v7.55.2

Fixed

• PDF extraction: duplicated content on multi-page documents. MinerU's end_page_id is inclusive, and nexus was passing a half-open range end. With the default one-page batching, every page but the first and last was extracted twice and both copies were concatenated into the document text. Documents indexed before this release carry the duplicate text: it is chunked, embedded and searchable, so the same passage can be returned more than once and crowd out other results. Nothing is lost or corrupted — identical chunk text collapses to a single stored row — so the symptom is a document whose chunk count exceeds its distinct chunks. To correct an affected document, re-index it (nx index pdf --force, or nx dt index --force for DEVONthink records); re-embedding is skipped for any chunk whose text is unchanged.

• The MCP stdio connection died whenever a request was cancelled. The mcp SDK's RequestResponder.cancel() answers a request it has just cancelled, but the client has already discarded that id, so the reply arrives as an unknown message id and the client tears down the whole transport — every tool on the server at once, with no auto-reconnect for stdio. Measured across 121 client log files: ten teardowns, five of them carrying Request cancelled. nexus now corrects the SDK at server startup so a cancelled request is answered with silence, which is what the SDK's own respond() already assumed. The patch verifies the defect is present before applying and no-ops if a future SDK fixes it. (nexus-rjmyk)

• Aspect rows built in batches carried no source_uri, so 481 of 2232 matched nothing. The two single-document builders minted the URI and the batch builder did not, because it was written as a transform of the response entry and source_uri is derived rather than returned. The aspect operators re-derive that key and match it byte-equal, so an absent one reported as a content verdict — "does not match" — rather than as an error. (nexus-vnz3d)

• A failed taxonomy rebuild left a collection with no centroids and no signal. rebuild_taxonomy deleted the existing centroids before computing the plan that replaces them, so any raise in between emptied the collection permanently. The empty state reports healthy through every layer: the nearest-centroid query returns no rows, the unmatched-chunk list is computed from chunk existence rather than assignment and so comes back empty, and the client tripwire only fires on a non-empty list.

• nx mineru status printed ? for every count: it read an active_tasks key the MinerU server's /health has never sent, while the server serves processing_tasks, queued_tasks and completed_tasks. The test fixture invented the same key, so the suite stayed green.

• A development checkout no longer sends the anonymous install ping. install_ping had three off-switches — NX_NO_TELEMETRY, telemetry.enabled and the 24h throttle — and none asked whether the process was running out of a git checkout, which nexus.db.service_endpoint.is_dev_checkout_process() already answers for the production-write guard. Container-based test harnesses run an installed wheel rather than a checkout, so NX_NO_TELEMETRY=1 remains required there and is enforced by a lint over every docker run under tests/e2e. (nexus-6doho)

Changed

• The install shell under src/nexus/_install/ now dispatches to stdlib-only Python cores instead of restating their logic. layout.sh, census.sh, gc.sh and shims.sh become thin dispatchers over layout_core.py, census_core.py, gc_core.py and shims_core.py; the cores import nothing from nexus, because the installer runs with nothing installed. The duplication being removed was not theoretical — collapsing it fixed a source-classification rule that had no Python twin at all, a census that under-reported holders for a relative path, and a shim pruner that treated a hostile-named shim as owned and so kept it forever while the writer refused to write it and nx doctor never walked it. (nexus-utpuw)

Engine

• Pinned engine identity unchanged: engine-service-v0.1.129.