Steer teacher prompts to portable, non-hacky kernels (dataset quality)

[ai-hpc]

Improves miner dataset quality at the generation source — a small, prompt-only change with no trust/hash-boundary impact that helps every miner immediately.

Problem (found in the live merged data)

Reviewing gittensor-model-hub/sparkproof-mining (1,029 rows) surfaced a validator-hack repair pattern: a kernel prompted "for Hopper SM90" adds a hard SM/compute-capability assertion to its self-test, which then fails on the actual SM120 (Blackwell) validation GPU:

validator: RuntimeError: This test targets Hopper SM90, found SM120
assistant: "The false architecture rejection is removed … does not require an exact-SM90 guard"

The "successful" repair passed by deleting the guard, not fixing the kernel — teaching a questionable debugging habit and burning a repair round. Root cause: the prompt induces an arch assertion that the real hardware can't satisfy.

Fix (4 prompt edits)

• default_system + IMPLEMENTATION_SECTION — teachers must write portable code with no device / compute-capability / SM / architecture asserts (SparkProof attests the GPU separately; the code must run on whatever GPU validates it). Kills the failure class at the source.
• validator_feedback_content + _repair_prompt — the repair step must fix the root cause, not pass by deleting assertions, weakening the torch.allclose test, or loosening tolerances. Discourages the validator-hack directly.

Impact

• Fewer wasted repair rounds (the arch-guard failure largely disappears).
• Repairs that do happen fix real bugs instead of learning to delete checks.
• Works even when a task's arch label doesn't match the runtime GPU, because kernels stop embedding hard guards.

Tests

354 pass, ruff clean. New test_prompts_discourage_arch_guards_and_validator_hacks pins the guidance so it can't silently regress.

Independent of the Kimi-K3 teacher swap (PR #50) — this branches from main and only touches prompt text. A natural follow-up (separate PR) is a release-gate/anti-cheat check that flags repairs which pass by removing an assertion.