Planning

What this role does, and does not

Does: a whole repository's plan from scratch (specs → decisions → open
questions → cycle map → cycle 0.0), or a cycle's README.md and its subcycle
map.

Does not: write the next subcycle file at a cycle's close. That belongs to
the worker who just finished, deliberately — the person who has just learned
what a cycle taught is the one best placed to plan the next, and handing it to a
fresh planner throws that away. /npk:worker's close checklist owns it.

The order, and why it is not negotiable

specs → decisions → open questions → cycle map → cycle 0.0 → then code.

The compiler project credits this order with catching most of its design holes,
and its cycle notes record the counter-example: a construct that parses is not
a construct that works, found by three cycles that were mostly repair.

The rule that keeps a plan honest

Every decision records the alternatives declined, not just the choice. The
alternatives are what a later reader will propose, and a decision that does not
say why they lost will be re-litigated. This is the single highest-value habit
in the ecosystem's decision logs.

Two more that follow from it:

• A settled decision's text is never rewritten. Supersede it with a new
  numbered one that says why. nitpick-tui's T-113 is the worked example: it
  corrects a false claim about signal disposition without erasing the claim,
  because the record of how the error survived is itself the lesson.
• A question that gets answered is struck through with its decision number,
  never deleted. The question is part of the record of how the answer was
  reached.

What a plan must contain

• Specifications that state numbered, normative rules — facts about the
  library, not intentions. Rationale paragraphs carry no obligation and are
  marked as such.
• A safety document first, stating what the language imposes: the
  failsafe contract, the error budget REACH-002 forces, the trapping
  arithmetic, the borrow and ownership rules. Most proposals that look
  reasonable in the abstract die on one of these.
• An error budget as a number, with the module decomposition that keeps it
  payable — REACH is import-scoped, so which module declares what is part of the
  budget.
• A cycle map where the riskiest thing is early, instruments precede what
  they guard, and cycle 0.0 is always the language probes.
• Open items with a recommendation each, or a stated reason they stay open:
  it is a measurement, it is data, it is gated, it waits for a consumer. "Open"
  with no reason is "forgotten".

Currency

A plan asserts things about the world outside the compiler tree — a
standard's edition, a data release, a corpus revision, a reference
implementation — and every one of them moves. The plan writes
meta/research/CURRENCY.md: one row per such dependency, with the version
pinned, the date it was checked, the source, and the decision that pins it.
The shape, the sources that count, and the request and digest formats are
skills/research/SKILL.md §7, §2, §3 and §5; cite them, do not restate them.

• A cycle whose currency rows are unchecked is not ready to start, the
  same rule as for open questions.
• A security-sensitive library has three mandatory research items before
  its cycle map — known defects in the prior art and reference
  implementations, the standards body's current guidance, the current test
  vectors — re-run at hardening (skills/research/SKILL.md §8). A finding
  there is a cycle-0.0 probe before it is anything else.
• The model's knowledge is months old. Anything the plan calls "current"
  without a date beside it is a currency row waiting to be found.

Probes, spikes, and the difference

• A probe answers "is this spellable?" — a small program asking the compiler
  whether a language shape the design depends on actually works. A probe that
  fails changes the design, which costs a day now and a cycle later.
• A spike answers "how big is this actually?" — a measurement, with its
  thresholds decided in advance, so a bad number produces a stop rather than
  an improvisation.

Both belong in cycle 0.0. Confusing them produces a probe with no pass mark.

Verify before committing

python3 ${CLAUDE_SKILL_DIR}/../check/scripts/check_refs.py .

A plan whose cross-references do not resolve is a plan somebody will follow into
a dead link. Every decision cited must be defined, and every decision defined
must be cited — an uncited decision is usually a specification that states a
rule and forgot to attribute it.

Do not copy content into the plan

The language constraints live in ../PLAYBOOK.md; the compiler's decisions live
in its own tree. Cite them; do not restate them. A fact with two homes drifts,
and this ecosystem's entire discipline is that a fact has one.