feat(recipes): add smollm3-3b-sft recipe (#271)

[Srinivasan8888]

Closes #271.

What

Adds a ready-made recipe for SmolLM3-3B — a genuinely missing popular model. The catalog had SmolLM2 in three sizes (135M / 360M / 1.7B) but no SmolLM3, which is a strong, widely-used small/edge model.

smollm3-3b-sft is a pure declarative RecipeMeta mirroring the other small SFT recipes (LoRA r8/α16, target_modules: auto, 8-bit, batch_size: auto, 2048 ctx) — no GPU, fully CI-validated.

Count sync

Adding a recipe moves the catalog 154 → 155, so this bumps every site the test_recipe_count_is_synced guard enforces:

• the # Recipe catalog (N recipes) comment in catalog.py
• CONTRIBUTING.md, docs/commands.md, and the two docs/serving-and-export.md count lines
• the four test_catalog_size_is_154 → _155 checkpoints (test_recipes.py, test_v07124.py, test_v07130.py, test_v07132.py), with a history line added to the running changelog docstring.

Acceptance criteria

• soup recipes show smollm3-3b-sft prints valid YAML
• soup recipes use smollm3-3b-sft writes a soup.yaml that load_config_from_string loads without error
• ruff check clean
• pytest -k "recipe or catalog or sync" passes (286 passed locally)
• Recipe count updated consistently across docs + checkpoints

[MakazhanAlpamys]

The catalog and documentation changes are present, but this head is not mergeable yet. Issue #271 requires the model to be claimed in an issue comment before opening the PR; I found no SmolLM3 claim from the author in the issue history. The branch is also DIRTY/CONFLICTING against current main, and gh pr checks 582 reports no checks.

I independently ran the recipe/catalog suites (438 passed, 1 skipped) and Ruff. The required mutation exposed a coverage gap: changing the new recipe model from HuggingFaceTB/SmolLM3-3B to HuggingFaceTB/SmolLM3-WRONG still left the focused new-recipe/count tests green (8 passed), so the exact model identity is not load-bearing. Please add an exact model-id assertion, add the required claim comment, rebase onto current main, and rerun the full checks before requesting re-review.

[Srinivasan8888]

All three addressed on the current head 69cc65a8:

• Exact model-id assertion. Added TestIssue271SmolLM3Recipe (mirrors the Add ready-made recipe: qwen3.5-4b-pretrain (continued pre-training, Base checkpoint) #278/Add ready-made recipe: glm-5.1-dpo (DPO variant) #280 recipe-test pattern) pinning HuggingFaceTB/SmolLM3-3B on both RecipeMeta.model and the parsed config.base, plus a recipes show/use check. Verified the mutation you flagged now fails: changing the id to SmolLM3-WRONG fails test_recipe_loads_with_exact_model_id.
• Rebased onto current main. Main had moved 154 → 158 recipes; rebased and re-derived the count as 159 across all five doc sites and the four test_catalog_size_is_158 → _159 checkpoints. PR now reports MERGEABLE.
• Claim comment. Posted on Add a ready-made recipe for a popular model (good first issue) #271.

Local: 455 passed, 1 skipped; Ruff clean. The remaining pending checks are the CI approval gate (fork PRs sit action_required until a maintainer approves the run).

[MakazhanAlpamys]

Re-review of current head 69cc65a against issue #271: the required claim is now present in the issue, the exact SmolLM3 model ID is asserted for both the recipe and parsed config, and the branch is rebased onto current main. No tests were deleted; docs and changelog are present. I independently ran the recipe/catalog suite (445 passed, 1 skipped) with PYTHONPATH=src and Ruff; mutating the model ID caused the exact-ID regression tests to fail. Approval is contingent on all 14 current-head checks completing successfully before merge.

[MakazhanAlpamys]

Thank you @Srinivasan8888 for adding a focused, catalogued SmolLM3-3B SFT recipe with an exact model-ID regression assertion. I independently ran 445 focused recipe tests with 1 skip, Ruff, the full required CI matrix, and a mutation replacing the model ID; the mutation was caught by two tests.