feat(architecture): add four agent hazards to the deepening scan and a one-paragraph shape to record-decision #8646

Refs: #7590

Agent Brief

Type: Feature
Summary: The deepening scan of /architecture:improve learns four design hazards that mislead agents, each judged by an open-file test, and may take a spec or plan for upcoming work as its scope; /architecture:record-decision offers a one-paragraph ADR shape and lists the kinds of decision that qualify.

Current behavior:

• The friction checklist (plugins/architecture/skills/improve/research/deepening/scan-briefing.md:22-48) restates the Phase 1 signals of actions/deepening.md for a subagent: six shared signals and three scan-only ones. None names a hazard by how it misleads an agent.
• Phase 1 scope (actions/deepening.md:22-30) comes from the arguments, the conversation, or recent history; a spec or plan for upcoming work is not a source.
• record-decision Step 2b offers "two or three common shapes" with no shape named (plugins/architecture/skills/record-decision/SKILL.md:96-104), and the admission test (SKILL.md:57-66) gives three questions and no examples of what qualifies.

Desired behavior:

• Four hazards added to the Phase 1 signals in deepening.md and to the checklist in scan-briefing.md, in our words:
  • Split ownership: more than one module writes the same state or keeps its own copy, so an agent that edits one writer cannot see the others. Remedy: one owner per piece of state; others read it or ask the owner.
  • Two ways to do one task: an agent copies whichever way it finds first, so every way keeps gaining callers. Remedy: keep one way; move callers off the others and delete them in the same change.
  • Importable internals: an agent takes the shortest path that compiles and imports internals directly, which makes them part of the interface. Remedy: make internals unreachable from outside, so such an import fails the build.
  • Hand-synced list: two or more places list the same items and adding one means editing every list; an agent updates only the list it sees. Remedy: keep one list and derive the others, or fail the build when they disagree.
• Open-file test, stated once beside the hazards: with only the file being edited open, can an agent see every other place it must change? A "no" confirms the hazard.
• Phase 1 scope sources gain "a spec or plan for upcoming work": when one is given, scan the areas it will change.
• record-decision: Step 2b offers, as one shape, a one-paragraph ADR: a short title and 1-3 sentences giving the context, the decision and why; optional status, considered options and consequences only when they add something. Under the admission test, one line of qualifying kinds: architectural shape; integration patterns between contexts; technology choices that carry lock-in; boundary and scope decisions; deliberate deviations from the obvious path; constraints not visible in the code; rejected alternatives when the rejection is not obvious.
• Eval cases: in improve/evals/evals.json, a fixture where two modules each write the same state yields a split-ownership candidate that cites the open-file test; in record-decision/evals/evals.json, a no-convention run lists the one-paragraph shape among the offered shapes.
• No existing sentence is removed.

Key interfaces:

• /architecture:improve deepening (Phase 1 signals, scan briefing).
• /architecture:record-decision Step 2b and admission test.

Acceptance criteria:

• The diff adds the four hazards and the open-file test to both deepening.md and scan-briefing.md, the spec-or-plan scope source to deepening.md, and the one-paragraph shape and qualifying kinds to record-decision/SKILL.md, deleting no line in any of them.
• plugins/skill-quality/scripts/check-evals-quality.sh plugins/architecture/skills/improve/evals/evals.json plugins/architecture/skills/record-decision/evals/evals.json exits 0, with the two new cases asserting the behavior named.
• scripts/check-changelog-fragments.sh --check passes, with one minor fragment for architecture naming pstack's architect design red flags and Matt Pocock's domain-modeling ADR format as sources.
• CI's check-skills passes on the PR (a PR check; this run does not run it).

Out of scope:

• The boundaries and conventions lenses (the other architecture: boundaries and conventions lenses #7590 slice): they hand findings to /corrections:codify (corrections: new plugin with /corrections:codify #7573), paved-path entries to /discipline:reuse-or-replace (discipline: paved-path file and principle placement #7589, feat(discipline): make reuse-or-replace the paved-path writer and add a codemod-first migration row #8545) and boundary rules to the architecture concern file (standards: per-concern standards layout #7583), none of which exists yet.
• docs/upstream/cursor-pstack.md and docs/upstream/mattpocock-skills.md rows (docs(upstream): re-audit the pstack record past cursor/plugins 99743e9c15 and re-pin it #8542 re-audits the pstack record).
• The research/enforcement-ladder.md copy (regenerated by feat(enforcement-ladder): ship ladder 2.0.0 with a runs-at column, four tiers and the new rung ids #8536).

Origin:

• Four agent hazards with the open-file test: Lauren's (pstack architect design-red-flags.md), Q48.
• A spec or plan for upcoming work as a scope source: Matt's, Q525.
• One-paragraph ADR shape and qualifying kinds: Matt's (domain-modeling ADR format), Q525.

Do not run with: none open.

Slice of #7590, drafted by the poteto-review session from the thread's recorded decisions.