feat(hip2hip): inject generic task contract at framework level

[mehdi-saeedi]

Summary

Follow-up to #35 (review thread). @irvineoy asked that the four generic task constraints — preserve kernel names/signatures, keep launch/config interface compatible, keep the build/test interface intact, handle shared-memory launch sizing — survive the cleanup of the MI300X-specific prompt that previously lived in tasks/hip2hip/others/points_in_boxes/config.yaml. Per the agreed plan, this PR adds them at the framework level so every hip2hip task gets the same contract, with no per-task YAML duplication.

No-touch guarantee

This PR is strictly additive for task_type == 'hip2hip' only. The following task types are untouched by this change — they render the exact same prompt before and after:

• triton2triton (incl. rocmbench/*, geak_eval/*, repository/rocprim/*)
• torch2hip
• cuda2hip
• instruction2triton
• repository

The injection sites in both src/prompt_builder.py and agents/geak_v3/launch_agent.py are gated by an if task_type_name == 'hip2hip': check; nothing else in either file moves.

Other things this PR does not touch:

• No task YAML edits (no config.yaml modified).
• No new dependencies.
• No changes to evaluator, scorer, task discovery, runner, or worktree handling.
• No changes to default_cheatsheet.yaml, the architecture cheatsheets, or the GPU arch-precheck directive.
• The cheatsheet injection added in Enable RDNA4 (gfx1201) HIP support: arch cheatsheets, prompt builder, preprocessing #35 is preserved verbatim — the contract is appended alongside it, not in place of it.

What changes

A new ### Task Contract (Generic) section is rendered into the agent prompt for every task_type == 'hip2hip' task. The contract enumerates:

1. Preserve kernel function names and signatures — listed by name from the task's own target_kernel_functions field so the agent sees exactly which symbols are part of the contract.
2. Keep the launch / configuration interface compatible — grid/block dims, stream usage, host-side launch helpers, Python bindings, extern "C" shims.
3. Output must remain directly compilable and runnable — the task's existing compile_command / correctness_command / performance_command must succeed against the modified source with no edits to the test runner and no extra build flags.
4. Handle shared-memory launch sizing correctly when __shared__ / dynamic LDS allocations are introduced.

The contract is injected from two places, mirroring the section ordering used by each prompt path:

• src/prompt_builder.py::prompt_builder — regular agents. Inserted right after the ### Source Code section.
• agents/geak_v3/launch_agent.py — GEAK-v3 (uses simple_prompt_builder, bypasses the regular path). Mirrors the cheatsheet injection added in Enable RDNA4 (gfx1201) HIP support: arch cheatsheets, prompt builder, preprocessing #35.

Why framework-level (not per-task prompt.instructions)

Across the 36 current hip2hip configs:

• 24 gpumode/* already have a generic role-preamble in prompt.instructions (no contract bullets).
• 12 others/* have instructions: null and cheatsheet: null.

Adding the contract per-config would require touching 36 YAMLs and inevitably drift over time. Hosting it in one function keeps the contract in lockstep across tasks, single-source-of-truth, and architecture-neutral by construction.

Diffstat

```text
agents/geak_v3/launch_agent.py | +17 -2
src/prompt_builder.py | +11 -1
src/prompts/task_type.py | +66
3 files, +94 / -3
```

Test plan

• All 36 hip2hip configs render with the contract present and the correct kernel name(s) listed.
• Sampled triton2triton configs (5/5) do NOT receive the contract (correctly scoped).
• Regular src/prompt_builder.py::prompt_builder path verified.
• GEAK-v3 agents/geak_v3/launch_agent.py path verified (logs incl. hip2hip contract when fired).
• End-to-end run on a representative hip2hip task confirms the agent receives the contract in task_prompt.md.
• Spot-check that an agent attempting to rename a target kernel function sees the contract section in its initial prompt.
• Spot-check a triton2triton run pre/post this PR — task_prompt.md is byte-identical.

[irvineoy]

Thanks!