[jayleaton]

What this changes

Extract #662's family-neutral lane host and lane core on main c7137bb31a7dad924fb541a7068422127b9a94b2, preserving the prompt-piece lifecycle, opt-in planner, proposer/row-policy hooks, deterministic pricing and halt behavior with their tests. This branch is independent: it requires neither sessions (#637) nor TP, CUDA or DeepSeek code.

Illustrative control flow:

LaneHost
  default loop (existing whole/chunked prefill)
  opt-in Rounds -> admit -> prompt pieces -> finals -> shared decode
  halt -> next round boundary -> fail queued/filling/live requests
lanes.Engine -> optional proposer, row policy, timed pricing

Receipt

• Environment, with the TensorFold commit, MLX or PyTorch version, machine and GPU, checkpoint and revision: local Linux CPU host, Zig 0.17.0; candidate 7c8ba10e622ed5695fb416cea160d23da8accb30, pinned base c7137bb31a7dad924fb541a7068422127b9a94b2. No checkpoint, GPU, MLX or PyTorch used.
• Exactness, from tools/bench_concurrent.py --alone --serial or token_sha against "draft": false: no device/model receipt. Scripted host tests exercise identical token output for queued prompts in pieces versus whole prompts, cancellation, failure isolation and halt cleanup.
• Decode speed before and after, from tools/bench_openai.py: not measured; this is a host-only extraction, and planner/backend performance needs a model/device gate.
• Prompt speed before and after: not measured; optional prompt-piece processing can affect prompt execution and needs backend qualification.
• Tests run, with pass, fail and skip counts: final clean-head checks below. Some targets used Zig’s cache; counts describe this invocation, not a full-suite total. Direct same-engine-tree receipts from the original extraction: engine API 117/117 and lane core 85/85 passed. The earlier uncached extraction host receipt recorded 753 passed, 5 skipped; those skips remain explicit. The app-only base advance leaves all engine sources unchanged.

Gresume-shared-lanes: commit 7c8ba10e622ed5695fb416cea160d23da8accb30 tree 663bcf6cfd0ff0c32fe3b05b18ad61e41c85c240 dirty 0 2026-10-15T01:06:55Z
install rc=0 Build Summary: 10/10 steps succeeded
test rc=0 Build Summary: 46/46 steps succeeded; 753/758 tests passed (5 skipped)
test-golden rc=0 Build Summary: 4/4 steps succeeded
fuzz-cortheon rc=0 Build Summary: 3/3 steps succeeded; 8/8 tests passed
test-cuda-host rc=0 Build Summary: 27/27 steps succeeded
test-server-cpu rc=0 Build Summary: 3/3 steps succeeded
zig fmt --check rc=0 0 files
git diff --check rc=0
7c8ba10e lean: 527 problems, 527 on the base, 0 new
FAILED STEPS 0

• Not run, and why: CUDA/Metal/model arithmetic, checkpoint exactness and speed checks are not run for this host-only extraction. Five host skips remain; host results do not qualify a device.
• Before: main lacks api.Rounds, prompt-piece hooks and Opened.halt. After: the optional APIs and lanes: a family's round planner drives the lane host with prompt pieces, and a halt ends the rounds at a drain deadline (DeepSeek-V4.1 stack 6/14) #662 lifecycle tests pass on main; the complete zig/src/core tree equals the rebased stack/06-lanes core tree.

Merge Danger — Door: two-way. Blast Radius: core. A code revert restores the previous API and host loop. Optional planner hooks affect shared request scheduling, admission, cancellation and state cleanup for any backend opting in; device consumers still require qualification.

Checklist

• Every token still goes through the lane rounds. No serial path, and nothing that needs drafts off.
• Drafted output equals "draft": false, a resumed prompt equals a fresh one, and concurrent equals solo.
• No precision traded for speed. If the bits change, the description says which sums change and why.
• Prompt processing is no slower than the last release.
• The description says which platforms I ran, Metal M1 to M5 and CUDA, and which I could not.
• New tests fail before the change, pass after it, and skip cleanly without their dependency.
• Comments and docstrings are one line. No measurements or history in the source.
• No personal data, machine names, internal hosts or local paths. No AI attribution lines.

[ashhart]

Thanks, @jayleaton. Needs changes at 7c8ba10. In zig/src/core/lane_rounds.zig:116, a seventeenth piece calls h.drop, freeing its job while its key remains in finals; runFinals then dereferences it. Please batch pieces or remove every remaining reference before freeing. Also, zig/src/core/lane_halt.zig:20-24 clears job lists even when copying them fails, losing requests without terminal events. Drain without allocating or transfer the existing lists' ownership. These are opt-in path defects; I found no separate default-loop token change.

[jayleaton]

Thanks, @ashhart. Both are fixed in 7293d9e7. The current head, 4c351a39, is an empty commit on the same tree, added to rerun CI after the fmt job timed out downloading setup-zig.

• Pieces past sixteen (lane_rounds.zig): pieces now run in bounded batches, and a job is no longer freed while a later piece or its final still refers to it. A failed repeated piece drops its job once, and that job's remaining pieces and finals are skipped.
• Halt (lane_halt.zig): halt now takes over the existing queued and admitted lists under the lock. It no longer copies them, so it can't fail on allocation, and every request gets its terminal event.

New regressions in zig/src/core/lane_review_test.zig, run by zig build test:

• "more than sixteen planned pieces retain their job until its final". Before the fix this faulted with the page allocator.
• "halt transfers queued and admitted jobs when every subsequent allocation fails". Before the fix it expected two terminal events, got zero, and leaked the jobs.
• "failed repeated pieces drop their job once and skip its remaining batches".

Host gate on tree a21d86cb (Linux x86_64, Zig 0.17.0): build 10/10, test 46/46, test-golden 4/4, fuzz-cortheon 3/3, test-cuda-host 27/27, test-server-cpu 3/3. zig fmt and git diff --check are clean, and lean_check shows 527 inherited findings and 0 new. desktop, mobile, zig and fmt pass at 4c351a39. No device or model run was done for this change.

[ashhart]

Thanks @jayleaton for the bounded-piece and allocation-free halt repairs, and for the three regressions. Both original lifetime findings are addressed in the common-runtime integration: a job survives every remaining piece/final, and halt transfers the existing lists so allocation failure cannot lose terminal events.

We retain the repeated-piece drop-once control and your authorship. The empty CI-rerun commit does not change that source result. Host coverage stays separate from each consuming family's device checks, and later window changes retain these shared planners rather than introducing another scheduling loop.

Posted by Ash's AI Assistant Synthena

[ashhart]

Thank you, @jayleaton, for fixing the planner findings and porting the round planner onto 1.0.12's host.

We carried your work into 1.1.0 in 4a9b8f7, preserving your authorship. Our integration adapts stream opening to the released host and makes failure, cancellation and close finish every held job without allocating or opening an unadmitted stream.

The fixtures exercise planner failure and allocation failure during cleanup. No family selects this planner in 1.1.0. We are closing the PR because its replayed commits have different IDs.

Posted by Ash's AI Assistant Synthena