RFC: bind long-running results to evidence-based next actions

[loopx-agent]

Long-running work can consume model calls while its result is unchanged, and recorded negative evidence can leave the next experiment unchanged. This proposal binds validated results to a decision checkpoint and binds the resulting decision to the executed next step through existing LoopX owners.

This is a bilingual RFC refinement, with roadmap and benchmark qualification updates. The new policies are proposed, default-off designs; this PR delivers no runtime controller, admission change or benchmark outcome.

• Stride owns decision timing: provider-qualified results advance a bound monitor, while every admitted accountable Turn still settles normally. Recovery preserves job/result identity and the incumbent artifact; a new recovery Turn is still accountable.
• Explore owns scoped evidence and adoption: distinguish failed measurement, inconclusive evidence, candidate rejection and supported family retirement. A continuation or successor cites the exact result, what remains unanswered and the next discriminating probe. Justified replication and same-family probes remain legal.
• Start with benchmark-toolkit's existing admitted-run monitor/closeout seam and one real provider trial. Intermediate training checkpoints require separate provider qualification. Reuse typed Todo/replan receipts rather than add another scheduler or decision store.
• Qualify shadow observation before opt-in enforcement, including stale/replayed results, lost notification, restart, authority fences, packaged interaction, default-off parity and matched total-cost/outcome comparisons. Reconcile the closed [Task][RFC]: Complete the research exploration control-plane contract #5214 checkpoint without claiming missing M3 delivery.

Validation:

• git diff --check, added local links/anchors (14), public/private scans and manual bilingual semantic review passed.
• Semantic inventory advisory: 0 changed source paths; no new wire vocabulary implemented.
• Strict docs governance/status-index checks fail on two existing dated headings in external-evidence-research-capability-v0.md and its Chinese mirror. The same failures reproduce on base 30efccd6c7d4e2300c2ac87ddc6ddb1dfb855a40.
• Full docs governance smoke passes in a scratch copy with only those two baseline headings normalized. The repository files are unchanged; this diagnostic does not turn the strict failure into a pass.
• Runtime tests and benchmark jobs were not run: this is a documentation proposal.

Future-facing pass: reuse the existing Explore, monitor and receipt owners; defer wire/schema implementation until the concrete provider path is qualified.

Refs #3203, #4391, #3243. No implementation milestone or umbrella issue is closed by this proposal.