fix: preserve historical memory source views

[dcellison]

Summary

• distinguish canonical ownership stamps from canonical transcript source locators
• keep complete historical JSONL provenance usable after Workshop memory ownership migration
• continue to fail closed when any canonical run/message source pointer is partially written

Installed finding

The Milestone 4 installed qualification found that the parser classified all
358 non-canonical rows as malformed because migrated ownership metadata shares
four field names with canonical source provenance. With this fix, the deployed
corpus classifies as 37 canonical locators, 247 historical locators, 111 rows
without a source locator, and zero malformed rows. Read-only resolution of a
representative canonical source and a representative historical source both
returned their exact owned user/assistant pair.

Tests

• make check
• make test (6,104 passed, 1 skipped)
• focused provenance/history tests (66 passed)

Follow-up to #1047 and #1056.