Track 2: LoRA lite slice (film LoRAs only)

[WalksWithASwagger]

Port only the LoRA capability the film uses, on top of Rafiki's existing lib/training.py (already dry-run Replicate FLUX LoRA + cost + manifest).

Scope (lite)

• train a character/style LoRA from an album (e.g. @lex, Mage, Mayumi, Kevin)
• generate an immediate sample gallery after training (review signal)

Archived in the studio (NOT ported — dormant since April; revisit only if reactivated)

• 6-variant experiment suite
• finale / preference tournament
• onboard_client client-onboarding flow
• multi-subject rating viewers

Rationale: git history shows the portrait-client business is dormant (0 commits/30d). Don't rebuild cold capability.

Part of epic #268 (value-first retirement). Pattern: lib/providers/replicate_provider.py + lib/video_jobs.py. Dry-run-first; --execute operator-only spend. Audit: meta/audits/2026-06-28-alex-samuel-consolidation-audit.md.

[WalksWithASwagger]

Backlog audit: the capability is only partial and remains potentially useful. Current main can plan/launch Replicate FLUX LoRA training from an input-images URL, but it does not package a local album or generate the post-training sample gallery requested here. Before agent implementation, a maintainer must fix the supported album input contract, sample prompt set, output/review contract, and provider-spend checkpoint; otherwise this is too ambiguous for agent:ready.