Release v0.3.0

[ai-hpc]

Cuts v0.3.0 — bumps pyproject.toml (0.2.0 → 0.3.0) and rolls [Unreleased] into [0.3.0] — 2026-07-26. generator_version stays 0.3.0 (the new preferences_sha256 manifest field is optional/additive; existing bundles verify unchanged).

Highlights

Dataset-quality + training-signal release (PRs #35, #51–#56):

• Closes the validator-hack / arch-mismatch pattern from the merged-data review — portable-code prompts (Steer teacher prompts to portable, non-hacky kernels (dataset quality) #51), retarget prompts to the actual validating GPU (Retarget prompts to the actual validating GPU architecture (root cause) #54), non-blocking guard-removal flag (Flag repairs that pass by deleting a hardware guard (dataset quality) #53).
• Denser preference signal — DPO correctness pairs (DPO exporter: correctness preference pairs (verified vs failed kernel) #52), now sha256-pinned + attested in the manifest (Pin DPO preference file in the bundle manifest (attested, secret-scanned) #56); token-efficient diff-style repair export (Diff-style SFT export for repair turns (cut token waste) #55).
• Broader corpus — license-gated external task seeds, PyTorch-only, re-proved from scratch (Import KernelBook external corpora as verifiable task seeds #35).

385 tests pass, ruff clean.

Compatibility

Additive only: the manifest gains an optional, attested preferences_sha256; bundles without it verify exactly as before. No stable_keys change, no generator_version bump.

After merge, tag v0.3.0 on the merge commit + publish the GitHub Release.