perf(review): shorter reviewer reports; draft verdict-independent records during review

gates/selftest.sh

@@ -45,12 +45,12 @@ budget='623 SKILL.md
 2335 skills/1-intent/SKILL.md
 663 skills/2-spec/SKILL.md
 901 skills/3-plan/SKILL.md
-1337 skills/4-build/SKILL.md
+1374 skills/4-build/SKILL.md
 1903 skills/5-ship/SKILL.md
 1816 skills/6-maintain/SKILL.md
-818 roles/adversary.md
+824 roles/adversary.md
 440 roles/researcher.md
-2420 roles/verifier.md'
+2456 roles/verifier.md'
 for p in "$kit"/SKILL.md "$kit"/AGENTS.md "$kit"/skills/*/SKILL.md "$kit"/roles/*.md; do
   rel=${p#"$kit"/}
   max=$(printf '%s\n' "$budget" | awk -v f="$rel" '$2 == f { print $1 }')

roles/adversary.md

@@ -75,7 +75,7 @@ adversarial diff pass still follows all three lenses on the same final source.
 - <objection>; evidence: <file:line / quote>
 ### Non-blocking (flag to human at gate)
 - <concern>
-### Checked and clean
+### Checked and clean   (one line per area, no narrative)
 - <area>: <what you looked at>
 ### Per changed file   (diffs only)
 | File | Severity | Finding |

roles/verifier.md

@@ -171,6 +171,8 @@ snapshotted`; live source unreachable → say so, the snapshot stands.
 
 Record every check as **command or tool · environment · scenario · observed
 result** (deciding lines verbatim; bulk into `.sdlc/work/<slug>/scratch/`).
+A current receipt stands for the recipe checks it covers: cite it once, and
+list a recipe check alone only when it failed or you read its log.
 Include report provenance — every item of the template's Provenance,
 Inputs/source and Evidence lines, with actual IDs when exposed. An
 unavailable identifier is `unknown — <why>`; never fabricate an ID or claim a
@@ -188,8 +190,8 @@ call, recorded in evidence.md.
 ### <E2E | Side effects | Intent match>   (fill your lens's lines)
 - Provenance: reviewer ID <actual ID | unknown — reason>; context ID <actual ID | unknown — reason>; tool calls <actual IDs | unknown — reason>; independent of implementer <observed relationship | unknown>
 - Inputs/source: <revision + current tree identity>; recipe <identity>; upstream <identities>; input paths <exact paths>
-- Evidence: executed <commands, exit codes, log paths, deciding lines>; independently inspected <current logs/receipts, or none>
-- Ran: <command> → <verdict line(s)> · receipt: <VERIFY line · coverage line>
+- Evidence: executed <commands beyond the receipt, exit codes, log paths, deciding lines>; independently inspected <current logs/receipts, or none>
+- Ran: receipt <path · recorded_at · VERIFY line · coverage line> · <other command> → <verdict line(s)>
 - Failures: pre-existing <ids> at <base ref @ sha> | none · regressions <ids> | none
 - Changed files: <path> → <R/scenario> | no behavior: <reason>   (one per file)
 - E2E: <command/tool> · <environment> · <scenario> → <observed>

skills/4-build/SKILL.md

@@ -122,6 +122,10 @@ With a `.sdlc/verify.md` recipe the E2E lens runs `tools/verify.sh run <slug>`
 unless rule 6's proof reuse applies, inspects deciding logs and
 expectations itself, and evidence.md quotes the receipt's deciding lines.
 
+While reviewers run, draft what does not wait on their reports — qa-guide.md,
+harvest.md, and evidence.md outside Verification and Adversarial code review —
+so ship adds the reports and approves. A finding or gap revises the drafts.
+
 A finding from any lens — a failing flow, a data skew, a missing origin
 detail — enters the **fix loop**: