[aryanorastar]

Purpose

Make the support plan easy to follow and bring every existing PR to a deliberate merge/repair/close decision. Record the separately requested public-data research without adding another feature PR. Documentation, metadata and unexecuted synthetic scenario specifications only: no runtime, provider, secret, setting, permission or deployment change.

Current outcome

• Preserve all 13 existing PRs, product milestones and unfinished acceptance items. Do not create more feature drafts or close useful work merely to lower the count.
• Record completed Keep support priorities and PR outcomes clear #69/Exercise reference memory through SQL and live context #73 review and checks, Prepare durable reviewed corrections for the fix inbox #49/Remove owned correction links when forgetting support history #74 independent passes for disabled foundations, and the concrete Bound optional planning and continue safely on timeout #68 language / Recover abandoned unsent replies after missed heartbeats #70 recovery-delivery findings. Keep Keep ordinary support and failed follow-ups in the same thread #67/Keep evidence for every part of a support question #71's failed speed gates, Keep customer corrections from becoming unrelated action requests #75's unverified flow/quality work and the Preserve staff-approved engineering updates across restarts #28/Prevent duplicate support sends with durable Discord receipts #29/Make tracked case closure and reopening receipt-safe #30 dependency/conflict/correction requirements explicit.
• Align instructions with the user-approved risk-based policy: verify the changed boundary, reuse valid evidence, retain repository-required comparisons/approvals and material safety gates. No irrelevant customer/hardware certification for inert documentation/test/foundation changes.
• Show 47/100: 14 built + 14 tested + 12 reviewed + 7 verified live. The only new point is milestone 18's independent Prepare durable reviewed corrections for the fix inbox #49 foundation review, not a deployment or accuracy gain. Full A–E acceptance remains 2/21; no memory or learning is enabled.
• Preserve release/rollback/comparison history behind explicit historical folds. Old process handles, manual-Restart notes and superseded review-pending checkpoints are not current commands or blockers. Do not claim global exactly-once delivery, causal aggregate latency results or a fresh production health check.
• Add docs/MIRA_DATA_STRATEGY.md: eight primary-owner dataset/license screens, a blocked-by-default metadata register, and 12 original support-evaluation specifications across English, Spanish, Chinese and Hindi scripts/code-switching. Distinguish candidate data, unresolved rights and non-commercial/academic exclusions. No external corpus rows, training job, executed new benchmark, prompt change, model migration or progress point. Later execution stays behind the existing finish queue in A1/A2/E2/Give Mira a consistent, considerate support personality #59; Discord messages remain outside training.

Final verification

Exact head: dda2cf6fb820e4da16f18a70cb6692aa316ddd39.

• npm test: 839/839, before push; offline customer check 8/8, no provider calls. Runtime/dependencies are unchanged. Exact-head hosted run 39948620027: Node 20, Node 24 and real PostgreSQL all pass.
• One user-authorized independent reviewer found stale status instructions, then re-reviewed their corrections. The new data-plan/metadata/scenario additions also pass independent review, reconfirmed at this exact head: no actionable findings. Earlier unaffected-document review evidence remains valid (working-diff SHA256 00df05438b998cfd63086907078d81c78fa0146dba53be53b1011ee659504584, base 6e27f2f). This is agent review, not a fabricated human approval or data-use/training authorization.
• Current-head GitHub Preview confirms the new data-plan table/checklists and updated master/architecture/roadmap links. Unchanged regions retain prior rendering evidence: score, all 13 PR rows, collapsed history, operations and risk-based instructions. Relative links resolve, folds balance, and git diff --check passes. All 16 external dataset source/terms links return HTTP 200; primary pages were read, not merely link-checked.
• JSON checks confirm eight unique blocked sources, 12 unique unexecuted scenario specifications across five locale/script labels, and training approval false. These are metadata/spec integrity checks, not executed answer-quality tests, fluent-speaker certification or corpus-level rights approval.
• Reproduce: check out this exact head, run npm test / git diff --check, open the changed Markdown files in GitHub Preview, expand the archive controls and follow the linked project documents. Never execute archived handles or enable a feature as a documentation test.

Merge versus deployment boundary

Ready for maintainer review of the documentation. Not merged or deployed. Aryan authorized deliberate completion of the PR queue, not a production setting change. The existing Railway browser session is readable, but its source-branch lookup reports "Auto deploy unavailable" / "Could not load branches" after one retry; CLI authentication is unavailable. This is not explicit OFF proof. Do not merge on an assumed automatic-deployment state or reuse an old active-deployment guard. No key value was displayed; no new key, permission, production setting or deployment was created.

Public content contains no customer text, private case identifiers, protected source cases, external corpus rows or secret values. Current readiness clears none of the runtime PRs' missing/failing gates, data-use/training approvals or optional-feature enablement conditions.