Add PR template and PR style rules

.github/pull_request_template.md

@@ -0,0 +1,36 @@
+<!-- Style rules: docs/PR_STYLE.md. Delete any section that adds nothing. -->
+
+## Purpose
+
+<!-- 2 to 4 sentences: the problem, then what this PR does. -->
+
+## How it works
+
+<!-- Mermaid diagram for flows. `text` file tree for new or changed files. -->
+
+```text
+path/
+└── file.py    # new: what it does
+```
+
+## Changes
+
+<!-- Numbered, grouped by area. Each item: what, Effect:, Before:. -->
+
+1. **What changed.**
+   Effect: what is different now.
+   Before: what happened before.
+
+## Usage
+
+```bash
+# exact commands or config
+```
+
+## Verification
+
+<!-- Real commands and real numbers. Never invented. -->
+
+## Notes
+
+<!-- Merge order, dependencies on other PRs, what is not done, decisions for the reviewer. -->

docs/CODE_DESIGN.md

@@ -0,0 +1,49 @@
+# Code design
+
+These rules decide where code goes and how much code a change adds. They come from code review. `docs/PR_STYLE.md` gives the rules for PR text.
+
+## Put code where its concern lives
+
+- Group code by domain, not by workflow step. A package called after a pipeline stage (for example `verify/`) with unrelated modules inside it is not a design.
+- Extend the package that already has the concern before you add a new one:
+  - images and containers go in `docker/`;
+  - task rendering and `task.toml` go in `harbor_adapter/`;
+  - LLM calls and prompt extraction go in `agents/`;
+  - pipeline steps go in `runners/`;
+  - pushing data to a registry or dataset goes in `publish/`.
+- Add a top-level package only for a new domain with several modules, for example `measurement/` for timing, the noise rule and the measure gate.
+- Keep code that runs in different places apart:
+  - code that runs inside a task container goes with the templates;
+  - long-running host services go in their own module;
+  - the library holds the logic;
+  - the CLI only parses arguments and calls the library.
+
+## Change the source, not the output
+
+- Fix behavior where it starts. If new tasks need a different prompt, render it that way: add a runner class (for example `RenderShortProblemsRunner`) and make it the default. Do not add a tool that edits rendered files after the fact.
+- Extend the existing patterns (a runner subclass, a publisher next to the other publishers) instead of a parallel copy.
+- Before you add a step, check whether the existing pipeline already does it. For example, publishing is a step of `fc-data`.
+
+## Write less code
+
+- Apply the ponytail rules (https://github.com/DietrichGebert/ponytail/blob/main/skills/ponytail/SKILL.md) to every component: the smallest change that fully solves the task, deletion before addition, no option or layer that nobody asked for.
+- Before you write a helper, check in this order:
+  1. this repository;
+  2. the standard library;
+  3. an installed dependency;
+  4. a maintained external package.
+- Generic helpers (JSONL files, shell calls, free disk space) are not main modules. Put them in `utils/` or in the module that uses them.
+- One-off scripts (reports, counts, operational runs) are not library code. Put them in `scripts/`, or in the repository that uses them, or delete them after use.
+- Archive old code and documents when nothing uses them. Do not keep them for history; git keeps the history.
+
+## Make it work on every host
+
+- A service or script must not depend on one machine: no fixed home paths, host names or local mounts. It must run the same way on an AWS host as on the local machine.
+- Settings come from arguments or environment variables with clear names, documented in the module that uses them.
+
+## Review checklist
+
+- Does each new module belong to the package that holds its concern?
+- Is any new module a utility, a one-off script or a duplicate of something that exists?
+- Does the change fix the source of the behavior, or edit its output?
+- Can a smaller change do the same job?

docs/PR_STYLE.md

@@ -0,0 +1,53 @@
+# PR style
+
+The PR template (`.github/pull_request_template.md`) gives the sections. This page gives the rules for PR text. `docs/CODE_DESIGN.md` gives the rules for the code.
+
+## Language
+
+- Use ASD-STE100 Simplified Technical English.
+- Use short active sentences in the present tense: 20 words maximum for steps, 25 for other text.
+- Write one instruction per sentence. Use simple words. Do not use idioms or needless hyphens.
+- A reader must understand the PR in 30 seconds.
+- Do not use these words: "contract", "fail loud", "fail closed", "straddle", "owns", "surface", "guard", "load-bearing", "robust", "seamless", "footgun". Say what the code does.
+- Do not write host names. Write "training node", "inference node", "this machine" or "the second machine".
+- Do not write secrets. Write credentials only as `${VAR}`.
+
+## Sections, in this order
+
+Delete a section if it adds nothing.
+
+1. **Purpose**: 2 to 4 sentences. The problem, then what the PR does.
+2. **How it works**: a Mermaid diagram for flows. A `text` file tree for new or changed files. The Mermaid must parse.
+3. **Changes**: a numbered list, grouped by area. Each item gives what changed, `Effect:` and `Before:`.
+4. **Usage**: a code block with the exact commands or config.
+5. **Verification**: the real commands and the real numbers. Never invent numbers.
+6. **Notes**: merge order, dependencies on other PRs, what is not done, decisions for the reviewer.
+
+Write math as GitHub math: `$…$` inline and `$$…$$` for display.
+
+Do not add a footer or any AI attribution.
+
+## Code in the PR
+
+- Use the minimum number of lines.
+- Comments: 1 or 2 lines. Say what the code does and why. No narrative.
+- Write one test for each real behavior.
+- Do not add co-author or session trailers to commits.
+
+## Example
+
+````markdown
+## Purpose
+
+The reward counts a timeout as a failure. A slow task then gets a negative reward. This PR gives a timeout a reward of 0.
+
+## Changes
+
+1. **Timeout reward is 0.**
+   Effect: a timed out trial does not push the policy away from the task.
+   Before: a timeout got the failure penalty $-1.25$.
+
+## Verification
+
+`pytest tests/test_reward.py`: 14 passed.
+````