[Refactor] Audit design conformance and implement missing runtime behavior

[dotnetpower]

Motivation

Review the current architecture and performance contracts against executable paths, preserve the uncommitted hardening work under #27, and implement reproducible missing behavior rather than equating standalone modules with integrated features.

Work

• Trace freshness, indexing, search, resource governance, and P2P contracts to their consumers.
• Add regression evidence and implement validated gaps incrementally.
• Measure meaningful bottlenecks and repeat critical review.
• Synchronize English/Korean documentation with implemented and unimplemented contracts.

Acceptance

Document an explicit design conformance matrix, focused evidence, supported Python test results, performance measurements, and any work requiring further implementation or operational validation.

Related Code

• infomesh/crawler/crawl_loop.py
• infomesh/services.py
• infomesh/index/local_store.py
• infomesh/search/query.py
• docs/en/02-architecture.md

Implementation Verification (2026-09-19)

Implementation checklist rechecked against the current local checkout. Python 3.13: 2,089 passed (54.48 s); Python 3.12: 2,089 passed (62.74 s). The supported gate excludes optional vector tests. Ruff lint/format, strict mypy, build, actionlint, editor checks and diff validation are recorded in the bilingual conformance report.

The latest #28 follow-up corrects owner-deletion starvation beyond the first 20 peers, with a durable per-URL cursor and 45-peer/restart/failure/churn/old-database regression evidence.

Evidence: docs/en/13-design-conformance.md and docs/ko/13-design-conformance.md, Open-Issue Acceptance Recheck. Physical LLM/audio quality, optional vector implementation, cross-platform and long-duration public-network validation are not certified by these local tests.

Status: implementation and local validation ready for maintainer review; no commit, push, deployment or issue closure performed. Checked boxes are not publication evidence.

[dotnetpower]

Design conformance review remains In progress.

Current local implementation covers scheduled refresh, stable-ID index updates, compressed-text consistency, metadata write reduction, search-cache isolation and accounting, distributed/explain filters, MCP resource gates, P2P inbound load limits, bounded crawler GETs, robots error handling, and auxiliary crawl admission.

Validation: Python 3.12.13 and 3.13.14 each passed 1,951 tests. Ruff lint/format, strict mypy, actionlint, wheel/sdist build, editor diagnostics and diff checks passed. Optional vector tests are excluded by the repository gate.

Synthetic measurements: 200 metadata updates on a 28 KB document improved from 20.982 ms to 0.908 ms. A 1,000-document local-search benchmark measured p50 0.620 ms and p95 0.971 ms. These are local microbenchmarks, not public-network performance guarantees.

Production integration for mandatory message authentication, document replication, audit/deletion enforcement, background summaries and wire bandwidth governance remains incomplete. A versioned protocol migration and legacy-peer compatibility policy is needed before changing the existing wire contract. This issue is not ready for closure.

Detailed bilingual evidence is in docs/en/13-design-conformance.md and docs/ko/13-design-conformance.md in the current local checkout. No commit, push or deployment performed.

[dotnetpower]

Parallel v1/v2 search and replication rollout is implemented in the local checkout.

New hosts offer v2 before v1 in one negotiation and retain legacy interoperability. A selected v2 exchange never retries v1 after authentication or parsing failure. Ed25519 signatures use the actual libp2p host identity and bind both stream challenges, recipient, protocol and direction. Ranked JSON exposes verified peer provenance. Replica requests validate source identity, content hash and positive storage ACKs.

Validation: real localhost libp2p hosts passed v2/v2, v2/v1 and v1/v2 search and replica exchanges. Python 3.12.13 and 3.13.14 each passed 1,984 tests; ruff lint/format, strict mypy, actionlint, wheel/sdist build, diagnostics and diff checks passed. Optional vector tests remain excluded by the repository gate.

Evidence and boundaries: docs/en/13-design-conformance.md and docs/ko/13-design-conformance.md. Default replica storage/repair, the remaining protocol families, audit/deletion and summary integration are still open work under this umbrella issue. Parallel v1 support does not claim universal authentication or content truth. No commit, push or deployment performed.

[dotnetpower]

Severity-ordered follow-up review: 40 critique items, 32 implementation improvements, 8 remaining architecture/integration items.

