Follow-up to #2721. It fixes how the remaining tests behave and changes no production code.
## CI
- Both `dotnet test` steps pass `--filter "Category!=Manual"`. Before this, Manual-tagged tests ran on CI anyway.
- The one surviving movement-throttle latency test is the only coverage of burst forgiveness over a full, wrapped history. It loses its Manual tag, so it runs on CI.
## Tests that passed on CI without testing anything
- **Corpse and throwing tests.** They relied on tiledata for the weapon `Layer`, so on CI the equip silently failed and they asserted against empty lists. They now set `Layer` explicitly and assert the equip took effect. `TestThrown` also pins its own strength and range values, so a shard retuning SoulGlaive no longer breaks them.
- **Healthbar poison rows.** Server.Tests registers no poisons, so every row tested `null`. They now use real poison levels.
- **Account packet tests.** The mock accounts reported `Length` 0, so the character-list tests only ever encoded empty slots. They now encode the real characters.
- **Pathing and AI tests that only mean something on real Trammel/Felucca terrain** are now guarded with `TileDataRequirement.SkipIfMissing()`: Britain desk and clutter scenes, house footprints (no MultiData on CI) and fingerprints. Tests that are valid on flat blank terrain stay unguarded. (A synthetic tiledata fixture to run these on CI is a planned follow-up.)
- **Assertions that couldn't fail now check their claim.** Affected tests include the `DefaultWalker` cache use, cache on/off parity requiring a path, the recorder capability flags, the idle-pet gate, the monthly DST recurrence, the AdvancedSearch `Record.Exception` wrappers, `CanFitItem` with no branch-guarded assert, the Firewall callback, and `Import_NoMap`. The Geometry Parse asserts also now use the same input as their TryParse.
## Shared state
Every test that changes process-wide state now restores it, even when an assert fails:
- **The clock.** A new `TimerWheelScope` saves and restores `Core._tickCount`, `Core._now` and the timer wheel. `SimulatedClockScope` covers decay tests.
- **StepCache settings and caches**, via `StepCacheStateScope`.
- **Network state.** The ManualAllowlist, LoginAllowlist, firewall and BanChannel reporters are restored. Global packet handlers are now registered per test and removed afterwards, so they no longer stay registered for later tests.
- **Data and config tables.** `ArchiveJournal` statics, `TileData.ItemTable`, `ObjectPropertyList.Enabled`, `Core.Expansion` and the localization table are restored. One `MobileStatusVersion` restore was resetting a different expansion than the one the test changed.
- **Entities.** Tests with fixed serials now delete what they create. Fixed mobile serials moved to `World.NewMobile`, because deleting a hard-coded serial could unregister a live entity.
- **Collection race.** `SpiralScanTests` no longer clears the spawn cache from a separate collection that could run in parallel with the cache-manager tests.
- **Shared statics.** Classes reading shared static allowlists joined the sequential collection.
## Duplication
New shared helpers replace copy-pasted harness code:
- **Server.Tests:** `PacketAssert.Sends`, used by about 130 packet tests; `TestClientHarness`, `PersistencePipelineScope`, `NetTestHelpers`, `EntityCleanup` and one `TestMulti`.
- **UOContent.Tests:** `TimerWheelScope`, which also holds one wraparound-safe `RunFor`/`RunUntil`; `AiTestStubs`, which merges six walker copies and five identical creature stubs; `MobileTestHarness`, `CommandTestHarness` and `TempDirectory`.
Dead oracle packet classes are removed.
## Cost
| Test | Before | After |
|---|---|---|
| `VirtualSerialTests` | 1.57s | 2ms (starts near the wrap via reflection instead of drawing ~17.9M serials) |
| Event scheduler recurrence loop | 473ms | 1ms |
| CrowdSec drain | 517ms | 16ms (event-driven instead of a fixed 500ms wait) |
| AuthId tests | 248ms | 99ms (SHA2 except the two GameLogin tests) |
| Container cap tests | 5,000 items | 3,449 / 3,277 (one past each protocol cap) |
| Bulk removal | 100k children | 20k children |
| OPL hash distribution | 60k builds | 10k builds |
## Shard-tuning decoupling
- AI assertions read `ReacquireDelay` and `FollowMoveSpeed` instead of hard-coding production tuning.
