Add LoRA artifact, resume, merge/export, and callback drift canaries

[Zigfreidish]

This was generated by AI during triage.

Parent

• Track post-closure #43 advanced fine-tuning follow-up notes #1519
• Source milestone: Milestone plan: long-context and advanced fine-tuning #43
• Related provenance/export trackers: M-Courtyard roadmap: P2.2 Training Monitor and Adapter History #1501, M-Courtyard roadmap: U2.2.2 Persist adapter provenance, loss series, and notes #1503, M-Courtyard roadmap: P3.2 Post-Export Smoke and Diagnostics #1508

What to build

Add focused correctness canaries for LoRA training artifacts, resume metadata, merge/export drift, callback/API drift, and teacher-forced round-trip behavior.

This slice should prove that saved artifacts remain loadable and semantically aligned with the base tokenizer and training API, without duplicating generic M-Courtyard provenance/export smoke work.

Acceptance criteria

• Tokenizer EOS preservation is checked and recorded for saved adapters.
• Resume metadata captures tokenizer/config/processor state needed to restart training safely.
• Adapter checkpoint canaries detect missing base config, missing tokenizer fields, and incompatible auxiliary module state.
• Merge/export canaries detect drift before publishing an unusable artifact.
• Callback/API drift evidence and teacher-forced round-trip smoke receipts are saved with the training run.

Governing docs

• docs/runbooks/phase-8-lora-adapter-workflow.md
• New implementation plan should be created before code changes.
• Related: M-Courtyard roadmap: P2.2 Training Monitor and Adapter History #1501, M-Courtyard roadmap: U2.2.2 Persist adapter provenance, loss series, and notes #1503, M-Courtyard roadmap: P3.2 Post-Export Smoke and Diagnostics #1508

Expected receipt fields

• source_eos_token
• saved_eos_token
• tokenizer_config_path
• base_config_present
• processor_resume_mode
• aux_modules_restored
• merge_export_canary_result
• callback_api_drift_result
• completion_loss
• round_trip_passed
• grad_norm

Focused tests

• Tokenizer/processor EOS fixtures.
• Processor-backed resume fixture.
• CPU merge/export static canaries.
• Callback arity drift test.
• Teacher-forced round-trip smoke test.

Blocked by

None - can start immediately. Coordinate with #1503/#1508 when artifact provenance or export smoke surfaces overlap.

[Akane-CN]

Watch-log follow-up — base-model export needs explicit merge/save-method canaries.

• Design principle: LoRA merge/export should not rely on library defaults whose meaning can drift between releases. The export receipt must record the save method and prove that the produced base artifact is loadable before publishing or marking the run reusable.
• Melix-specific goal: extend merge/export canaries so base-model export distinguishes adapter-only save, merged 16-bit save, quantized export, and hub/upload paths, with the selected save method captured in run evidence.
• Measurable target: missing or wrong save-method selection fails a pre-publish canary; local export and upload/export paths produce identical declared save method, tokenizer, and config receipts; a post-export smoke load succeeds before the artifact is promoted.
• Next executable slice: add a fixture that exports a tiny LoRA run through local and publish-style paths, asserts saveMethod=merged_16bit when producing a merged base artifact, and blocks promotion on mismatch.

[Akane-CN]

Watch-log follow-up — quantized-base LoRA needs train/resume/export canaries, not just admission flags.

• Design principle: training against a quantized base is only product-safe when module targeting, frozen/base parameter handling, optimizer state, resume metadata, and export layout are validated end to end. A successful training loop is not enough if the saved adapter or merged artifact cannot be reloaded under the same quantization contract.
• Melix-specific goal: extend LoRA artifact canaries so QLoRA and quantized-base fine-tuning record the base quantization profile, target-module rewrite decision, tokenizer/config snapshot, resume compatibility, and post-export load smoke.
• Measurable target: one minimal quantized-base train→resume→export→activate fixture should pass without dense-base fallback, preserve adapter metadata, and fail closed when an unsafe target module or missing base config is used.
• Next executable slice: add a synthetic quantized-base manifest to the LoRA test harness, then require the export receipt to include base_quant_profile, adapter_target_modules, resume_config_hash, and post_export_load_status.

[Akane-CN]

Watch-log follow-up — adapter save fallback should be a product canary, not a best-effort trainer side effect.

• Design principle: training completion is not enough; the selected adapter save path must prove that weights, tokenizer/processor state, and fallback serialization are loadable after the trainer exits. If the preferred save API changes or fails, fallback behavior needs an explicit receipt and post-save smoke test.
• Melix-specific goal: extend LoRA artifact/resume/export canaries with preferred_adapter_save_path, fallback_adapter_save_path, fallback_reason, saved_weight_file_set, processor_snapshot_present, and post_save_load_smoke_result.
• Measurable target: SFT and preference-style trainer fixtures force the preferred save branch and the fallback branch; both must produce a reloadable adapter with matching tokenizer/config receipts, and a missing fallback artifact fails the run before activation/export.
• Next executable slice: add a trainer double whose save method raises after partial output, then verify cleanup, fallback serialization, and post-save reload evidence in one minimal LoRA run.

[Akane-CN]

Watch note: checkpoint resume fallback receipts

• Design principle: adapter resume should prefer the accelerated loader only when it can actually restore the checkpoint; otherwise it must fall back to the standard loader with an explicit receipt rather than failing late or mutating model state ambiguously.
• Melix-specific goal: strengthen LoRA resume/export canaries so a saved adapter records whether the accelerator path loaded it, why fallback was needed, and whether tokenizer/config/auxiliary module state remained compatible after fallback.
• Measurable target: add a resume smoke where an unsupported accelerated checkpoint path returns a typed accelerated_resume_unsupported reason, standard loading succeeds, and the saved receipt includes resume_loader, fallback_reason, base_config_present, tokenizer_config_path, and round_trip_passed.
• Next executable slice: add one fixture to Add LoRA artifact, resume, merge/export, and callback drift canaries #1529 that deliberately marks an adapter as accelerator-preferred but unsupported for the current backend, then verifies fallback loading plus post-resume generation before merge/export canaries run.