fix(run): key the resume cache on steps + guidance

[pedropaf]

Problem

modl run skips sub-jobs whose output is already on disk, but the completion key was only (prompt, seed, model). Re-running the same prompt/seed at a different guidance or step count matched the old key, so the run skipped generation and handed back the stale image — while printing ✓ complete and exiting 0.

Any guidance/steps sweep was silently corrupted. Same failure class as the silently-dropped configs in #145: it looks accepted, nothing errors, and the run is quietly not what was asked for.

Found the hard way: a guidance 3.0 → 2.5 re-run returned the 3.0 images and reported success. Only caught by diffing the output-image count.

Fix

• Completion key is now (prompt, seed, base_model, steps, guidance).
• The generate spec is built before the skip check, so the resolved steps/guidance (after fast/lightning/default precedence) go into the key — then reused for execution, no double build.
• guidance is f32 (neither Eq nor Hash), so it is quantized to an integer (×1000, rounded) via quantize_guidance, stable across the YAML-sidecar round-trip.
• Edit steps have no skip path — no change there.

Tests

• completion_index_keys_on_steps_and_guidance — distinct settings → distinct keys; unseen settings miss.
• quantize_guidance_is_stable_and_distinct.
• Verified end-to-end with SD-1.5: identical guidance skips, changed guidance regenerates (sidecar confirms the new value).

🤖 Generated with Claude Code