Tasks

1. Adapter implementation

• 1.1 Encode the hit connection into search result ids ({connection_id}/{stream}:{record_id}) in normalizeSearchResults, wrapping only record-shaped base ids.
• 1.2 Parse both id grammars in fetch (parseRecordResultId splits the connection segment at the first /; every segment passes requireSafeName) and forward the embedded connection as the canonical connection_id query parameter.
• 1.3 Reject a self-contained id whose embedded connection disagrees with an explicit connection_id argument with a typed conflicting_connection_id error before any RS call.
• 1.4 Stop repeating connection_id= in search content[] preview lines when it is embedded in the id; raise the preview id bound from 80 to 200 chars so handles stay complete; update the fetch hint line.
• 1.5 Carry the embedded connection into constructed citation URLs (?connection_id=...).
• 1.6 Update tool descriptions, the fetch id input description, the search output-schema description, server instructions, and the package README for the new grammar.

2. Regression coverage

• 2.1 Multi-source journey test per the model-visible-journey canon rule: consume only content[] text, extract the id, complete search→fetch with no connection_id argument against a fixture whose unscoped record reads return typed 409 ambiguous_connection (test/self-contained-result-id.test.js).
• 2.2 Backcompat pins: legacy stream:record_id unscoped still surfaces the typed 409; legacy id + explicit connection_id unchanged (new file plus existing connection-id-forwarding.test.js).
• 2.3 Conflict, traversal/malformed-segment, matching-explicit-argument, and no-connection passthrough pins.
• 2.4 Update existing pins to the new contract (server.integration.test.js, record-payload-token-budget.test.js) and re-run the footprint/token-budget suites within their budgets.

3. Validation

• 3.1 Full packages/mcp-server suite green (127 tests; was 120).
• 3.2 openspec validate --all --strict green.