Integrate desktop training fixes on the API completion branch

.agents/skills/fine-tuning/SKILL.md

@@ -56,6 +56,8 @@ Train language models with SFT, KTO, and GRPO locally or on supported cloud prov
 | Prepare verified raw-text SFT data | `python tuner.py prepare-dataset --config <config.json> --json` |
 | Check SFT loss masking (tokenizer only) | `python tuner.py doctor sft-mask --dataset-path <jsonl> --model <tokenizer>` |
 | Inspect native training host candidate | `python tuner.py doctor native-training --json` — read-only platform/Python/package metadata; see `reference/native-local-training.md` |
+| Diagnose an offline SFT worker rejection | Read the exact retained `trainer.stderr.log`; closed phase/category/milestone labels locate a boundary, not a cause or retry authority; see `reference/sft-training.md` |
+| Native MLX prepared SFT candidate | Internal exact-token batch/loss and runtime comparison only; no native launch CLI or API provider yet; see `reference/native-local-training.md` |
 | Plan a derived training image | `python scripts/qualify_derived_training_image.py plan --config Trainers/image_profiles/<profile>.yaml` |
 | Warm Space scaffold | `python3 Trainers/cloud/scripts/manage_space.py render --template vllm_warm --output-dir /tmp/my-space --base-image ghcr.io/<org>/<image>:<tag>` |
 | Warm Space deploy | `python3 Trainers/cloud/scripts/manage_space.py deploy --space-id <user>/<space> --template vllm_warm --base-image ghcr.io/<org>/<image>:<tag> --hardware a10g-small --sleep-time 3600 --var BASE_MODEL=<model>` |

.agents/skills/fine-tuning/configs/qwen35_4b_token_profile.yaml

@@ -8,7 +8,7 @@ model:
 runtime:
   provider: modal
   lock_path: tuner/execution/providers/modal/modal-runtime-v1.lock.json
-  expected_lock_sha256: 42c25ecab7848fd5002bc27711e342d1b8fac0761621af65cba3736ad9110727
+  expected_lock_sha256: 4f3fa81c9aeb9130258dec23f2b841e6dbdf2d686d1cf6b50249e743c3d496ca
 input:
   jsonl_path: /replace/with/private/prepared-dataset.jsonl
   mode: text

.agents/skills/fine-tuning/reference/grpo-training.md

@@ -198,6 +198,14 @@ end-to-end provider deadline or public cancellation qualification. The first
 executable candidate requires identical model and tokenizer revisions because
 both consume one verified local snapshot.
 
+The packaged runtime resolves `GRPOConfig` precision before loading the model.
+When the retained model dtype is `null`, it passes the effective FP16 or BF16
+dtype to Unsloth's loader. Pinned Unsloth checks that loaded dtype against
+trainer precision during `GRPOTrainer` construction; allowing the loader to
+auto-select BF16 on BF16-capable hardware while the recipe requests FP16 can
+fail at `TRAINER_SETUP`. A closed `TypeError` marker alone cannot establish
+that this was the historical cause, and it does not authorize replay.
+
 GRPO has a separate fixed 131-member offline inventory and 13-member outer
 closure. Existing SFT inventories remain unchanged. Maintenance uses the
 checked-in command, offline first because the outer closure binds its bytes:

