plugin-quality audit follow-ups: sweep-all internal tension, wait-what evals gap, orchestrate fork probe

[kyle-sexton]

Small deferred items surfaced as blindspots/optional remediations by the 2026-08-15 plugin-quality audits applied in #2695 (audit packets under plugin-quality data, session a86db903, nonce 20260815T104849Z). None was applied there because each needs its own judgment or authoring pass rather than a mechanical edit:

1. discipline/sweep-all internal tension (flagged, unaudited): its body says "The runbook never hardcodes member names" yet names scrutinize-dont-coast inline; and whether batch_promote can promote a never-tier corrector is unspecified between the userConfig text ("situational corrector skill names") and sweep-all's body. Reconcile the two (either the hardcode is a declared exception or it moves to tier metadata; state the never-tier promotion rule explicitly).
2. discipline/wait-what ships no evals/ while the other 16 skills in the plugin do — consistency gap; author its eval suite.
3. session-flow/orchestrate fork probe (auditor remediation 3 on the fork-claim finding): empirically probe whether a below-limit Agent-tool fork can spawn a non-fork subagent (the docs imply it via the depth-limit carve-out but never state it); record the outcome with a date in context/sources.md, per the skill's own confirm-from-behavior doctrine. feat(docs-hygiene): repo-wide sweep escalation for audit-derivability + apply the sweep's confirmed verdicts #2695 already narrowed the prose claim to what the docs state.
4. discipline declared-delta conformance of the other 11 skills was not audited (the audit covered the five exercised correctors; do-your-research-deep visibly runs an extended loop with a "step 6"). A cheap conformance pass against the method doc's declared-delta rule would close the set.