loader rejects outright. Beyond triplicated builders and plumbing, two things were actually broken:
- The three near-identical rejection tests each cited a `MinSupportedVersion` that had since moved (`"version 5 < MinSupportedVersion 6"`, `"6 < 7"`, `"7 < 8"`). They passed for the wrong reason.
- `AssertBaseEqual` (used by V7 and V8) **silently skipped the swim and strata trailers**. A regression dropping either would not have failed those tests.
Now one `StepCacheFileFormatTests`, named for behavior — predictive-Z elision, compression, compact index — with a single `AssertIdentical` that does check both trailers, the three rejection tests folded into one theory that also covers a future version, and a zero-chunk case the delegate-based writer never had coverage for.
## 4. `test`: consolidate the parity and lifecycle tests
Three files tested "parity" and none of the names said *which*. They were three different layers, and the seams are the useful part, so they are now one `StepCacheParityTests` that names them:
| Test | Compares | Answers |
|---|---|---|
| `ProbeMatchesSlowPath` | StepProbe vs MovementImpl | Is the bake right? |
| `CacheMatchesProbe` | StepCache vs StepProbe | Is it stored and returned intact? |
| `CacheServesReachableWalkStates` | StepCache vs MovementImpl | End to end, over the states A* visits |
Merging removed a duplicated stub `Mobile`, duplicated region seeds, and a filename/class mismatch (`StepProbeParityTests.cs` declared `StaticWalkabilityParityTests`). `SwimBake_ProducesWetCells` moved with it — it lived in the cache parity file but never touched the cache.
Tests reached into `StepCache._chunks` via `GetField` in **9 places**, each rebuilding the key encoding and cell-index arithmetic by hand. `StepCache` now exposes `GetResidentChunk` and `ResidentIndexInSync` alongside the internal test hooks it already had (`LazyReaderHasChunk`, `CurrentFindGeneration`), and the shared arithmetic moved to `PathingTestSupport`. All 9 reflection blocks are gone.
`StepCacheLifecycleTests` is regrouped by what it covers — promotion gate, fallthrough routes, strata, swim layer, eviction — with the `Tier4*` names dropped. Removed `Singleton_IsAvailable`, which asserted an inline-initialized static property was not null; that is the entire 123 → 122 test-count delta.
---
## Verification
Tests were mutation-checked rather than just run, since round-trip and parity tests can pass while a transform silently no-ops:
- Injecting an off-by-one into the `IndexOffset` patch fails **15 of 123** — the format tests are load-bearing.
- Offsetting the cache's cell index by one fails **7 of 10** parity cases, and the 3 that stay green are exactly the ones that do not touch the cache. The layering localizes a fault rather than just reporting one.
b852bca41e