.agents/skills/fine-tuning/reference/native-local-training.md

@@ -26,3 +26,41 @@ Resolve with the documented platform/interpreter/index and install only in a
 fresh private environment with hashes required and source builds disabled.
 This candidate does not select a model, enable a provider, or qualify Metal,
 SFT or GRPO. Keep CUDA/Modal environments and the legacy Mac checkout unchanged.
+
+## Prepared SFT worker candidate
+
+`tuner/runtime/native_mlx_sft.py` is an internal CPU-tested boundary. It consumes
+the existing shared `PreparedSFTExample` objects after host verification and
+tokenization; it does not introduce a dataset format or public execution API.
+Pass the exact prepared-input identity, tokenizer revision, retained row IDs,
+declared train/validation tuples and explicit `MLXSFTConfig` controls. The
+projection copies all tokens/labels immutably and records a deterministic digest.
+The host must still bind those rows to the authenticated source and tokenizer.
+
+No MLX text/chat loader, extra EOS, resplit, truncation, packing, tail dropping,
+or single-offset masking is allowed. Failed masking diagnostics, already
+truncated rows, attention masks other than all-ones unpadded causal rows, and
+rows without shifted target supervision reject. Arbitrary ignored token spans
+remain ignored. Training shuffles deterministically within its own split;
+validation retains every row in declared order, including partial batches.
+
+The custom hooks target pinned MLX-LM 0.32.2. One batch is one optimizer update;
+gradient accumulation and checkpointing are deliberately unavailable in this
+increment. Upstream accumulation averages microbatch means, which needs a
+separate objective review for unequal supervised-token counts. Model, LoRA,
+optimizer and learning-rate policy must be prepared explicitly by the future
+child composition; this worker does not infer or silently apply those controls.
+
+`tuner/runtime/native_mlx_runtime.py` compares host-supplied exact interpreter
+and complete distribution pins against the running candidate before worker ML
+imports. `CANDIDATE_MATCH_ONLY` means metadata and executable bytes match.
+It does not authenticate wheel files, imported module origins, Homebrew's linked
+Python framework, other native libraries, Metal, or launch authority. Retain the
+original hash lock and use the existing `uv pip freeze`, `uv pip check`,
+`shasum -a 256` and `otool -L` tools for review evidence; do not install packages
+into the shared interpreter or treat a newly observed inventory as approved.
+
+There is no native launch CLI yet. The internal worker may run only in an
+independently admitted, host-owned bounded child. The desktop plan lists the
+missing native runtime material, child lifecycle, artifact roles and API joins.
+Do not call the worker as an ad hoc substitute for that composition.

.agents/skills/fine-tuning/reference/sft-training.md

@@ -460,3 +460,45 @@ The CPU `probe-filesystem` command measures actual native writer modes and host
 readability under a reviewed finite container plan; it is not training, GPU,
 provider or cancellation qualification. Source inventory and five-role native
 verification remain closed; no default capability or UID policy is promoted.
+
+Before short-lived source preparation, run `model-preflight` with the exact
+interpreter selected for `prepare-model`; the base consumer may lack the optional
+Hub SDK. This checks only the existing SDK signature contract and creates no
+client or claim. `prepare-model` repeats it before its one-use claim. Use an
+independently verified helper environment; do not upgrade the pinned trainer.
+Create and verify necessary private source/artifact parent directories first.
+Preserve any consumed model-preparation claim after failure. New failures retain
+the existing allowlisted `model_preparation_stage`; never infer that stage or a
+cause for an older generic diagnostic, or replay an uncertain attempt.
+
+### Offline SFT worker rejection
+
+The isolated source worker retains exit code 2 on failure and emits one bounded
+diagnostic line to the existing `tracking/trainer.stderr.log`:
+`OFFLINE_SFT_WORKER_REJECTED phase=<closed> category=<closed>`.
+The phases are `BOOTSTRAP`, `CLOSURE`, `BOOTSTRAP_ORIGIN`, `ARGUMENTS`,
+`IMPORT_GUARD`, `TRAINER`, and `POSTCHECK`. Categories are `REJECTION`, `IMPORT`,
+`OS`, `VALUE`, `SYSTEM_EXIT`, and `OTHER`. During `TRAINER`, the last valid
+existing packaged-trainer hook may append `milestone=<closed>` (for example
+`TORCH_IMPORT` or `MODEL_LIBRARY_LOAD`). Unknown hooks clear that milestone.
+The fallback remains the bare rejection marker. This diagnostic does not print
+exception text, arguments, paths, environment values, or traceback frames;
+library-generated stdout/stderr is a separate surface.
+
+Use the labels to choose the next targeted reproduction. An import category
+does not identify the missing library, and a model-load milestone does not
+establish OOM. Do not interpret an older bare marker retrospectively. Preserve
+the consumed claim, roots, container and retained logs; the diagnostics never
+authorize replay. See `docs/review/offline-sft-worker-diagnostics.md` for the
+October 20 evidence boundary, CPU regressions and fresh-run prerequisites.
+
+After an intentional worker source change, check and refresh existing
+manifests with the checked-in maintainers in dependency order:
+`regenerate_offline_sft_worker_closure.py`,
+`regenerate_packaged_training_worker_closure.py`,
+`regenerate_grpo_worker_closures.py --kind offline`, then `--kind outer`,
+`regenerate_hf_parent_closure.py`, `regenerate_modal_runtime_lock.py`, and
+`regenerate_modal_inference_lock.py` (all under `scripts/`). Run each read-only
+check first, review its drift, use `--write` only for intentional hash refresh,
+then recheck. Preserve inventories and non-hash pins. Sync canonical skills
+with `.skills/scripts/sync_skill_trees.py` and verify its `--check`.

