[virtue039]

The current extraction path collapses same-name subject references into one name-based mapping. This documentation PR proposes a source-mention protocol for #1136: per-mention refs, server fragments bound to a version/hash, subject/object/qualifier/alias references, and separate legacy/new entry points. It changes no production parser, prompt, public API or database. The proposal is not an accepted ADR.

Please review three boundaries in the internal protocol draft:

1. A ref identifies a source occurrence, not a cannot-link constraint or authorization for a model-supplied entity binding. The known list retains multiple same-name records.
2. The new mode validates all used references before graph writes. Broken references and cold-start identity ambiguity that cannot be represented honestly use existing unfinished signals with explicit coverage-loss accounting; they do not silently fall back to name-based resolution.
3. New-format entity objects may have their own locating references, but when that text may become identity context requires explicit agreement. Preserve the legacy object-only protection and the existing semantics of statements/null quotes, identity sources and memories awaiting confirmation.

Validation and limits:

• The reference prototype covers 33 scenarios in both input orders (66 rows): repeated sentences, PDF whitespace, Unicode, abbreviations, decimals and wrong-chunk/fabricated/dangling references. Expected rejections fail explicitly. This is not production-parser acceptance.
• Three prototype-validated examples were passed to the unchanged production resolver: two namesakes, one person holding two roles, and same-name subject/object mentions in one sentence. All 12 rows across seed and mention orders selected the expected existing IDs and created no entities. Four same-person controls retaining the old response_claims exclusions incorrectly split the person and each created one entity.
• New production extraction/apply and complete fact writes did not run. These results do not mean A02/A03 are fixed in production. The four historical error states contain only zero or one exact-name holder, so controls with two existing candidates do not establish that all four historical failures are solved.
• The 19 existing statements-related tests, 34 original namesake checks and 48 named adjudication measurements passed with DB/PDF requirement guards enabled. Original inputs, gold labels and named-failure lists are unchanged; all 24 baseline gold errors remain visible.
• New real-model/embedding calls: zero. No accuracy improvement is claimed. The independent cross-script adjacency experiment is outside this protocol implementation; recall and adjudication safety remain under A name in another script or a former name never meets its entity #709/Require source-grounded positive identity evidence before automatically merging namesakes #1193.

All six checks in CI 40640830074 passed on signed commit 004965709e583c1e8934f51711501d69f5c93927. The subsequent English-language revision translates the proposal and protocol draft, retaining original Chinese regression quotations and adding English explanations; it changes no protocol decision or executable code. Its checks are reported separately from that earlier run.

#1377 and #1425 have landed. #1193's source-input/second-look proposal still awaits agreement on scope and added human cost. #1192/#1199 remain closed on hold. After contract discussion and a landed ADR, the production change would be a separate PR with a same-session 16-KB comparison.

Refs #1136, #1193, #709.