Priority corrections include P2P fragmented-frame memory amplification, bounded gzip/deflate expansion, redirect-body and total-response budgets, exact framing and unique message keys, single-use authenticated channel state, strict MCP inputs, safe authentication failures, queue cancellation recovery, indexer ACK accounting and crawler scope/DNS responsiveness.

Measured synthetic allocation peaks: 100,000 one-byte frame fragments, 8,901,521 to 206,641 bytes; 4,000,000-byte gzip expansion with a 4,096-byte limit, 9,661,263 to 64,191 bytes.

Final gates: Python 3.12.13 and 3.13.14 each passed 2,053 tests. Ruff lint/format, strict mypy, actionlint, wheel/sdist build, changed-file diagnostics and diff checks passed. Optional vector tests remain excluded by the supported gate.

Remaining High items: default N=3 replica storage/repair, end-to-end audit/isolation/deletion integration, and authentication for other protocol families. Remaining Medium items: connector-IP pinning, wire bandwidth governance, summary jobs, conditional refresh, and remote language/date metadata. These have not been downgraded to Low. Approved parallel v1 compatibility is preserved.

The full 40-item bilingual ledger is in docs/en/13-design-conformance.md and docs/ko/13-design-conformance.md in the local checkout. This umbrella issue remains In progress. No commit, push, deployment or issue closure performed.

[dotnetpower]

Runtime integration checkpoint (not closure).

Implemented: default v2 owner-scoped replica storage and durable repair; shared Trio stream budgets including Kademlia; persistent owner withdrawal outbox and tombstones; signed replica audit evidence and isolation; DNS-pinned HTTP connections with original TLS hostname; conditional refresh; leased content-versioned local summaries; remote search metadata.

Evidence: Python 3.13: 2072 passed (53.27s). Python 3.12: 2072 passed (58.84s). Optional vector tests excluded. Ruff lint/format (268 files), strict mypy (167 modules), Pylance summary call compatibility, actionlint, wheel/sdist and diff checks pass. Actual localhost tests cover four nodes, three replicas, replica search, owner deletion, tamper isolation and governed Kademlia traffic. Source recrawl responses and the LLM backend are controlled test doubles.

Remaining contracts: High universal authenticated ingress and complete enterprise submission; High source consensus and authorized legal deletion/relay/offline coverage; Medium independent peer summaries and atomic reward recovery. Five of the historical eight integration areas are connected; three remain partial. See docs/en/13-design-conformance.md and docs/ko/13-design-conformance.md for the current ledger.

Board stays In progress for #28; #19-#27 stay In review. No commit, push, deployment or long-duration soak performed.

[dotnetpower]

Final implementation checkpoint for maintainer review.

This supersedes the prior partial checkpoint. All eight tracked runtime integration areas now have executable paths. Added original-issuer legal authorization (empty-by-default trusted_deletion_keys), localhost /legal/notices, durable multi-hop legal relays, all-owner replica blocks and optional-vector cleanup. Enterprise split now uses the actual multiaddr v2 bridge and default search-node receiver with storage/publication ACK and resource limits. New /infomesh/ DHT writes validate signatures/key/issuer and bounded keyword admission. Source audits aggregate three distinct signed recrawl responses; changing web content remains inconclusive. Independent replica LLM jobs produce version-bound reports, update trust/reputation once, and award local/peer credits through durable outboxes and idempotent ledger receipts.

Final evidence: Python 3.13 2086 passed / 55.93s; Python 3.12 2086 passed / 58.22s. Ruff lint/format 269 files, strict mypy 168 modules, Pylance call compatibility, editor diagnostics, actionlint, wheel/sdist and diff checks pass. Actual four-node tests cover v2 enterprise ingestion, three replicas, search, owner/legal multi-hop deletion, peer-summary reports and tamper isolation. The final bilingual conformance checkpoint records 20 additional severity-ordered hardening findings and their corrections.

No High/Medium defect remains identified in the reviewed integration paths. This is not universal defect-free certification: physical LLM quality, optional vector implementation, long-duration/public-network load and cross-platform checks remain follow-up validation. Approved v1/read compatibility, operator-selected legal keys, N=3 capacity and unreachable-peer deadlines are explicit operating constraints.

Code and bilingual docs are ready for maintainer review. No commit, push, deployment or issue closure performed.