.claude/skills/fine-tuning/SKILL.md

@@ -56,6 +56,8 @@ Train language models with SFT, KTO, and GRPO locally or on supported cloud prov
 | Prepare verified raw-text SFT data | `python tuner.py prepare-dataset --config <config.json> --json` |
 | Check SFT loss masking (tokenizer only) | `python tuner.py doctor sft-mask --dataset-path <jsonl> --model <tokenizer>` |
 | Inspect native training host candidate | `python tuner.py doctor native-training --json` — read-only platform/Python/package metadata; see `reference/native-local-training.md` |
+| Diagnose an offline SFT worker rejection | Read the exact retained `trainer.stderr.log`; closed phase/category/milestone labels locate a boundary, not a cause or retry authority; see `reference/sft-training.md` |
+| Native MLX prepared SFT candidate | Internal exact-token batch/loss and runtime comparison only; no native launch CLI or API provider yet; see `reference/native-local-training.md` |
 | Plan a derived training image | `python scripts/qualify_derived_training_image.py plan --config Trainers/image_profiles/<profile>.yaml` |
 | Warm Space scaffold | `python3 Trainers/cloud/scripts/manage_space.py render --template vllm_warm --output-dir /tmp/my-space --base-image ghcr.io/<org>/<image>:<tag>` |
 | Warm Space deploy | `python3 Trainers/cloud/scripts/manage_space.py deploy --space-id <user>/<space> --template vllm_warm --base-image ghcr.io/<org>/<image>:<tag> --hardware a10g-small --sleep-time 3600 --var BASE_MODEL=<model>` |

.claude/skills/fine-tuning/configs/qwen35_4b_token_profile.yaml

@@ -8,7 +8,7 @@ model:
 runtime:
   provider: modal
   lock_path: tuner/execution/providers/modal/modal-runtime-v1.lock.json
-  expected_lock_sha256: 42c25ecab7848fd5002bc27711e342d1b8fac0761621af65cba3736ad9110727
+  expected_lock_sha256: 4f3fa81c9aeb9130258dec23f2b841e6dbdf2d686d1cf6b50249e743c3d496ca
 input:
   jsonl_path: /replace/with/private/prepared-dataset.jsonl
   mode: text

.claude/skills/fine-tuning/reference/grpo-training.md

@@ -198,6 +198,14 @@ end-to-end provider deadline or public cancellation qualification. The first
 executable candidate requires identical model and tokenizer revisions because
 both consume one verified local snapshot.
 
+The packaged runtime resolves `GRPOConfig` precision before loading the model.
+When the retained model dtype is `null`, it passes the effective FP16 or BF16
+dtype to Unsloth's loader. Pinned Unsloth checks that loaded dtype against
+trainer precision during `GRPOTrainer` construction; allowing the loader to
+auto-select BF16 on BF16-capable hardware while the recipe requests FP16 can
+fail at `TRAINER_SETUP`. A closed `TypeError` marker alone cannot establish
+that this was the historical cause, and it does not authorize replay.
+
 GRPO has a separate fixed 131-member offline inventory and 13-member outer
 closure. Existing SFT inventories remain unchanged. Maintenance uses the
 checked-in command, offline first because the outer closure binds its bytes:

