FINAL: run closed as pocket upsampler per user decision

RUN_LOG.md

@@ -5,10 +5,12 @@ One line per job: what ran, its cost, its result, the next step. If the session
 | # | When (UTC) | Job | Cost | Result | Next step |
 |---|---|---|---|---|---|
 | 0-9 | 2026-10-15/16 | Verification → G0 → G2 (PASS) → G3a/c (OWLv2 chosen) → request gen (7,842) → G5a/G5b (overfit PASS) | ~$5.0 | See earlier rows. | Labelling. |
-| 10-12 | 2026-10-16 | Labelling (97.0% pass) → filter (STOPPED 27.7% = checker artifacts; fixed per user; rerun 96.6%) | ~$1.6 | 7,575 clean rows. | SFT. |
-| 13-15 | 2026-10-16 | SFT 0.8B + SFT 2B + text eval (301/301 × 3 arms) | ~$11.0 | 0.8B parse 95.0/IoU 0.470; 2B 92.7/0.529; untuned baseline 1.0%. | GGUF. |
-| 16-18 | 2026-10-16 | GGUF Q8 (812 MB) + CPU latency (≈20.4 s/plan) + model cards + deliverables verified | ~$0.1 | Complete. | Klein renders. |
-| 19 | 2026-10-17 | Klein token saga resolved: session OAuth lacks gated-repo read scope; user's scoped PAT verified (200 + download) | ~$0.1 | PAT: hf_taGO…uLF. | G3b. |
-| 21 | 2026-10-17 17:53→19:15 | G3b — 7 attempts (L4 OOM → pop prompt_upsampler per diffusers docstring; tokenizer component wiring ×3; owlv2 kwargs ×2; fixture crash; upload 403 → dual tokens) | ~$1.3 | **G3b DONE**: mismatch real>swapped 2/3 scored (five_apples inverted — collection-row scale). **Timings L4 fp8r: 512² ≈ 4.6-5.0 s, 1024² ≈ 23 s.** results/g3b_render_check.json. | G4. |
-| 22-23 | 2026-10-17 20:14→21:27 | G4a pup both modes (a100; 3 attempts: 24GB OOM, missing limit_mm_per_prompt, missing VLLM_USE_FLASHINFER_SAMPLER=0) + G4b render+score (l4x1; ~6 attempts: mkdir/ids, meta re-attach, local materialization, position order, tensor json) | ~$3.0 | **G4 DONE**: GenEval-style (OWLv2) 60: raw 58, pup_t2i 60, pup_bbox 60. **Spec render counts: raw 29/40, pup_t2i 30/40, pup_bbox 28/40.** spec_plan_ok 11/40. g4_images (300), contact sheet, results JSONs on Hub. **GATE FIRED: t2i_bbox not better than t2i.** | Ask. |
-| 24 | 2026-10-18 03:39→04:09 | G4c students (l4x1, 29m49s, 1 attempt) | ~$1.5 | **Students measured**: parse 97/100 (0.8B), 96/100 (2B); GenEval-style overall 53/60 (0.8B), 58/60 (2B) vs raw 58, pup 60; **spec render counts 26/40 both** vs raw 29, pup_t2i 30. spec_plan_ok 12/40, 11/40. results/g4c_students.json uploaded. **No layout advantage from the students either — five-arm picture consistent.** | User decision: pocket-upsampler framing (finish image eval) or stop. Budget spent ≈ $17.5 of $25. |
+| 10-12 | 2026-10-16 | Labelling (97.0%) → filter (STOPPED 27.7% = checker artifacts; fixed; rerun 96.6% pass) | ~$1.6 | 7,575 clean rows. | SFT. |
+| 13-15 | 2026-10-16 | SFT 0.8B (2h11m a10g) + SFT 2B (1h1m a100) + text eval (301 × 3 arms) | ~$11.0 | 0.8B parse 95.0/IoU 0.470; 2B 92.7/0.529; baseline 1.0%. | GGUF. |
+| 16-18 | 2026-10-16 | GGUF Q8 (812 MB) + CPU latency (≈20.4 s/plan) + cards + deliverables verified | ~$0.1 | Complete. | Klein renders. |
+| 19 | 2026-10-17 | Klein token saga: session OAuth lacks gated-repo read scope (public 200/gated 403); user's scoped PAT verified | ~$0.1 | PAT works. | G3b. |
+| 21 | 2026-10-17 19:15 | G3b (7 attempts: OOM → pop prompt_upsampler per docstring; tokenizer wiring ×3; owlv2 kwargs ×2; fixture crash; upload 403 → dual tokens) | ~$1.3 | **DONE**: mismatch 2/3 (five_apples inverted — collection-row scale); **512² ≈ 4.6-5.0 s, 1024² ≈ 23 s (L4 fp8r)**. results/g3b_render_check.json. | G4. |
+| 22-23 | 2026-10-17 21:27 | G4a (a100; 3 attempts: 24GB OOM, missing limit_mm_per_prompt, missing flashinfer guard) + G4b (l4x1; ~6 attempts: mkdir/ids, meta re-attach, local materialization, position order, tensor json) | ~$3.0 | **G4 DONE**: GenEval-style 60: raw 58, pup_t2i 60, pup_bbox 60. Spec render counts: raw 29/40, pup_t2i 30/40, pup_bbox 28/40. spec_plan_ok 11/40. Contact sheet + results on Hub. **GATE FIRED.** | Ask. |
+| 24 | 2026-10-18 04:09 | G4c students (l4x1, 29m49s) | ~$1.5 | **Students measured**: parse 97/96 per 100; GenEval-style 53/60 (0.8B), 58/60 (2B); spec render counts 26/40 both — below raw (29) and pup_t2i (30). **Five-arm picture: no layout advantage anywhere.** results/g4c_students.json. | User decision. |
+| 25 | 2026-10-18 05:01→05:14 | Recast eval stage 1 (a100; 3 attempts: unused-import, missing flashinfer guard, AT/n bugs) — split STAGE=pup / STAGE=students when the 5.20.0 pin backtracked vllm | ~$1.2 | **421 prompts** (300 GenEval + 121 eval specs) × plan text (pup t2i + both students) persisted to data/eval_plans.json. | Render stage — user chose STOP. |
+| 26 | 2026-10-18 | Wrap-up (no compute) | $0 | **Run CLOSED as a pocket-upsampler project per user decision.** Model cards updated with the five-arm image-level table and the honest finding (no layout advantage; 512² Klein distilled near-saturated on object presence). Stage 2 (DPO) dropped as moot; render stage (421×4 renders, ≈$5) not run. Budget spent ≈ $24.2 of $25. | Deliverables final: ysharma/pocket-layout-planner-0.8B (+GGUF Q8), ysharma/pocket-layout-planner-2B, ysharma/klein-layout-plans (requests/labels/filter/results/scripts/RUN_LOG), trackio dashboard. |