[k0valik]

Fork replay — 🟢 READY

upstream	https://github.com/Niko1221/Strata.git (main)
foreign fork	https://github.com/gputier/StrataGP (main)
foreign commits	1138 (61b3fb5dd3f1..ee952078114d)
their fork point	387 commit(s) behind upstream; -387 (their fork point is older than our accepted baseline)
replayed onto	93caa81eb4b9 (HEAD)
raw change	7 files changed, 901 insertions(+), 14 deletions(-)
replayed change	** 2 files changed, 418 insertions(+), 3 deletions(-)**
filtered by policy	5 dropped, 0 ours, 0 protected, 0 unclassified
replay branch	replay/hub.lumenfield.work-gputier-StrataGP-main — 1 squashed commit(s), 2 file(s)
verification	0 structural leak(s); pristine files: 2 exact / 0 mismatched; both-sides-changed: 0; conflict-marked: 0
published	replay/hub.lumenfield.work-gputier-StrataGP-main pushed to origin — a pull request opens from this report

What changed

• serve — 3 suppressed by fork policy
• tools — 2 integrated
• docs — 1 suppressed by fork policy
• (repository root) — 1 suppressed by fork policy

Foreign fork commits (1138, showing the oldest 10 and newest 25)

• f2a08d4b2da9 Strata: Qwen3.8-Flash-Next on a 12-24 GB NVIDIA GPU + 64 GB RAM, one-click install
• 28c47b103ea9 Swift 1.5 support (UkisAI's fine-tune): split shards, per-model chat template, engine v0.1.1
• 1ee8b666c13e README: a short, beginner-friendly front page (details in docs/DETAILS.md); pictures in WebP/TIFF/AVIF
• 6da1f667e865 Engine v0.1.2: streaming tool calls, live status, real cancel; fix requests stalling on a full GPU
• dd03276deaa9 serve: fix zero-token replies from engine-queue race between requests
• 23f6b15b6b80 serve: conversation cache - keep the chat between requests, read only what is new
• c47e290d35ce cpu pool: sleep between requests instead of spinning (fixes replay (review only): hub.lumenfield.work-eelgaev-Strata-AC922-ac922 — filtered squash of 920 commit(s) from https://github.com/eelgaev/Strata-AC922 #4)
• 9d9351b9abdd serve: read short prompt parts through the decode windows
• 80a1f1327b52 Merge branch 'pr-7'
• 30acb0c7a89e Merge branch 'pr-8'

1103 commits omitted; the whole range is 61b3fb5dd3f1..ee952078114d.

• 5c068599b63b Merge branch 'pr/prompt-encoder-0.1.37'
• 76b916e2b545 tools: llama.cpp's qwen35 vectors are checked with the qwen35 vocabulary
• 33b4e4d47533 Merge branch 'pr/prompt-encoder-0.1.37'
• c319b1137b07 serve: the prompt is tokenised from the last shared prefix, not from the start
• b98660e9ec64 serve: serve: tokenise the prompt from the last shared prefix, not from the start Niko1221/Strata#567's server tests take their reference from a full encode_prompt
• 854564d580ec tools: llama.cpp's qwen35 vectors are checked with the qwen35 vocabulary
• 8941f1f46ba5 serve, tools: serve: tokenise the prompt from the last shared prefix, not from the start Niko1221/Strata#567's tests and ByteTokenizer without repeated code
• 96e6139bfe65 serve: serve: tokenise the prompt from the last shared prefix, not from the start Niko1221/Strata#567's kept prompts, one per conversation, none refused for the context
• a62a1924ac9a Merge branch 'pr/prompt-encoder-0.1.41' into fork/main-0.1.41
• ca560719e951 docs: FORK.md, the fork restarts from v0.1.41 with serve: tokenise the prompt from the last shared prefix, not from the start Niko1221/Strata#1852 only
• fe23c36fe24f Merge branch 'main' into fork/main-0.1.41 (old main kept by history only)
• c73b389b31b8 serve: the prompt is tokenised from the last shared prefix, not from the start
• 4a1bab0c0567 serve: serve: tokenise the prompt from the last shared prefix, not from the start Niko1221/Strata#567's server tests take their reference from a full encode_prompt
• 62ec02d31f89 tools: llama.cpp's qwen35 vectors are checked with the qwen35 vocabulary
• e13bb1f12832 serve, tools: serve: tokenise the prompt from the last shared prefix, not from the start Niko1221/Strata#567's tests and ByteTokenizer without repeated code
• 6f6ebc156ef7 serve: serve: tokenise the prompt from the last shared prefix, not from the start Niko1221/Strata#567's kept prompts, one per conversation, none refused for the context
• da7fa640c5f7 serve: keep one prompt per engine conversation cache slot
• 401cc508410b serve: count tokens without storing the prompt in the encoder
• 3569e1e164ff Merge branch 'pr/prompt-encoder-0.1.42' into fork/main-0.1.42
• 94bdaa67364e docs: FORK.md, the fork restarts from v0.1.42 with serve: tokenise the prompt from the last shared prefix, not from the start Niko1221/Strata#1852 only
• eb111e866afe Merge branch 'main' into fork/main-0.1.42 (old main kept by history only)
• 9b9253061e70 serve: size the kept prompts from --conversation-cache-slots on a lazy start too
• 80ce1c0dd505 Merge branch 'pr/prompt-encoder-0.1.42' into fork/main-0.1.42
• 3e03db842bcb docs: FORK.md counts the slot fix for lazy starts and the new test totals
• ee952078114d docs: FORK.md names main, fork/main-0.1.42 is deleted

dropped — 5 path(s) the foreign fork changed that this fork does not take

  SECURITY.md (M)
  docs/FORK.md (A)
  serve/server.py (M)
  serve/test_prompt_encoder.py (A)
  serve/test_server.py (M)

mirror — 2 path(s) taken from the foreign fork

  tools/strata_tokenizer.py (M)
  tools/test_strata_tokenizer.py (M)

Groups (one squashed commit each)

• tools — 2 applied (2 clean, 0 conflict-marked) file(s)

Where the mass is (top 10 files by churn on the replay)

Start reading here or skip accordingly: one rewritten file explains more
than ten thousand added lines of data.

     300  +298/-2  tools/test_strata_tokenizer.py
     121  +120/-1  tools/strata_tokenizer.py

Inspect locally

git log --oneline HEAD..replay/hub.lumenfield.work-gputier-StrataGP-main
git diff HEAD...replay/hub.lumenfield.work-gputier-StrataGP-main --stat
git log --oneline 61b3fb5dd3f1..ee952078114d   # the original commits

Review artifact — produced by fork-replay.sh. The replayed commits exist
for reading: building and testing them is a human step, exactly as for a
synchronization candidate. Nothing here is evidence that any of it works.

---

No upstream code was executed while preparing this synchronization: it reads git
objects and applies .fork/path-policy.

Building and running the engine against a candidate is a manual step (see
FORK_NOTES.md); this workflow runs no engine CI.