.claude/skills/fine-tuning/reference/native-local-training.md

@@ -26,3 +26,41 @@ Resolve with the documented platform/interpreter/index and install only in a
 fresh private environment with hashes required and source builds disabled.
 This candidate does not select a model, enable a provider, or qualify Metal,
 SFT or GRPO. Keep CUDA/Modal environments and the legacy Mac checkout unchanged.
+
+## Prepared SFT worker candidate
+
+`tuner/runtime/native_mlx_sft.py` is an internal CPU-tested boundary. It consumes
+the existing shared `PreparedSFTExample` objects after host verification and
+tokenization; it does not introduce a dataset format or public execution API.
+Pass the exact prepared-input identity, tokenizer revision, retained row IDs,
+declared train/validation tuples and explicit `MLXSFTConfig` controls. The
+projection copies all tokens/labels immutably and records a deterministic digest.
+The host must still bind those rows to the authenticated source and tokenizer.
+
+No MLX text/chat loader, extra EOS, resplit, truncation, packing, tail dropping,
+or single-offset masking is allowed. Failed masking diagnostics, already
+truncated rows, attention masks other than all-ones unpadded causal rows, and
+rows without shifted target supervision reject. Arbitrary ignored token spans
+remain ignored. Training shuffles deterministically within its own split;
+validation retains every row in declared order, including partial batches.
+
+The custom hooks target pinned MLX-LM 0.32.2. One batch is one optimizer update;
+gradient accumulation and checkpointing are deliberately unavailable in this
+increment. Upstream accumulation averages microbatch means, which needs a
+separate objective review for unequal supervised-token counts. Model, LoRA,
+optimizer and learning-rate policy must be prepared explicitly by the future
+child composition; this worker does not infer or silently apply those controls.
+
+`tuner/runtime/native_mlx_runtime.py` compares host-supplied exact interpreter
+and complete distribution pins against the running candidate before worker ML
+imports. `CANDIDATE_MATCH_ONLY` means metadata and executable bytes match.
+It does not authenticate wheel files, imported module origins, Homebrew's linked
+Python framework, other native libraries, Metal, or launch authority. Retain the
+original hash lock and use the existing `uv pip freeze`, `uv pip check`,
+`shasum -a 256` and `otool -L` tools for review evidence; do not install packages
+into the shared interpreter or treat a newly observed inventory as approved.
+
+There is no native launch CLI yet. The internal worker may run only in an
+independently admitted, host-owned bounded child. The desktop plan lists the
+missing native runtime material, child lifecycle, artifact roles and API joins.
+Do not call the worker as an ad hoc substitute for that composition.

.claude/skills/fine-tuning/reference/sft-training.md

@@ -460,3 +460,45 @@ The CPU `probe-filesystem` command measures actual native writer modes and host
 readability under a reviewed finite container plan; it is not training, GPU,
 provider or cancellation qualification. Source inventory and five-role native
 verification remain closed; no default capability or UID policy is promoted.
+
+Before short-lived source preparation, run `model-preflight` with the exact
+interpreter selected for `prepare-model`; the base consumer may lack the optional
+Hub SDK. This checks only the existing SDK signature contract and creates no
+client or claim. `prepare-model` repeats it before its one-use claim. Use an
+independently verified helper environment; do not upgrade the pinned trainer.
+Create and verify necessary private source/artifact parent directories first.
+Preserve any consumed model-preparation claim after failure. New failures retain
+the existing allowlisted `model_preparation_stage`; never infer that stage or a
+cause for an older generic diagnostic, or replay an uncertain attempt.
+
+### Offline SFT worker rejection
+
+The isolated source worker retains exit code 2 on failure and emits one bounded
+diagnostic line to the existing `tracking/trainer.stderr.log`:
+`OFFLINE_SFT_WORKER_REJECTED phase=<closed> category=<closed>`.
+The phases are `BOOTSTRAP`, `CLOSURE`, `BOOTSTRAP_ORIGIN`, `ARGUMENTS`,
+`IMPORT_GUARD`, `TRAINER`, and `POSTCHECK`. Categories are `REJECTION`, `IMPORT`,
+`OS`, `VALUE`, `SYSTEM_EXIT`, and `OTHER`. During `TRAINER`, the last valid
+existing packaged-trainer hook may append `milestone=<closed>` (for example
+`TORCH_IMPORT` or `MODEL_LIBRARY_LOAD`). Unknown hooks clear that milestone.
+The fallback remains the bare rejection marker. This diagnostic does not print
+exception text, arguments, paths, environment values, or traceback frames;
+library-generated stdout/stderr is a separate surface.
+
+Use the labels to choose the next targeted reproduction. An import category
+does not identify the missing library, and a model-load milestone does not
+establish OOM. Do not interpret an older bare marker retrospectively. Preserve
+the consumed claim, roots, container and retained logs; the diagnostics never
+authorize replay. See `docs/review/offline-sft-worker-diagnostics.md` for the
+October 20 evidence boundary, CPU regressions and fresh-run prerequisites.
+
+After an intentional worker source change, check and refresh existing
+manifests with the checked-in maintainers in dependency order:
+`regenerate_offline_sft_worker_closure.py`,
+`regenerate_packaged_training_worker_closure.py`,
+`regenerate_grpo_worker_closures.py --kind offline`, then `--kind outer`,
+`regenerate_hf_parent_closure.py`, `regenerate_modal_runtime_lock.py`, and
+`regenerate_modal_inference_lock.py` (all under `scripts/`). Run each read-only
+check first, review its drift, use `--write` only for intentional hash refresh,
+then recheck. Preserve inventories and non-hash pins. Sync canonical skills
+with `.skills/scripts/sync_skill_trees.py` and verify its `--check`.

