[Re-land] #118 jemm native worker: 9 deduped commits onto main, conflict map and fidelity check #154

Re-land #118: cherry-pick 9 commits onto a branch from main

Why

#118 ("jemm: add native decision worker and pinned weight export") was merged at 2026-10-21T14:09:41Z with base=codex/qwen-vision-27b, one minute after #117 merged that branch into main. The result: compare/0410e84...codex/qwen-vision-27b = ahead_by 19 / behind_by 42, and the JEMM native worker (plus follow-ups) is not in main. #122, #123 and #143 still have that branch as base and modify files that do not exist in main, so they cannot land until this content does.

Decision (2026-10-21): re-land by deduplicating the commits by patch-id and cherry-picking the remainder, in order, onto a new branch from main. Do not merge the whole branch (#117 is already in main; the branch is 42 commits behind) and do not drop (#118 is real merged work).

Dedupe result (patch-id)

git cherry -v origin/main origin/codex/qwen-vision-27b -> 9 non-merge commits, all +: no patch-equivalent exists in main. git rev-list --no-merges origin/main..origin/codex/qwen-vision-27b = 9, so the 10 merge commits carry no additional commits. However, they carry conflict-resolution content ("evil merges"); see the fidelity check below - that content is not reproduced by a flat cherry-pick alone.

Branch tip at snapshot: 8945af81cbb747033836c90d85ca31907c0ec1c5.

Commit list (cherry-pick order, oldest first)

Conflict map (measured in a scratch worktree from 0410e84)

Cherry-picking the sequence above conflicts 6 times (recorded, then resolved mechanically by taking the re-land side to finish the simulation):

Fidelity check (do not skip)

Reference tree = the conflict-free branch->main merge result at 0410e84:

git merge-tree --write-tree origin/main origin/codex/qwen-vision-27b
# -> 7054a0f02b50b37c12e073ebe4cd7d25f56d7c69  (no CONFLICT, 60 files, +10129 -76)

After the cherry-pick branch is finished, git diff 7054a0f0 <re-land-branch> must be empty (or every difference explained). A simulation that resolved all 6 conflicts by taking the re-land side still differs from the reference tree in 14 files (+62 -331), which is what the merge commits' resolution content buys; the executor must hand-merge these, not just take one side:

Cargo.lock
README.md
docs/supported-models.md
mkdocs.yml
recipe/README.md
src/backends/cuda/qwen3_5/README.md
src/backends/cuda/qwen3_5/qwen3_5.backend.json
src/models/qwen3_5/native/src/model.rs
src/models/qwen3_5/native/src/vision/model.rs
tests/jemm/contract.rs
tests/qwen3_5/lora.rs
tests/qwen3_5/lora_fake.c
tests/qwen3_5/vision_batch.rs
tests/qwen3_5/vision_batch_fake.c

Acceptance

1. New branch re-land/118 from 0410e84 (or newer main; regenerate the reference tree if main moved) with the 9 commits in the order above; git diff <reference-tree> <branch> empty or explained.
2. CI rust + benchmarks green on the PR head: record run id + tip SHA in the PR.
3. Local on the branch tip: cargo fmt --all --check and cargo test --workspace --locked (macOS Apple M4, no GPU) - green; CUDA / accuracy / latency / e2e are not run (no NVIDIA GPU, no model weights locally).
4. New PR base=main, review approval; repo owner merges (bots do not merge/push). After merge, #127 -> #122 -> #123 -> #143 can retarget per [Triage] 30 open PRs: merge / rebase / author buckets, merge order and acceptance criteria (2026-10-21) #153.

Notes

• Do not cherry-pick merge commits; they are maintenance merges (content covered by the 9 commits, resolutions covered by the fidelity check above).
• The full-branch merge is conflict-free, but it was rejected in favour of the flat cherry-pick to keep main's 42 newer commits out of the PR history; the trade-off is exactly the 14-file residual above.
• Snapshot: main = 0410e84, 2026-10-21 ~23:10 +08. Re-verify before executing.

Executor: OpenCode/Codex (bots only produce the plan). Related: #153 (merge manifest), #118, #122, #123, #143.