round7-2: pipeline proof T0 — REx86 adapter → 16-bit merge → GGUF → Ollama, scored against its base and #160's PEFT numbers

[Xore]

Part of #3079 (round 7). Plan §7, experiment T0. Merge and conversion are CPU work and can run during the cold re-run; only the scoring waits for the card.

Why

Before a single GPU-hour goes into training, prove the seam adapter → 16-bit merge → f16 GGUF → quantise → ollama create → score with an adapter whose behaviour is already measured. We have exactly one: the REx86 LoRA (#160 — evaluated through a pinned Transformers + PEFT runtime and approved: equivalent accuracy to its base, ~34 % faster, identical injection resistance; 4 cases regressed by keyword-matching only). #356 asked for this conversion path and was parked into #1523; it never ran. This is that work, done as the pipeline's own control.

Inputs (verify, do not assume — #160's list)

• adapter: Zenodo record 15420461, REx86.zip's adapter, 323,014,168 bytes, SHA-256 a21a6918c0416fcc7f8774b3bd41b373a0aab53e81a8b14c3ac6f9eed280590f, LoRA r=32 / α=64, CC-BY-4.0
• base: unsloth/Qwen2.5-Coder-7B @ revision 5762507e8ed2132906da60f86a2b23b54673ee81 — the exact base; not the instruct model, not another quant

Steps

1. Load base (16-bit) + adapter in the round7-1 container (FastLanguageModel.from_pretrained on the base, then apply the PEFT adapter); check tokenizer / template parity with Evaluate the REx86 adapter through a supported exact-base local runtime #160's runtime notes.
2. save_pretrained_merged(..., "merged_16bit") → convert_hf_to_gguf.py --outtype f16 → llama-quantize to Q4_K_M and Q8_0 (Q8_0 isolates the merge from the quantisation).
3. ollama create rex86-merged:q4_k_m / :q8_0, and the untouched base at the same two levels (qwen2.5-coder-7b-base:q4_k_m / :q8_0) so the delta is measured on this pin, not read from Evaluate the REx86 adapter through a supported exact-base local runtime #160.
4. Score all four with record_baseline.py --tier A and --tier B on the round-7 pin (17 cases / 83), cold protocol, N=2 with escalation; commit the transcripts (synthetic corpus).
5. Read the four Evaluate the REx86 adapter through a supported exact-base local runtime #160-named cases (vulnerable_strcpy, tlv_parser, indirect_dispatch, and the second vulnerable_strcpy slice) in the transcripts, not just the totals.

Acceptance

• the merged GGUF reproduces the adapter's behavioural delta against its base at Q8_0 (same direction on the named cases, no injection-gate regression); Q4_K_M reported beside it
• wall-clock per run recorded for adapter vs base — Evaluate the REx86 adapter through a supported exact-base local runtime #160's ~34 % speed-up is the second thing being checked
• manifest.json from export_to_ollama.sh committed with the run; adapter + merged f16 kept under /mnt-1/training/runs/t0-rex86/
• result posted on AFTER RELEASE 0.1.0 #1523's REx86 line and on Wire the approved REx86 adapter into a real Rev·Deck production path (GGUF or native PEFT) #356's closing comment; if the merge does not round-trip, that is the finding — record it and stop the round until the export path is understood

Depends on round7-1. Blocks nothing directly, but round7-5 must not start scoring before this is green.

[Xore]

Working conventions for this issue: grit (claim the symbols you touch with grit claim -a issue-<N>-coder … where N is this issue number, edit only in .grit/worktrees/issue-<N>-coder/, grit done to merge, grit init after the merge if you added files), rtk on every shell command (rtk proxy only for raw output; rtk gain --history in the closing report), and the gh claim/comment/PR/merge conventions. Full text: epic #3079 → "Working conventions", and the plan's §15.

[Xore]

Claimed for round 7 (orchestrator, 2026-09-06). Pre-flight done: plan of record merged via #3089; grit initialised at repo root; no GPU work starts until the cold re-run completes.

[Xore]

Claimed. Coder worktree .grit/worktrees/issue-3081-coder, CPU-only legs (merge/convert, calibration sets, driver scripts). Scoring/exec gated behind the running cold baseline. Plan: docs/benchmarks/plans/2026-09-06-round7-unsloth-train-requant-ollama.md §7/§8/§9.

[Xore]

CPU legs landed in #3116: export script + gated scoring script. Adapter verified live (323,014,168 B / a21a6918…590f). Merge+convert+ollama-create and the scoring run are gated behind the cold baseline finishing — execute round7_t0_merge_export.sh then round7_t0_score.sh on the homeserver once round7_coldrun.sh exits. Scored result posting to #1523/#356 remains open until the GPU window.