.skills/fine-tuning/SKILL.md

@@ -56,6 +56,8 @@ Train language models with SFT, KTO, and GRPO locally or on supported cloud prov
 | Prepare verified raw-text SFT data | `python tuner.py prepare-dataset --config <config.json> --json` |
 | Check SFT loss masking (tokenizer only) | `python tuner.py doctor sft-mask --dataset-path <jsonl> --model <tokenizer>` |
 | Inspect native training host candidate | `python tuner.py doctor native-training --json` — read-only platform/Python/package metadata; see `reference/native-local-training.md` |
+| Diagnose an offline SFT worker rejection | Read the exact retained `trainer.stderr.log`; closed phase/category/milestone labels locate a boundary, not a cause or retry authority; see `reference/sft-training.md` |
+| Native MLX prepared SFT candidate | Internal exact-token batch/loss and runtime comparison only; no native launch CLI or API provider yet; see `reference/native-local-training.md` |
 | Plan a derived training image | `python scripts/qualify_derived_training_image.py plan --config Trainers/image_profiles/<profile>.yaml` |
 | Warm Space scaffold | `python3 Trainers/cloud/scripts/manage_space.py render --template vllm_warm --output-dir /tmp/my-space --base-image ghcr.io/<org>/<image>:<tag>` |
 | Warm Space deploy | `python3 Trainers/cloud/scripts/manage_space.py deploy --space-id <user>/<space> --template vllm_warm --base-image ghcr.io/<org>/<image>:<tag> --hardware a10g-small --sleep-time 3600 --var BASE_MODEL=<model>` |

.skills/fine-tuning/configs/qwen35_4b_token_profile.yaml

@@ -8,7 +8,7 @@ model:
 runtime:
   provider: modal
   lock_path: tuner/execution/providers/modal/modal-runtime-v1.lock.json
-  expected_lock_sha256: 42c25ecab7848fd5002bc27711e342d1b8fac0761621af65cba3736ad9110727
+  expected_lock_sha256: 4f3fa81c9aeb9130258dec23f2b841e6dbdf2d686d1cf6b50249e743c3d496ca
 input:
   jsonl_path: /replace/with/private/prepared-dataset.jsonl
   mode: text

.skills/fine-tuning/reference/grpo-training.md

@@ -198,6 +198,14 @@ end-to-end provider deadline or public cancellation qualification. The first
 executable candidate requires identical model and tokenizer revisions because
 both consume one verified local snapshot.
 
+The packaged runtime resolves `GRPOConfig` precision before loading the model.
+When the retained model dtype is `null`, it passes the effective FP16 or BF16
+dtype to Unsloth's loader. Pinned Unsloth checks that loaded dtype against
+trainer precision during `GRPOTrainer` construction; allowing the loader to
+auto-select BF16 on BF16-capable hardware while the recipe requests FP16 can
+fail at `TRAINER_SETUP`. A closed `TypeError` marker alone cannot establish
+that this was the historical cause, and it does not authorize replay.
+
 GRPO has a separate fixed 131-member offline inventory and 13-member outer
 closure. Existing SFT inventories remain unchanged. Maintenance uses the
 checked-in command, offline first because the outer closure binds its bytes: