feat: complete verified silicon local loop and final B consent

[Shik-Utau]

Summary

Deliver the MatCreator silicon local loop against Ai-ready_Database #1, covering #3, #4, #5, #6 and both retrospectives.

material request -> public local lookup
  compatible record -> export, validate and reuse
  no compatible record -> original authorized VASP work -> validate -> local archive/readback
  query failure -> explain unknown data existence and duplicate-computation risk -> pause
    explicit current failure/task-bound confirmation -> resume within original authorization
    refusal, no confirmation, stale confirmation or preparation-only authorization -> remain bounded

Final policy B replaces development policy A only after verified T4 acceptance at df6116d. Goal/condition changes and new lookup failures invalidate old confirmation, including user update turns without tool calls. Runtime hooks, tracked-job tools, local archive/reuse checks and repeatable CI are included. No cloud database upload or new real computation was performed during T5, retrospectives or PR preparation.

Evidence

交付口径（2026-10-08）： 本轮以本地验收为准，交付范围为 #1 主规格下的 #3–#6 硅闭环及两次 retro。隔离回归 150 项通过、零跳过；两条保留真实记录的只读复用核验通过；最终 B 授权边界和旧确认失效检查已完成。完整本地测试仍有既有的 39 项失败、2 项收集错误，与已记录基线一致，不宣称全量测试通过。远端后端与前端检查通过，硅闭环作业在读取私有依赖时失败、尚未执行测试；该状态单独保留，不作为本轮本地验收完成的阻塞。本轮不配置或上传跨仓库凭据，不重提真实计算，不上传云端数据库。PR 保持未合并，现有未提交改动保留。后续模糊自然语言检索需求另行形成规格，不追加到本次交付。

• Before: skill loading alone did not establish a verified compute/archive/reuse loop.
  After: T2 retains the real silicon relaxation/static calculation evidence; T3 verifies local archival; T4 and acceptance JSON record independent natural-session discovery/reuse with zero newly owned jobs. Raw outputs remain local. Reuse is evidence matching, not a claim that a denser mesh proves scientific accuracy.
• Before: development A continued on failed lookup; stale consent survived standalone ENCUT, KPOINTS and functional changes.
  After: T5 and acceptance JSON retain the A→B prerequisite/switch, 29 passing consent cases and dual-axis review (Standards 0 unresolved; Spec 0 unresolved). Issue #6 retro turns the reviewer finding into deterministic boundary checks: historical hook 4 failures/2 passes; repaired hook 6 passes, zero skips (proof).
• Before: optional dependency skips and manual acceptance obscured repeatability.
  After: Issue #5 retro adds one check CLI, zero-skip enforcement, retained JSON/JUnit/logs, read-only reuse auditing and a public-bridge CI job. Clean pinned aidb checkout: 150 isolated tests passed, zero skips; fresh readback of both real records passed. Dependency is published at 87148777fc9995a6d2e3bd85571ee8e20ff53181. The fea:add ft_agent #6 retro reuses this infrastructure instead of adding another harness.
• Full local suite: 864 passed, 39 failed, 2 collection errors, zero skips; exact failures match the existing baseline. This is not an all-green suite; unrelated failures remain unassigned and need triage. Dedicated silicon CI does not use this baseline.
• Remote CI: push run 37710820263 passed backend and frontend checks. The silicon job failed before tests at dependency checkout (Repository not found). GitHub API confirms Zikkying/Ai-ready_Database is private: a published branch is not publicly readable, and this repository's default Actions token cannot read it. PR run 37711016873 reproduces the checkout failure; backend and frontend passed. No silicon test/artifact success is claimed. If remote silicon CI is required in a future delivery, a maintainer must provision a repository-scoped read credential for the pinned private dependency and wire it to the dependency checkout, then rerun CI. This is deferred under the local-only acceptance scope for this delivery. This account has push access but no repository administration permission; no local credential was uploaded.

Merge Danger

Door: two-way

Code can be reverted; already submitted scientific jobs and local archive side effects cannot be undone by reverting code. Real jobs are not rerun by the regression checks. Keep this PR unmerged pending review.

Blast Radius: runtime

Registered agent callbacks affect skill preflight and silicon computation authorization. Final B intentionally pauses query failures until valid explicit confirmation. CI relies on the pinned public aidb bridge; aidb main alone lacks that interface. Existing unrelated working-tree changes are excluded from this PR, while prior delivery commits and their environment/tutorial guidance are retained.