feat(storage): add bounded ledger archive reconstruction reader

[likun666661]

Summary

Refs #4071. Reader-first follow-up after #4963 and #4965.

Tool Result archive bytes are derived from durable model projections. Before changing writers, this PR adds a read-only reconstruction path that proves when an existing v1 archive can be recovered from the ledger alone.

• Add a lease-bound SQLite evidence reader for one Session/event target. Select sizes and at most 65 transition metadata rows before loading JSON; reject more than 64 transitions or more than 2 MiB total evidence. A read is complete or refused, never a silently truncated history.
• Add an indexed Session/target lookup and refuse unscoped malformed transition evidence. Use one read transaction for metadata admission and body reads. Keep node:sqlite loading lazy.
• Reuse the existing reducer to accept only applied archive transitions. Reconstruct the source immediately before the matching replacement, including predecessor chains and concurrent sibling selection.
• Freeze v1 archive serialization independently of future provider/UI formatting. Verify the stored placeholder identity, hash and byte length; do not expose raw tool results or run the legacy/live tool projector when a durable projection is missing.
• Publish narrow reader entrypoints, but do not wire them into production ArchiveRead, change writers, delete payloads or activate a legacy fallback in this slice.

Verification

• Core/storage/runtime builds and typechecks passed.
• Full storage suite: 1196 passed, 8 skipped, 0 failed.
• Final focused reader/evidence/reducer/ArchiveRead suites: 52 passed, 0 failed.
• Real SQLite close/reopen reconstruction with no Artifact payload; cross-Session refusal; predecessor-source recovery; sibling/duplicate transitions; unknown versions; missing projections; hash/size limits.
• 12k unrelated transition records do not enter the target evidence result. EXPLAIN asserts the target index; oversize evidence returns before any ledger JSON is materialized. v8-to-v9 migration preserves transition bytes.
• Negative control: bypassing the reducer's applied-transition filter makes the refusal regression fail. Restored before final verification.
• Root lint, format check, ASF headers and diff check passed.

Full runtime suite is not green locally: 3299 passed, 13 skipped, 2 failed in unchanged Responses reasoning coverage:

• Alibaba Responses keeps multiple streamed reasoning items distinct through replay.
• The pinned SDK maps official summary events across raw byte chunks.

The second reproduces when run alone. No Responses/provider implementation is changed here. Full Host/Desktop suites, Windows/Linux execution and large-workspace latency/RSS benchmarks were not run.

Migration and Rollout

Core-execution schema advances from 8 to 9 to create the partial expression index over existing transition records. This is a rebuildable index, not another history authority. JSON-invalid rows are indexed as unscoped instead of crashing index extraction.

Index construction has a one-time schema-upgrade cost. This PR does not claim index creation itself is bounded or that #4027 startup acceptance has passed. Normal reads use indexed target lookup and bounded metadata/body retrieval.

The reader is intentionally conservative: missing durable projections, incomplete/corrupt evidence and evidence exceeding the internal 64-transition/2-MiB limits are refused. The evidence byte limit includes the containing RuntimeEvent, so a small projection inside a very large raw-result event may also be refused. Existing Artifact reads remain unchanged.

The next writer slice must introduce its own new ledger-reference protocol and copy/remap contract using this reader foundation. This PR only reconstructs current v1 archive identities; it does not emit new durable refs or claim all historical archives are reconstructible. No old payload can be reclaimed merely because this reader exists.

AI use

• No generative tool made a substantive contribution
• Generative tooling made a substantive contribution

Tool(s) and scope: Codex inspected the ledger and storage contracts, implemented reader/index/test changes, ran local checks and prepared this PR. Commit carries Generated-by: Codex.

Checklist

• Tests cover the change and fail without it
• Lint, format, typecheck and the affected suites pass locally (focused and full storage pass; full runtime failures documented above)

Does this PR entail a change in behavior?

• Yes — adds a schema index and opt-in read capability; production archive routing is unchanged.
• No

[Astro-Han]

Thanks for the reader-first implementation. Re-reviewed e6ec11f33d1a6dc559c6483c09a2034c77826847 with independent design, storage and reconstruction reviews. One API error-classification finding is inline, independently reproduced against the existing lease-revocation test. The production ArchiveRead route remains outside this slice.

AI disclosure: this review was performed by Codex agents; the coordinating agent checked the finding and reran the counterexample. This is not an independent human review.

中文摘要

已对当前提交重新进行设计、存储和重建路径交叉审查。发现一项 reader API 的错误分类问题，主审已通过现有 lease 撤销测试独立复现，详情见行内评论。生产 ArchiveRead 切换仍属于后续范围。本次为 Codex 自动化审查，不冒充人工验收。

[Astro-Han]

LGTM.