- The guard friendly-fire test uses an `AI_Mage` stub instead of WhiteWyrm.
- Argon2 assertions pin the `argon2id` variant and `v=19` but read the cost parameters from the configured hasher.
## Tests
| Run | Server.Tests | UOContent.Tests |
|---|---|---|
| With client files | 812 pass | 1,123 pass |
| Without client files (CI simulation) | 800 pass, 12 skip | 1,021 pass, 102 skip |
- The UOContent fixture reads `MODERNUO_TEST_DATA_DIR`; Server.Tests reads `MODERNUO_CLIENT_PATH`.
- The 15 new UOContent skips are the map-dependent guards above.
- Total run time with client files is about 5s, down from about 8s.
Reduces the world-save freeze window from ~740ms to ~78ms (measured on a synthetic 10M-entity / 1.7GB world, 24 cores, through the real pipeline classes) by removing the per-entity handoff between the game loop and the serialization workers, fixing how large indivisible payloads are scheduled, rewriting the BufferWriter hot path, removing per-entity placement state entirely, and finally replacing the global serialized-types tracking with a per-file type table (idx v4) that also shrinks idx files by ~21% and speeds the background write phase. The pipeline has also been validated end-to-end on live-copy worlds in the multi-million-entity range, where the freeze is drain-bound (real `Serialize()` costs far more CPU per byte than synthetic writes) — the same structural wins hold, and entity/file round-trips are byte-clean across both load paths.
## The problem
The freeze window is `max(main-thread handoff, slowest worker drain)`:
1. **The producer was the bottleneck.** The main thread round-robined every entity through per-worker `ConcurrentQueue`s — two interlocked ops per entity, ~740ms of freeze floor at 10M entities before any serialization happened.
2. **Round-robin distributes count, not cost.** "Deep" systems (50MB generic persistence blobs) and "thick" entities (100K-item storage keys) landed on arbitrary workers, producing lopsided drain times on large worlds.
3. **Worst-case scheduling.** `GenericEntityPersistence.Serialize` pushed its self-payload *after* all entities, and generic persistences sort last in the registry — so the biggest indivisible blobs started serializing at the very end, extending the freeze by their entire duration.
## The fix
**Commit 1 — chunked handoff + LPT scheduling + heap pre-sizing:**
- Pooled 4096-entity chunks published to one shared queue; workers pull chunks and load-balance dynamically (a worker busy with a thick entity simply takes fewer chunks).
- `Persistence.SerializeAll` pushes systems largest-first (LPT) using the previous save's payload size (or loaded file size on first boot); self-payloads get dedicated single-entity chunks so they overlap the entity stream instead of ending it.
- Worker heaps pre-size from the loaded save's `.bin` totals, eliminating copy-on-grow inside the first save's freeze.
- `SpinWait` backoff in the drain loop (never `Sleep(1)`), per-worker balance stats logged in debug builds (the call site is compiled out of Release), 1MB snapshot write buffer.
**Commit 2 — workers iterate the dictionaries directly + main thread joins the drain:**
- `GenericEntityPersistence` publishes 4096-slot ranges over its dictionary's backing entries array; workers serialize occupied slots (`value != null`) directly through a `ShadowEntry<TValue>` struct mirroring the runtime's private `Entry` layout. Safe because the dictionary is frozen during `Saving` (mutations divert to the pending safety queues).
- The layout is **proven at startup before any code reads through it**: validation measures the true `Entry` stride via precise allocation accounting (guaranteeing all shadow reads are in-bounds), then verifies every key/value of a churned, resized, freelist-exercised dictionary — reading value slots as raw pointer bits only, never materializing a managed reference until the layout is proven. If a future runtime changes `Dictionary` internals, validation fails with a logged warning and saves fall back to the (fully maintained) enumerate-and-push path.
