Fewer authored decisions in agentic-system analysis

Type: reference/types/design-proposal.md

Archived (see archive README). Adopted by ADR 097: the overview, memory report and generated review type specs and the analysis job instructions carry the live design. The pre-adoption state of the workflow's authored decisions remains here — design texture only.

The operator selected five cuts on 2026-09-29 from the workshop's simplification candidates (kb/work/analysis-offload-to-code/simplification-candidates.md, Decisions 1, 2, 3, 5 and 6). This proposal stated the contract changes they needed.

Current state (as of 2026-09-29)

After ADR 096:

• A scoping job ran between the runtime pass and the lenses. It chose brief or full depth for each lens and named trigger records; no code read the depth. The overview rendered its output as ## Lens scoping. Both lens members stated their scope again: the memory profile's scope, and block 1 of the epistemic report.
• Code wrote memory-input.md (boundary, source, the scoping record's memory part, and the runtime member). The memory method (analyse-agent-memory.md) described a private handoff: the specialist hashed its input and method by hand, and the member carried canonical-register-sha256, method-sha256 and worker-model. _verify_memory_member checked only the first, against memory-input.md, which was gitignored run state, so no clean checkout could verify it. Nothing checked the other two beyond presence.
• A review job wrote review-body.md after verification: a description, an Evidence basis: line and prose. review_body_refusals checked only the H1, that line and the description.
• The caller gave source-identity at opening, and the boundary worker wrote source.identity independently; nothing compared the two. _check_incumbent compared identities by exact string.
• The workflow continued only when the verification's Blockers text, lowercased and without a final period, equalled none.