Add a named lede-voice rule barring how-to-invoke usage in changelog entries

[williamthorsen]

Status (2026-07-23): On hold — do not implement yet. This ticket adds a named rule to prose that the #1055 reframe is expected to replace; the reframe's direction (kind-keyed lede archetypes vs the density framing) is being settled by the #1056 gold-set eval. Fold this ticket's substance (no how-to-invoke register; flag naming vs usage enumeration) into the reframe, or revive it standalone if the reframe stalls.

Problem

Changelog and release-notes ledes (the ## What section) keep coming out in an instruction-manual register — sentences that enumerate flags and tell the reader how to invoke a feature. A recent example shipped on #681: the ## What ended with "Use --kb to choose which knowledge base to scan and --stale-after to set how old verification may be before it counts as stale." That is reference-doc content in the wrong channel: it buries the actual delta and makes the lede less likely to be read.

Context

• lede-voice.md (source: packages/agents/content/skills/_data/lede-voice.md) is the single source of truth for the changelog-writer subagent. That agent is instructed to read only that file — not CLAUDE.md, AGENTS.md, or saved house-style memories — and explicitly not to work from recall.
• The doctrine already catches the instructional case, but only by inference: such a sentence fails Rule 1 (it is neither outcome nor migration) and the "Allowed is not the same as worth it" test. There is no sharply named proscription against how-to-invoke / per-flag usage enumeration.
• In the #638 Add kb-curate skill for vault-wide KB hygiene #681 case the audit caught it after the fact, but the changelog-writer had self-approved the draft — soft, inferential coverage was easy to rationalize past.

Relevant considerations

• The rule must live in lede-voice.md to take effect. The agent is single-sourced on that file by design; pointing it at CLAUDE.md is both ineffective (subagents do not auto-load it) and contrary to the agent's one-doctrine-file model.
• A named rule with a concrete bad→good example is far harder to miss than inferential coverage, consistent with the principle that agents imitate the nearest concrete example more reliably than a directive read once.
• The rule must distinguish permitted flag naming (e.g., --apply named as the load-bearing report-vs-fix distinction the reader acts on) from prohibited usage enumeration (teaching how to invoke). The existing "earns its place" guidance is the natural hook.
• lede-voice.md in the consuming environment is a generated/synced file; edits go to the repo source and propagate via codeassembly-agents install.

Proposed solution

Add a named rule to lede-voice.md stating that the lede narrates the change, not how to invoke it: no how-to-invoke phrasing and no per-flag usage enumeration. A flag may be named when the reader must act on it as a behavioral distinction, but the entry must not enumerate flags and their usage — that is reference-doc content, routed to the docs or ## Details. Include a bad→good pair drawn from the #681 case:

• Bad: "...pass --apply to also have it fix the two cases... Use --kb to choose which knowledge base to scan and --stale-after to set how old verification may be before it counts as stale."
• Good: "...It reports findings without touching the vault by default; run with --apply and it additionally repairs the two issues it can fix safely..."

Acceptance criteria

Must have

• lede-voice.md (repo source) gains a named rule proscribing how-to-invoke phrasing and per-flag usage enumeration, with a bad→good example.
• The rule distinguishes permitted flag naming (reader acts on it as a behavioral distinction) from prohibited usage enumeration.

Should have

• The worked example is the #638 Add kb-curate skill for vault-wide KB hygiene #681 --kb/--stale-after case (or an equivalent), so the nearest concrete example models the rule.

Nice to have

• A cross-reference between Rule 1 / "earns its place" and the new named rule, so the inferential and named paths reinforce each other.