acdream/docs/research/2026-07-24-slice-e-cost-budgeted-streaming-report.md

# Modern Runtime Slice E - Cost-Budgeted Streaming Report

**Date:** 2026-07-24
**Exact source/binary:** `91e82c3c6850fcedbf20322b2ecf4fdcd11a2b2a`
**Configuration:** Release, physical local AMD display, 2560 x 1440
**Result:** PASS

## Scope

Slice E replaces portal-time whole-window work with one typed, frame-scoped
streaming scheduler. Completion admission, retained CPU payloads, entity work,
requested GPU bytes, and GL-retirement admissions share one budget. Stable
destination/control/unload/Near/Far queues preserve exact FIFO order and retry
receipts. Old generations become unavailable and stop simulation, collision,
targeting, effects, and owner audio immediately while physical retirement
converges over later frames.

Publication now advances through retained render, physics, static, building,
EnvCell, and spatial receipts. Destination work receives a reserved share of
the same ceilings; it does not bypass them. Login and portal transit remain in
the authored portal scene until the canonical render/composite/collision
predicate is complete, with retail's centered five-second wait notice instead
of forced partial materialization.

The final connected gate also corrected two ownership bugs found only under
real destination quiescence:

- CPU mesh-cache hits may stage GPU work only while their exact owner token is
  still live. An evicted owner can no longer recreate stale upload work.
- A destination live object may acquire spatial/render presentation while the
  world is quiesced, but remains invisible to draw, collision, picking, radar,
  status targeting, effects, and audio until reveal. Loaded residency and
  gameplay visibility are separate facts.

Raw JSON, CSV, PNG, counter, and frame-history artifacts remain in the ignored
local `logs/` tree:

- `connected-r6-soak-20260724-202244.*`
- `connected-r6-soak-20260724-203225.*`
- `connected-dense-town-20260724-204109.*`

## Connected results

| Route | Result | Elapsed | Failures | Warnings | Graceful |
|---|---:|---:|---:|---:|---:|
| Capped nine-stop | PASS | 561.516 s | 0 | 13 | yes |
| Uncapped nine-stop | PASS | 518.748 s | 0 | 12 | yes |
| Dense Arwic uncapped | PASS | 69.741 s | 0 | 1 | yes |

Warnings were expected authoritative population changes, known DAT-driven VFX
diagnostics, and empty world-edge landblocks. No warning represented a reveal,
streaming, residency, rendering, or shutdown failure.

## Frame cost and allocation

The values are complete-route frame-history percentiles. Allocation is the
largest positive frame-thread allocation.

| Capture | CPU p50 / p95 / p99 | GPU p50 / p95 / p99 | Frame allocation max |
|---|---:|---:|---:|
| Capped Slice A | 4.287 / 6.972 / 9.404 ms | 1.702 / 3.261 / 3.847 ms | 203.5 MiB |
| Capped Slice C | 4.247 / 5.873 / 7.814 ms | 1.836 / 2.694 / 3.078 ms | 46.8 MiB |
| Capped Slice E | 4.213 / 4.940 / 6.139 ms | 1.650 / 2.576 / 2.921 ms | 31.1 MiB |
| Uncapped Slice A | 1.520 / 5.694 / 7.825 ms | 0.897 / 3.239 / 3.406 ms | 202.9 MiB |
| Uncapped Slice C | 1.409 / 4.909 / 6.496 ms | 0.832 / 2.510 / 2.706 ms | 39.9 MiB |
| Uncapped Slice E | 1.216 / 4.216 / 5.450 ms | 0.725 / 2.486 / 2.666 ms | 36.0 MiB |
| Dense Arwic Slice A | 3.672 / 6.561 / 14.121 ms | 2.476 / 3.284 / 3.752 ms | 28.5 MiB |
| Dense Arwic Slice C | 3.487 / 6.044 / 12.432 ms | 2.126 / 2.768 / 3.122 ms | 26.9 MiB |
| Dense Arwic Slice E | 2.652 / 4.153 / 5.736 ms | 1.576 / 2.504 / 2.662 ms | 30.0 MiB |

Against Slice A, capped CPU p99 improved 34.7%, uncapped CPU p99 improved
30.4%, and the corresponding largest frame allocations fell 84.7% and 82.2%.
Against the package-only Slice C checkpoint, capped/uncapped CPU p99 improved
21.4%/16.1%, GPU p99 improved 5.1%/1.5%, and largest frame allocation improved
33.5%/9.7%. Dense Arwic CPU/GPU p99 improved 53.9%/14.7% from Slice C while
exercising 15,530 entities and 111 animated entities.

