Context

PR #962 designed the domain-specific embedding fine-tuning pipeline and shipped EmbeddingFineTuneConfig as a config stub. The config validates its invariants (required fields, path safety) but the Mem0 adapter does not consume it -- fine_tune.enabled=True raises "not yet supported" via a fail-fast validator.

This issue implements the actual pipeline.

Requirements

1. Wire checkpoint lookup into Mem0 adapter

When fine_tune.enabled=True and a valid checkpoint_path exists:

• Load the fine-tuned model instead of the base model at backend initialization
• If checkpoint doesn't exist at the path, log warning and fall back to base model
• Remove the fail-fast validator on Mem0EmbedderConfig._reject_unimplemented_fine_tune

2. Fine-tuning CLI command

Implement the 4-stage offline pipeline as a CLI command (or Python script):

1. Synthetic data generation -- LLM generates query-document pairs from org documents
2. Hard negative mining -- base model embeds all passages, selects top-k confusing negatives
3. Contrastive fine-tuning -- biencoder training with InfoNCE loss (tau=0.02, 3 epochs, lr=1e-5)
4. Deploy -- save checkpoint, update config to point to fine-tuned model

Pipeline design is documented in docs/reference/embedding-evaluation.md.

3. Update docs

• Remove "not yet implemented" notes from EmbeddingFineTuneConfig docstring, design spec, and reference page
• Update docs to reflect actual behavior instead of planned behavior

Acceptance Criteria

• fine_tune.enabled=True with valid checkpoint loads fine-tuned model
• Missing checkpoint falls back to base model with logged warning
• Fail-fast validator removed
• CLI command runs the 4-stage pipeline end-to-end
• Pipeline works on single GPU with no manual annotation
• Tests cover checkpoint lookup (found, not found, disabled)
• Docs updated to reflect implemented behavior

References

• PR docs: LMEB embedding evaluation + CSP accepted risk (#695, #925) #962 -- EmbeddingFineTuneConfig, pipeline design, LMEB evaluation
• src/synthorg/memory/backends/mem0/config.py -- config model + fail-fast guard
• docs/reference/embedding-evaluation.md -- pipeline design and expected improvements
• NVIDIA domain-specific fine-tuning blog