Add bounded knowledge retrieval

[Jesper Schulz-Wedde (JesperSchulz)]

Problem

Review skills repeatedly carried broad knowledge-index/model context and relied on report transport that could be incomplete or silently repaired. That made retrieval expensive, continuation state weak, and leaf-report acceptance difficult to validate deterministically.

Solution

• Add lossless deterministic catalog paging capped at 16,000 serialized UTF-8 bytes, including the output newline.
• Return every applicable or conditionally applicable catalog row without ranking, top-k filtering, query omission, or basename deduplication.
• Retrieve exact article or sample bodies in stable chunks of at most eight paths, preserving full text, byte count, and SHA-256 with strict UTF-8 and path-safety checks.
• Treat Entry's prepared v1 index (sourceSnapshot / sourceSha256) as the catalog freshness boundary; bind continuations to the exact raw index and normalized request/candidate state while validating selected live bodies on retrieval.
• Make applicability and unknown dimensions explicit in the additive v2 catalog format.
• Warn and skip invalid partner articles per file so valid siblings remain usable, while retaining complete search-time validation for corrupted or external prepared indexes.
• Give each review leaf private GUID-scoped scratch/report paths and require deterministic validation of the exact Task return before rollup.
• Apply the generic bounded Source workflow to all 16 canonical Microsoft review leaves without changing their Relevance, Worklist, Action, or Output rules.

Safeguards

• No ranking, top-k selection, truncation, summarization, query-based omission, or other catalog/body data loss.
• No runner framework, scanner, evidence ledger, model policy or allocation, domain shortcut, Shopify fixture, or domain-specific analyzer rule.
• Invalid source metadata, noncanonical sentinels, invalid version ranges, unsafe paths, stale continuations, selected-body changes, broken sample links, oversized batches, and invalid UTF-8 fail explicitly or warn/skip at the documented boundary.
• PR Separate deterministic analyzer rules from BCQuality #178's analyzer/knowledge separation is preserved: no deleted knowledge article or sample is restored, including the obsolete ApplicationArea guidance.

Validation

• Knowledge index: 282 current articles, deterministic and complete.
• Retrieval: all 282 articles and 457 samples round-tripped exactly; bounded pages, applicability, continuations, prepared-index boundaries, mutation detection, malformed/stale/unsafe/oversize failures, strict UTF-8, invalid sibling skip, semantic metadata parity, and layer validation covered.
• Frontmatter validator: 0 errors, 0 warnings.
• Review fixtures: 34 cases across 17 leaf domains.
• All changed PowerShell files parse successfully.
• Tracked and untracked whitespace checks and git diff --check pass.
• All 16 leaf diffs remain Source-only, all 52 upstream deletions since the reviewed old base remain absent, and no generated artifacts are included.

Measured-experiment caveat

Experiments motivated these conservative bounded-retrieval and exact-transport mechanics, but they did not prove a production model mix or quality equivalence. This change therefore includes no model allocation or model-policy decision.