[shtse8]

Markdown examples can describe another API edition bundled in the same npm package. Docs search previously admitted all prose and could quote an incompatible setup as parent context. Sections now inherit example imports before chunking, and retrieval and parent context use the same import scope. Unknown or mixed-edition examples are excluded conservatively; identical entrypoints and transparent barrels share scope, while root package searches also admit known public helpers whose exported names are disjoint from the root surface.

Questions naming an exact public import narrow a root search without overriding an explicit alternate selection. An explicit call resolves through that entrypoint before same-named transitive type members: on Next 14, cookies() from next/headers now returns cookies(): ReadonlyRequestCookies first instead of NextApiRequest.cookies.

Includes index format 16 and synchronized version 0.7.5 so the normal release workflow publishes the repair. Homebrew/Scoop previously stayed on 0.7.1 while their release job succeeded after skipping absent App credentials. The release now consumes published mcp-kit 0.8.7 and runs its OIDC/Broker publisher in a caller-owned job with explicit destinations and failure on missing authorization. Separate tokens cover only the tap and bucket; their manifests now travel through normal PRs, required checks and merge queue, followed by exact merged-byte readback. No long-lived App key reaches the runner.

Validation: core and workspace tests, clippy, manifest checks and benchmark grader tests passed locally. Twenty-one public-API regressions cover edition inheritance, chunks, repeated headings, child overrides, aliases, unknown/mixed imports, caches and explicit calls. The real-package CLI/MCP check passes three Zod versions and Next 14, including the new scope cases. The caller workflow parses and its version step executes successfully. The unchanged whole benchmark is checked separately: the first repair scored74/105 versus fresh unchanged-main76/105. Those regressions traced to over-filtered public helpers and duplicated citation paths consuming the answer budget; both have concrete local controls. Revised run41369000899 recovers all tuning regressions: first-use/fetched76/105 matches fresh main; keyword49 vs48, hybrid51 vs49, fetched-keyword73 vs72. The existing96 floor still fails; held-out fetched aggregate13 vs14 remains a limitation, without held-out diagnosis or tuning. No ranking weights changed; the React form-action paraphrase remains a measured miss. Actual package-manager delivery still requires the release's CI workload principal and narrow Broker policy to be verified on main.