Wire temporal-retrieval-filter into live retriever / re-ranker

[mdheller]

Follow-up runtime work for the temporal fact-supersession contract landed in #20 (schema + reference filter + teeth). The contract and reference temporal_retrieve filter exist; this issue is the runtime wiring into the live retrieval path, which is out of scope for the contract PR.

Context

• Contract: schemas/search/temporal-fact.schema.json
• Reference filter: tools/validate_temporal_supersession.py :: temporal_retrieve(facts, entity, relation)
• Behavior: high-recall candidate pass → suppress facts marked superseded (superseded_by/superseded_at) → most-recent (max valid_from) surviving fact per (entity, relation) is authoritative.
• Cross-ref: prophet-workspace#76

Runtime work

1. Project graph edges / epistemic-edge records into temporal-fact records at index time (map valid_time.from/to / temporalScope.validFrom/validTo → valid_from/valid_to; carry superseded_by/superseded_at from ACR decision-ledger / epistemic-edge promotionState: superseded).
2. Apply the supersession suppression in the retriever's high-recall pass (exclude, or hard-downrank, facts marked superseded).
3. Apply max-valid_from-wins tie-break per (entity, relation) in the re-ranker.
4. Keep claims_canonical_truth = false end to end; canonical authority stays with ACR / epistemic-edge.

Acceptance criteria

• For the John-Smith→Jenna-Brown fixture running against the live retriever, the query "Who is the CEO of ABC?" returns Jenna Brown, and the superseded John-Smith fact is excluded (or ranked below all surviving facts).
• A (entity, relation) with no supersession returns its single fact unchanged.
• Both facts are present in the high-recall candidate set before suppression (suppression is genuine, not an indexing gap).
• Malformed temporal records (missing valid_from, superseded_at < valid_from, valid_to < valid_from) are rejected at index time — the live path reuses/mirrors validate_temporal_supersession.py.
• An offline eval shows temporal-answer accuracy improves vs. the semantic-only baseline on a supersession test set (report granular before/after).