Phase 2C.3: ablate SSD-1B prompts and sampling parameters - #32

Summary

This draft PR adds a bounded, reproducible prompt and sampling-parameter ablation for the existing SSD-1B plus project-authored flat-cartoon LoRA route. It is an evaluation only; production Model Package, Workflow Pack, prompt profile, sampler, scheduler, steps, Provider Hub, and teacher UI are unchanged.

Experiment

• Fixed SSD-1B revision 60987f37e94cd59c36b1cba832b9f97b57395a10 and existing LoRA SHA-256 03dd1968d37e7fdb5f1d86a5ad3ac09a961acd37b39353aed27f398e6266f785.
• Stage 1: P0/P1/P2/P3/P4 prompt-only, A0 fixed, 4 cases × 2 seeds = 40 real MPS images.
• Stage 2: A0/A1/A2/A3 parameter-only, P0 fixed, 4 cases × 2 seeds = 32 real MPS images.
• Stage 3 was intentionally not run: neither layer produced a stable semantic improvement, so the protocol’s confirmation gate did not pass.
• Every image has fixed request/plan/profile/case/seed metadata, PNG SHA-256, provenance, runtime/model/LoRA identity, and pending teacher-review fields. No automatic better_than_baseline or teaching_usable values are written.

Real result

• 72/72 technical successes; 0 failures; 0 near-solid outputs.
• P1 had a local exact-three-person-looking success in one inspected seed, but still failed the one-person wave and classroom-pair cases. P2/P3/P4 changed composition/style without stable count/action improvement.
• A1/A2/A3 did not consistently fix count, action, or classroom relation; A1/A2/A3 were materially slower than A0. No clearly better prompt+parameter combination was confirmed.
• The evidence supports the current hypothesis that SSD-1B semantic control is the main bottleneck for these precise teaching cases. No production update is recommended.

Artifacts

Generated model, LoRA, images, cache, runtime, reports, and review package remain under ignored runtime/phase2c3-prompt-parameter-ablation/; none are committed. The tracked report is docs/phase2c3-prompt-parameter-ablation.md; the ignored HTML comparison package has 72 images and 72 pending teacher reviews.

Validation

• Focused Phase 2C.3 contract tests: 2 passed.
• Full API suite: 647 passed, 3 skipped.
• TypeScript noEmit and Vite production build: passed.
• Web state tests: passed.
• Playwright: 14 passed.
• Ruff and git diff --check: passed.

This PR intentionally remains Draft pending teacher review and any future model-level decision.