[grantoverton]

Part of #1190. Stacked on #1125 and #1189 (the score/decide queue); base for review: the #1189 head 227b961 — this branch carries its parents' commits until they merge. Review the last three commits: ee2ed50 (the head format and the decide-pass read), 9a2f2ef (option positions derived from the render), 55b51c0 (lean-check comments).

A --joint-head <dir> serve flag loads a small trained side head and the decide/score passes read its option scores off the trunk's post-final-norm hidden rows — each option scored at its own final-token position instead of label logits. #1154 scores labels off the prompt's last row; this PR's option-final map is a different position contract, derived from the render (prompt_format_version semantics) — flagged here so the convention settles in review.

What changes

• native/qwen27_joint.zig (new): joint_head_config.json (format bf16-safetensors-v1) gives the head's shape — the trunk's hidden dim, the head's width, and its Linear→GELU residual layers. Weights: proj (dim→width), the residual blocks, out (width→1); bf16 words read from the safetensors file.
• native/metal.zig: the open path passes --joint-head through to the qwen27 backend, and the module's tests load the head.
• native/qwen27_served.zig: the host loads and owns the head (its dim checked against the checkpoint), closes it, and hands it to the decide pass.
• native/qwen27_decide.zig: with a head loaded, labels are option-final token positions rather than label ids. forward prefills the ids in chunks and scores each positions entry on that row's hidden state — for score, positions inside the prompt; for decide, each branch's option-final positions as absolute offsets into the whole sequence (the trunk's length on a branch's rows). Without a head the pass is unchanged (label logits at row 0).
• server/decisions.zig: optionPositions derives each option's final-token position from the render — options sit one a row between input and the closing inside content.
• server/model_text.zig: indexAt maps a text span's end offset to the covering token's index.
• server/systemone.zig: a side head's readout — one group an option, its line's last token's index based on the trunk's length; every option rendered (the template's lone-description case gets its bare answer words).
• server/cli.zig: --joint-head flag.

Checks

• zig build native -Dcpu=apple_m1 -Dversion=1.0.0 passes clean on the head.
• zig build test test-golden -Dcpu=apple_m1: test-golden 0 unexpected differences (306/319 + 26/26 equal). The test step's device-dependent binaries fail on this machine while its GPU is serving a model — the same test binaries fail identically on the base head 30e5dcb (run for comparison), so the failure is inherited, not introduced by this diff.
• tools/zig/lean_check.py: 444 findings, as on the base (zero added).

Receipts (serving validation B3 with the head loaded, receipts/DECIDE-RECEIPTS.md; Apple M3 Ultra, 256 GB, macOS 26.4.1, zig 0.17.0; local Ornith-1.5-122B A1-affine 4-bit + DFlash2 drafter; stub head dim 3072 / width 1024 / 4 layers)

• /v1/decisions returns real scores with the stub head on a live serve: choice + yes_no + score all answered, label_mass 1.0, on a 270-token prompt at 4-way and 8-way load. /v1/systemone shares the scorer and the position map but has no recorded run on this stack.
• Chat traffic completes normally with the head loaded (B3: drafted chats at the recorded rates, ~0.55 acceptance). The head reads only during decision passes, and a decision landing mid-drafted-round leaves the token stream unchanged (qwen27: score and decide passes queued onto the lane host's idle engine thread #1189's integrity receipt); no with/without-head token-sha comparison was run.
• A trained head of the same format drops in without code changes (position-invariant joint head; option order shuffled, pick ordering stable).

Known limits

• Each decide branch re-prefills the shared trunk per question on the no-head path; with a head the trunk state is kept once and restored per branch (see qwen27: score and decide passes queued onto the lane host's idle engine thread #1189's note — trunk KV reuse is future work either way).
• Decisions with pictures are refused (d.pictures.len > 0) — image-aware scoring is not claimed.

Not run

As #1189: no new GPU jobs for this cut; the figures are from the recorded validation of the stack tip. Cold-prefill tables and bench_concurrent comparisons would be re-run at file time on the exact head if wanted.

[ashhart]

Thanks @grantoverton. This sits on the unmerged #1125 stack and on #1189, which was sent back, and it conflicts in 9 files on current main. It keeps #1189's second decision queue (decisions/decided) instead of main's lane_host_score.zig, sets the decider on every qwen27 host (so /v1/decisions and /v1/systemone answer by default), and its wait has no cancel or timeout. Once #1189 is reworked on main's score queue as an opt-in, please rebase this onto it; we'd then check that drafted equals plain and concurrent equals solo with a decision in flight.

[ashhart]

Review is on #1189, covering the whole stack #1189–#1215.

[ashhart]

Thanks @grantoverton for the JointHead idea. We're closing this with #1189, as our 26 Oct review hasn't had a reply. If you pick it up again, a qwen27 implementation of #1361's hidden call, with the head in the server, would be the best fit.