Dense Arwic's 30.0 MiB allocation maximum is 3.1 MiB above Slice C's dense
capture. It remains far below the portal baseline and is not hidden by a
changed threshold; Slice H's whole-world partition and unconditional
frame-owner work remain the program-level allocation gate.

### Wall-clock budget interpretation

`StreamingWorkMeter` measures elapsed wall time on a non-real-time Windows
thread. A frame may therefore report an overrun when the OS preempts an
indivisible operation. The meter never admits another operation after the
deadline and records the exact stage. Count/byte/entity ceilings likewise
permit only the documented first-indivisible-operation progress rule.

The uncapped route's largest observed streaming operation was 32.055 ms in
`publication-physics-statics`; the capped route's was 45.446 ms in
`publication-spatial-commit`. These are named observer-atomic boundaries, not
an unmetered continuation loop. Complete-route p99, per-destination turn
windows, and repeated-route behavior do not show a persistent 32-45 ms
streaming cost. Slice F/G is the approved architectural boundary for replacing
the remaining render-scene spatial commit with prepared deltas; retail
per-object collision registration remains one atomic physics owner operation.

This interpretation reconciles the plan's absolute wall-clock wording with
its earlier first-operation progress contract. It does not claim that Windows
can guarantee a 2 ms wall-clock maximum across thread preemption.

## Convergence and ownership

Every canonical checkpoint reported:

- `invariantFailureCount = 0`;
- reveal ready, materialized, complete, and world viewport observed;
- zero pending publication and retirement transactions;
- zero worker, destination, control, unload, Near, and Far backlog;
- zero staged mesh uploads and staged bytes;
- zero composite warmup candidates;
- zero residency staging, requested-GPU, and retiring bytes.

On the uncapped third Caul visit, cache-attributed committed CPU/physical GPU
residence was 41.1/310.8 MiB, below the second visit's 48.8/292.3 MiB in CPU
and within the configured GPU budget. The difference in physical GPU residence
tracks the changed authoritative/visible population and composite ownership;
no staged or retiring owner remained. Final process working/private memory was
1,716.4/1,461.3 MiB, below Slice D's 1,776.5/1,543.8 MiB final uncapped sample.
The prepared package remains memory mapped, so working set includes discardable
clean file-backed pages and is not substituted for ownership accounting.

## Automated gates

- Deterministic budget, reservation, pressure, cancellation, stale-generation,
  retry, callback-reentrancy, and GUID-reuse suites pass.
- Destination quiescence tests prove retained live objects can prepare without
  becoming drawable or gameplay-visible.
- Mesh-cache tests prove an unowned cache hit cannot recreate staged upload
  work and exact ownership reacquisition stages once.
- Release solution build passes.
- Complete Release suite passes: 8,164 tests, 5 intentional skips.
- Exact source and binary commits match for all three physical runs.
- Every client completed graceful character logoff and transport shutdown.
- Slice E changes scheduling and hidden destination preparation only; the
  existing fixed-camera harness found no pixel-change gate requiring new
  subjective approval.
- Divergence audit found no new retail behavior difference. AD-2 already owns
  the asynchronous readiness/quiescence adaptation; E6 only corrects its
  internal residency and cache-owner implementation.

## Conclusion

Slice E's scheduler, lifecycle, reveal, convergence, allocation, and connected
performance gates are complete. The client now spreads destination publication
and old-world retirement across explicit work budgets without exposing partial
worlds, replaying owners, or reducing visual quality. Slices F/G remain behind
their explicit user approval gate.