docs(tasks): scope the record protocol to both sides of the contract

[ethan-scitix]

Type

• docs — documentation only

Summary

Follow-up to #68, which is where the sizes below were noticed. This is a scoping fix, not a compression one — for a task author the loaded total goes 1485 → 1460 words, which is noise. Three separate things:

• The record protocol was visible to only one party. sieval/tasks/CLAUDE.md held the vocabulary for types defined in sieval/core/tasks/records.py, and CLAUDE.md loads by directory proximity — so someone editing records.py, profiler.py or runner.py never saw it. That is concrete for grader_output: the key name is a contract the runner reads back to route grader spend into profile.json, and the side most likely to rename it is the side that could not see the rule. Moved to .claude/rules/records.md, whose frontmatter globs sieval/tasks/** and sieval/core/tasks/**.
• ## Data Flow — Async & Concurrency was byte-identical in .claude/rules/tasks.md, which globs the same sieval/tasks/**/*.py — so both copies load together. That is not the intended rules-are-a-checklist / CLAUDE.md-is-principles overlap; it is the same three bullets twice. Dropped from CLAUDE.md, kept in the rules file.
• Constructor knobs restated what an enforcer already checks. check_task_shot_knobs machine-verifies the spellings across every constructor under sieval/tasks/. Trimmed 234 → 172 words, keeping only what the checker cannot say: assigning self._n_shot instead of self.n_shot leaves the class default standing, so meta.json reports the declared shot count rather than the one the run used.

Net: sieval/tasks/CLAUDE.md gets a coherent charter — how a task is declared and where it lives — at 948 words instead of 1485.

Related Issues

Refs #68.

Ordering: #68 has since merged (1c15c00c) and this branch is rebased onto it. The prediction held — #68 touched only the ## Naming Conventions / ### Variants region of sieval/tasks/CLAUDE.md, this PR touches the sections below it, and the rebase was conflict-free. Word counts above are against post-#68 main.

Test Plan

Automated

• Lint/format clean — markdownlint passes (pre-commit)
• check_preflight.py --level quick — 22 PASS, 0 FAIL (only the deep-link reachability SKIP)
• No code changed, so no test impact

Manual

• Coupled sites checked per .claude/rules/engineering-infra.md: the inbound references to sieval/tasks/CLAUDE.md (.claude/rules/tasks.md ×2, sieval/tasks/arc/_base.py, sieval/core/tasks/meta.py, scripts/check_preflight.py) point at the "Task Metadata" and "ppl vs clp" sections, both kept. CONTRIBUTING.md does not mirror the record protocol.
• Duplication verified before deleting — the three Data Flow bullets compared byte-for-byte against .claude/rules/tasks.md.
• New rule frontmatter matches the existing convention (paths: glob list, as in tasks.md / datasets.md).
• No enforcer behaviour changes — check_record_key_access and check_task_shot_knobs are untouched and still enforce what the prose now defers to.

Checklist

Required (all PRs)

• PR title follows conventional format (type(scope): description)
• No internal paths, credentials, or personal info in committed files
• AI-generated code has AI-Generated Code - <model> (<provider>) in module docstring — N/A, no code
• No new upper-layer dependencies added to core/
• Deleted code verified — prose only; every deletion is either a verbatim duplicate that survives in .claude/rules/tasks.md or a restatement of a machine-checked rule

🤖 Generated with Claude Code