FDP PoC Summary

현재 기준 작업공간은 /home/ny/LMCache/private/work/fdp이다.

Status

Stage 1 branch audit, Stage 2 environment audit, Stage 3 build/sanity, Stage 4 harness dry-run 기록은 이 디렉터리 아래에 보관한다.
PoC source checkout은 $FDP_POC_SRC로 표기하며, 기본값은 /home/ny/workspace/LMCache-fdp-poc-src이다.
FDP 평가용 GitHub repo는 private nayeonikim/LMCache-fdp-eval이다.
최종 FDP 효과 결론은 steady-state / overprovisioning gate를 통과한 측정 결과가 들어온 뒤 갱신한다.

Repo: https://github.com/nayeonikim/LMCache-fdp-eval (private)
Baseline: DongDongJu/LMCache:fdp-waf-agentic-replay-poc @ 35784a3c4e52d1b3d73969bb6cd2abb7e7484440
Refs in FDP repo:
- main = imported baseline at 35784a3c
- upstream/dongdongju-fdp-poc = immutable upstream baseline branch
- baseline/dongdongju-fdp-poc-35784a3c = immutable baseline tag
Push access: nayeonikim only. Do not grant smrc/shared-server credentials push access.
smrc/shared server policy: the server is only a temporary authoring and execution environment. Local commits are allowed there for checkpointing evaluation docs/results, but durable management happens on the development PC. Pull docs/results back to /home/ny/LMCache/private/work/fdp and/or /home/ny/workspace/LMCache-fdp-eval, review, then push from the development PC only.

private/work/fdp: personal SSOT for planning, status, analysis, interpretation, and links to result artifacts.
private/work/fdp/CHANGELOG.md: FDP-only changelog for smrc evaluation docs/results and pullback history. Do not mix these entries into the private root changelog unless the private repo structure itself changes.
LMCache-fdp-eval: reproducible evaluation workspace for configs, scripts, run manifests, small reports, and server-authored patches that should be shareable.
Local development clone: /home/ny/workspace/LMCache-fdp-eval
smrc: temporary run location. It may create local commits or patches, but those are transport artifacts until pulled back to the development PC.
Raw traces and bulky results stay outside Git. Keep only manifests, summaries, small logs, and paths/checksums in Git.

아직 FDP 효과를 확정 주장하지 않는다. 기존 32GiB window-local 결과는 harness 동작 검증으로만 취급하고, 다음 평가는 05_eval_redo/01_steady_state_overprovisioning_plan.md를 기준으로 진행한다.