- The main thread joins the drain via an inline worker after publishing, instead of idling — worth a full worker share, proportionally more on low-core hosts.
**Commit 3 — 2.2x faster BufferWriter write path, single-pass short strings:**
- PGO already devirtualizes and inlines every `IGenericWriter.Write` callsite (interface vs concrete measured identical) — the real per-write cost was the non-inlinable `Index` setter (range-check throw path + per-write high-water tracking) plus span bounds checks. Writes now reserve capacity once, then do an unaligned store through a ref with a raw index increment; the high-water mark folds at Seek/Resize instead of per write.
- Class-level implementations of the hottest default interface methods keep nested writes inlined (a DIM re-dispatches on `this` even at a devirtualized callsite).
- Strings of 85 chars or fewer encode once into a stack scratch instead of walking the string twice (`GetByteCount` + `GetBytes`). Byte output is identical.
- Measured: 34.4 → 15.7 ns/entity on a generated-style write mix; end-to-end freeze ~99ms → ~74-82ms.
**Commit 4 — branch-free fallback push loop:**
- A bare `foreach { PushToCache(entity); }` runs at 2.3ns/entity; the same loop carrying a per-entity heavy-check runs 2.3x slower — the cost is the fatter loop body defeating tight-loop codegen. Entity-level >1MB payloads are rare enough to ride in shared chunks; system self-payloads (the large ones) are still explicitly scheduled largest-first.
**Commit 5 — drop the 9-byte per-entity placement state; snapshots write from worker segment logs:**
- Every `ISerializable` carried `SerializedThread/SerializedPosition/SerializedLength` so `WriteSnapshot` could gather each entity's bytes from the worker heaps in dictionary order. But the idx records absolute positions — bin order is free — so the snapshot is now written in worker-heap order and the join inverts: workers log segments (owner, slot range, heap start) plus one length per record as they serialize; positions are implicit because a worker's writes are contiguous, and identity comes from re-walking the same snapshot slots in the same order (stable until `PostWorldSave`).
- Chunks are persistence-homogeneous (the partial chunk publishes at each `SerializeAll` boundary) so segments route to files by owner with zero per-entity state. Self-payloads keep placement as three private fields on the handful of persistence instances.
- Net: 9 bytes (plus padding) of resident state removed from every item, mobile, guild, and account on every shard; three interface-property stores per entity leave the drain hot path (stamping dirtied one cache line per entity mid-freeze — the lengths log is one sequential stream); each segment's bytes hit the bin as a single span write instead of one copy per entity, speeding the background write phase; and `IGenericSerializable` shrinks to just `Serialize(IGenericWriter)`. Transient cost: ~4 bytes per entity in pooled per-worker logs, released after each write. The save format was unchanged at this point (idx v3, same loader); the v4 bump comes later in the branch.
**Commit 6 — staged file writes replace memory-mapped snapshot writing:**
- `MemoryMapFileWriter` is removed. `FileBufferWriter` composes through the full `BufferWriter` raw write path into a pooled staging block that drains to the file as large sequential positional writes (`RandomAccess.Write`); seeks flush the block and move the file offset, so backwards patches (the idx entity count) become small positional writes.
- Memory-mapped composition paid a soft page fault on every composed page plus unpredictable dirty-section teardown stalls at dispose — measured ~4x slower end-to-end than staged writes at snapshot sizes.
**Commits 7–11 — idx v4: per-file type table replaces SerializedTypes.db and all runtime type tracking:**
- Previously every `Write(Type)` from every worker enqueued into a shared `ConcurrentQueue<Type>` during the freeze (interlocked writes on a shared cache line, millions of mostly-duplicate entries), the background phase drained and deduped it all into a `HashSet`, and the snapshot recomputed `xxHash64(Type.FullName)` once per entity record (~5.4M redundant hashes per save on a large world) to write 9-byte tag+hash idx records plus a global `SerializedTypes.db`.
