feat: Add scope split, followups stage, and detection standing checks

[jirispilka]

What we're solving

Mined human review feedback across recent apify-mcp-server PRs (#1081, #1094, #1077, #1076, #1073, #1089, #1080): shepherd runs keep collecting the same finding classes despite passing the panel — untested new code arms (one shipped a data-corruption bug, #1094), undisclosed behavior deltas, convention/layering drift, and total silence about adjacent debt and scope boundaries. One run's own review scope even instructed reviewers to ignore carried-over defects — which the human reviewer then flagged.

How

• Detection: standing checks gain new-arm test coverage (both sides + the old path's input classes), undisclosed-delta disclosure (incl. silently inverted helper defaults), and conventions/layering; per-lens detail goes into the base-registry reviewer scopes (uncapped), not orchestrator prose.
• Scope: the design template gains ## Scope split (This PR / Prerequisite refactor / Follow-ups); a non-empty prerequisite is an explicit human gate decision. Reviewers tag adjacent pre-existing defects pre-existing instead of silencing them; a new followups stage (dispatched with fulfillment — file contract, not orchestrator memory, so a weak orchestrator model can't drop it) compiles them into a ledger the human dispositions at the create-PR gate: fix-here | issue | pr-note | drop, issues only on human approval. A hard rule forbids scopes from suppressing reporting.
• Claims: PR-body numbers and references must match the final oracle run and repo state.

Line cap 450 → 500 (now 482). Contract tests pin the prose; 97 pass.

Stacked on #6-followup branch shepherd-skill-slop-checks (PR #7) — merge that first; this PR's base is that branch, so the diff shows only this change. Part of the same effort as #9 (refiner stage) and #10 (test-coverage lens).

Field test (2026-07-18)

This branch ran a full production run end-to-end: the apify-mcp-server #1084 task-machinery refactor — triage → verify → design (6 human decisions) → design gate → 3 inner iterations → final review → fulfillment → create-PR gate → two stacked PRs shipped (apify/apify-mcp-server#1105, #1106). 24 stage dispatches, ~2.36M subagent tokens. Per-feature verdicts, each grounded in the run's _progress.md ledger:

• Conventions/layering standing check: caught a real major — the CONTRIBUTING.md >3-params rule broken by two extracted signatures (exactly the mined finding class; under the old prose this ships). The finding forced a design amendment and spawned a CI-enforcement follow-up issue (refactor: Convert >3-positional-param functions to object params apify-mcp-server#1111). New-arm and undisclosed-delta checks were unexercised (pure-move diff, no new conditionals) — no signal either way.
• pre-existing tag + followups ledger: worked end-to-end, and the gate is live — 4 adjacent defects surfaced across 3 review rounds without blocking convergence; the human's dispositions diverged from the stage's recommendations (3× fix-here vs pr-note/drop), reopening the loop; zero issues auto-created.
• Scope split: representation worked, detection missed — the architect declared a "refactor first, then move" plan as internal commit sequencing ("Prerequisite refactor: none") despite the repo's refactor-lands-first policy; the human had to force the split. Fixed in the follow-up commit.
• Claims rule: complied — both shipped PR bodies carry the final oracle's exact numbers.
• Suppression ban: held — pure-move scopes downgraded/routed, never silenced; all pre-existing items reported every round.

Second commit (80b4c2f) folds the field-test lessons back in: human-picked light re-review for trivial gate-authorized fixes; prerequisite classification must consult the repo's refactor-separation policy; the conventions doc settles style questions before they reach the human; delta-mode verify on same-base re-runs; lens-fit panel nudge at every gate (the orchestrator never edits the panel itself). Related follow-up out of scope here: #15 (local commits on a run scratch branch; gate remote writes only).

Alternatives considered

• Orchestrator prose instead of a followups stage — rejected: compiling findings across iterations needs memory and judgment, exactly what a weak orchestrator drops; a dispatched stage re-reads the persisted review files fresh.
• A comment-analyzer reviewer lens — rejected as redundant: comment-truth standing check + PR feat: Tighten reviewer standing checks, design Q&A, and PR body rules #7's slop checks already cover it.

🤖 Generated with Claude Code

https://claude.ai/code/session_017ZmSKShwVcKG2P7MY22X8k

---

Generated by Claude Code

[Pijukatel]

Three more: is somewhat strange here. I am not sure what is meant by it

[Pijukatel]

Instruction with a question mark at the end?