erik/acdream
1
0
Fork 0
Code Issues Pull requests Projects Releases 6 Packages Wiki Activity Actions

docs(overhaul): defer FPS work by owner direction

Browse source
This commit is contained in:
Erik 2026-09-05 12:55:02 +02:00
parent 25e2702d33
commit 9c33535e45
5 changed files with 141 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

4
docs/research/2026-09-01-overhaul/481-entry-profile-recipe.md
View file

@ -1,5 +1,9 @@
# #481 — existing-profiler diagnostic recipe (not executed)
**OWNER-DEFERRED 2026-09-05 12:53+02:** "Dont fix the fps now. We do that
later." Do not execute this recipe or implement FPS changes now. Retained as
future preparation only; no measurement, fix or acceptance is claimed.
Wegener's read-only trace, lead verified actual RenderingDiagnostics startup
flags, FrameProfiler boundary/allocation sampling, VulkanFrameGpuMeasurement
timestamp pairing and PerformanceTools summarizer. No instrumentation change.

79
docs/research/2026-09-01-overhaul/c2-pre-overhaul-artifact-inventory.md Normal file
View file

@ -0,0 +1,79 @@
# C2 — surviving pre-OVERHAUL performance artifacts
**OWNER-DEFERRED 2026-09-05 12:53+02:** FPS/performance work is for later.
This completed read-only inventory is preserved; do not start the comparison
or optimization now. C2 performance is deferred, not passed or removed.
2026-09-05 read-only preparation; NOT a C2 run or PASS. Wegener searched
bounded campaign docs/logs/artifacts; lead independently read the anchor
reports, environment disclosure, frame summary and startup settings, checked
referenced artifact existence, and hashed all four reports below. No old
binary was built or launched. Current owner-fix graphics remain first priority.
## Accepted pre-OH anchor
FW closeout (`docs/plans/2026-08-30-campaign-fw-frame-walk.md:674`) names
`b8befded8b7bfa49628dcf3ded9436a19dd22c70`. Both reports declare source/binary
identity, PASS and graceful exit0:
- `logs/connected-r6-soak-20260831-083908.report.json`: nine checkpoints,
eighteen turn/stationary samples, 16,447 recorded frames. All ten non-null
named artifact paths exist (including report, logs, CSV, summary, counters,
environment disclosure and checkpoint timeline).
- `logs/connected-world-gate-20260831-083305/report.json`: six capped
checkpoints and ONE fresh uncapped-reconnect checkpoint; both artifact
directories exist. This is not a complete uncapped nine-stop route.
The R6 profile is retail/off render pack, 1280x720 FIFO, RX9070XT driver
2.0.395, quality High overridden to near4/far8/MSAA4/aniso16/A2C=true/
maxCompletions4. Profiler, runtime counters and movement-truth logging are on;
contention capture is off. The available report does not establish whether
the session used a physical display or RDP. This is NOT the current owner
High atmospheric / 1600x900 / far25 profile. Preserve historical route labels,
but use coordinates for scene identity (see479-rynthid-location-evidence.md).
Do not treat the route-wide p50/p99 or headline FPS as a stationary C2 window:
the full CSV includes preparation, travel and checkpoint work. Recompute
the exact matched windows and distinguish missing GPU samples from zero.
## Other surviving dense-town results
Both have actual companion artifacts, PASS, two turn/stationary samples and
one checkpoint, with the same profile except Immediate presentation:
| Report under logs/ | Exact commit | Coverage |
| --- | --- | --- |
| connected-dense-town-20260830-162046.report.json | 3a05d659074b28e37a9d866910bd69f1a1de3270 | Uncapped Arwic; 2,003 frames (agent-read summary) |
| connected-dense-town-20260830-170737.report.json | 1da178d9ba4e765564fde236151bc9857bb479eb | Uncapped Arwic; 2,761 frames (agent-read summary) |
Neither commit is the accepted b8befded anchor. The intervening170133/170414
reports failed before samples and must not supply a baseline. Agent also
located older VM0 real client logs, but those are pre-AR rather than pre-OH,
MSAA0, observer-free and a different workload; no substitution is authorized
by their mere existence. August22 Medium/off reports cited elsewhere were
not found in campaign logs during this bounded search.
No verified pre-OH capped-Arwic, Tusker, cathedral-stair or Facility performance
baseline was recovered here. This is a coverage gap, not a claim that those
files cannot exist anywhere. C2's complete20% comparison remains unproven.
## Execution constraints for the eventual comparison
The existing R6/lifecycle wrappers are useful route/artifact references but
their timeout tails force-kill clients. Do NOT execute them unchanged under
the owner's graceful-only rule. Their preflight also needs the separate
retail-client check. Use a bounded graceful-only adaptation or the current
selfgate runner, pin the session/package/settings and exact rebuilt HEAD,
and disclose observers. Never run old and new clients side by side. Script
PASS does not establish the C2 regression threshold, retained graph/membership
bytes, alpha high-water or absence of long preparation tails. Current #481
entry profiling is separate from this whole-campaign comparison.
## Lead SHA-256 seals
| Report | SHA-256 |
| --- | --- |
| R6 20260831-083908 | D4030362F0A482F00B1F23D95FBE1C02FEB5BBF10F787376017E4A40503092C3 |
| Lifecycle 20260831-083305 | 2C22E2EA4B0C48C93191118508108845E36D1B2F416B4472CEED4CC629C42D5C |
| Arwic 20260830-162046 | 6F5EB5F21730A048798035A069CE19C018A73ED5BF9ED7CB5B51573F0971466C |
| Arwic 20260830-170737 | 5BD761CF83CA653EA9851D64110E7BE9048225D3A18162C81C7AA1E4FA6F5AF2 |

19
docs/research/2026-09-01-overhaul/s5-consumers-material-closeout-packet.md
View file

@ -4821,3 +4821,22 @@ doorway nowPASS and full25-frame testPASS. First missing-opt-in attempt retained
and excluded (276/6/97). c1c-integrated-verification.md preserves classification.
c1a-geometry-activation.md now activates only§37.3–37.4/§45's test-only witness
in a fresh exact-base scratch, with graphical owner-issue work taking priority.
## 54. Owner defers FPS; correctness work continues
2026-09-05 12:53+02 owner: "Dont fix the fps now. We do that later."
#481 entry profiling/FPS repairs and C2 performance comparison/optimization
are deferred. The unexecuted481 recipe and completed read-only C2 artifact
inventory remain future preparation, not acceptance. No further performance
work now. This owner scope decision changes no retail algorithm or register
row; no renderer performance PASS is claimed.
Continue #477 crash/stability, #478 dungeon lighting, #479 shifted structure,
#480 visible held-item verification and prerequisite #476 reliable capture.
C1a is correctness-only and must not delay those graphical checks. Its first
full production-world witness reports four ramp cells where the earlier
partial-engine comparator reported three; preserve that failed contract
assertion, isolate the setup difference, and do not weaken/pin a golden over
the mismatch. No production change follows from it without a separate proof.
G4 remains FAIL; FPS deferral is not acceptance of the other owner findings.
No merge to main.