- The db's only real job was diagnostics: the string name behind "Type `<X>` was not found. Delete all of those types?" during idx loading. That map now lives in the idx itself: each `GenericEntityPersistence<T>` keeps an insertion-ordered `Type -> ushort` table, hydrated at `AddEntity` and on every deserialize path — one dictionary `TryAdd` per entity add on the game thread, amortized across gameplay, and provably immutable while the background writer reads it (adds divert to the pending queues during saves).
- idx v4 layout: the table (names only) is written before the records; records reference it by 2-byte index, shrinking from 33 to 26 bytes (−21%). The loader resolves each table name **once** (`FindTypeByHash(ComputeHash64(name))` — semantically identical to v3 resolution, `TypeAlias` included) into a constructor array, and each record becomes an array index instead of an 8-byte hash read plus dictionary probe. The unresolved-type prompt now surfaces once per type, with the name.
- Deleted outright: `World.SerializedTypes`, the drain/dedupe pass in `WriteFiles`, `BufferWriter`'s type tracking (its `Write(Type)` is now pure — payload format unchanged: tag byte + xxHash64), `FileBufferWriter`'s typeSet parameter, `Persistence.WriteSerializedTypesSnapshot`, and the adhoc db write. SerializedTypes.db is no longer produced.
- **Backward compatibility:** v0–v3 saves (including their SerializedTypes.db and legacy tdb files) load exactly as before, and every legacy load path hydrates the new table so the first v4 save after an upgrade is complete. Stale db files in existing save folders are simply ignored. Verified live: a v3 save boots, saves as v4 (Items.idx −20.2% on a dev world), and reloads with identical entity counts.
## Measured (synthetic 10M entities / 1.7GB, 64+64+32MB system blobs, 24x 2MB thick entities, dense write profile, 24 cores)
| Metric | Before | After |
|---|---|---|
| Steady-state freeze | ~740 ms | **~78 ms** |
| Main-thread publish cost | ~740 ms | **~0.1 ms** |
| Steady-state allocations | 0 | 0 (by iter 2) |
| Worker byte-load spread | 2x | ~1.15x |
The freeze is now bound by pure serialize throughput (payload / cores).
## Tests
- 779 Server.Tests + 501 UOContent.Tests pass.
- New across the branch: chunk fill/flush/owner-boundary tests, pool reuse/clear tests, an end-to-end multi-worker drain through the real wake/push/flush/pause protocol, a 50K-entry churn equivalence test for the shadow iteration re-walk (the exact pairing the snapshot writer relies on), byte-level BufferWriter output/position pins, `RuntimeLayoutIsSupported` so a silent fallback on a future runtime upgrade fails loudly in CI, and a full snapshot **round-trip test** that serializes 25K entities plus a self-payload through real workers, writes the idx/bin from the segment logs, and reloads them through the standard loader (now in v4 format).
- For idx v4 specifically: `FileBufferWriter` staging/drain/seek-patch and oversized-item tests, type-table registration tests, a hand-written v4 fixture proving an unresolvable type name skips only its own records through the console confirmation flow, and a hand-written **legacy v3 fixture** proving old saves still load and hydrate the type table for their next save.
## Trade-offs
- Chunk scheduling is nondeterministic, so worker heaps ratchet to each worker's max-ever draw rather than a fixed share. With slot ranges the balance is tight (~1.15x), so the effect is small; a shared slab pool remains an option if production shows retention creep.
- Entity-level heavy items inside slot ranges are serialized wherever they're encountered (no LPT for them); worst-case tail is one thick entity's serialize time (~ms). System self-payloads — the large ones — are still explicitly scheduled largest-first.
- Snapshot-write error granularity is per segment rather than per entity (heap-bounds bugs were the only thing the per-entity catch ever caught; idx metadata reads keep per-record granularity).
- idx v4 is a save-format version bump: old saves load unchanged through the preserved legacy paths, but saves written by this branch require this loader. Per-persistence type tables cap at 65,535 distinct entity types per boot (hard throw, orders of magnitude of headroom), and a type's table slot persists until restart even if its last entity is deleted — a few stale name entries per file, by design.
Kamron Batman