ADR-0167 - Bounded campaign guardrails

Context

ADR-0162 established campaign orchestration. ADR-0163 removed review theater from pull requests. Later guardrails bounded pull-request size and review breadth so an agent could not turn ordinary delivery into an endless ceremony.

The direction stands. A skeptical review found eight defects in its wording and boundaries: mandatory third choices polluted binary decisions; retained dependencies could go stale; linked worktrees were called isolated; prerequisite pull requests were called independently mergeable; reviewers could miss governing authority; repairs could expose risks without creating them; every finding was forced into a line-or-summary split; and final verification named no exact merged state.

Evidence supports small, coherent review units but supplies no universal line, file, option, or reviewer optimum [DIBIASE19] [BOSU15] [CHOICEOVERLOAD]. Suspec therefore labels its numbers honestly: circuit breakers against runaway execution, not laws of software quality.

Decision

1. Choices stay genuine. Material ambiguity gets three options by default and two when genuinely binary. Recommendation and cost remain explicit. Inventing a third answer is not rigor.
2. Lane reuse checks dependency identity. Ignored dependencies survive only while the lockfile and toolchain match the next branch. A mismatch triggers refresh. Preliminary CI evidence illustrates both reuse gains and stale-state failures; it does not directly measure local worktree dependencies [CICACHE].
3. Worktrees are separate, not isolated. They remain reusable write lanes; Git's shared repository state is not hidden behind a false guarantee.
4. Pull requests are independently reviewable. They merge in declared dependency order. The 500-reviewable-line and 15-handwritten-file limits remain default stop signals, never quality scores or universal thresholds.
5. Fresh review does not erase authority. Each reviewer receives governing requirements and accepted decisions. Prior reviewer prose and pull-request conversation remain hidden. Existing comments have not shown broad negative priming in controlled human review, so this separation is a method boundary, not an empirical claim [REVIEWPRIMING].
6. Risk discovery may follow repair. A frozen stance pool grows only when a repair creates or exposes a material trust boundary, public contract, failure mode, or previously unknown risk.
7. Comments follow defect scope. Local defects use changed-line threads, file-wide defects use file comments, and cross-cutting defects use short review-body items. Repeated instances collapse only when one repair and owner can resolve them together.
8. Final proof names merged state. Campaign closure verifies acceptance conditions against every merged SHA. No Suspec manifest, registry, or duplicate completion record is added.

Narrowed and upheld decisions

• ADR-0145 and ADR-0157: the three-option default now admits genuinely binary choices.
• ADR-0162: reusable lanes, fixed repository pools, one writer, adaptive routing, sequential repair, immediate authorized merge, and native project state stand.
• ADR-0163: internal stances, concise pull requests, actionable comments, implementation-owner repair, and in-place resolution stand.

Consequences

• Campaign remains bounded without pretending its guardrails are scientific constants.
• Reused lanes keep their economy without trusting stale dependencies.
• Reviewers receive authority, not each other's narration.
• Pull-request discussion matches defect scope and closure proves the code that actually merged.

Status

Accepted (2026-07-20). Narrows ADR-0145, ADR-0157, ADR-0162, and ADR-0163.