test(retrieval): dense-leg-only regression pin for table cell-value query

[henryle97]

This was generated by AI during triage.

Follow-up from #43 and PR #45.

The table half of #43 is implemented: the table branch in Retrieval Text composition now includes raw table_text before any summary/question augmentation, so literal cell values reach the dense embedding payload.

What remains unshipped is the explicit regression proof that the dense leg alone retrieves the gold table chunk for a pinned cell-value query without BM25 masking the result.

This follow-up tracks only that verification slice.

[henryle97]

This was generated by AI during triage.

Agent Brief

Category: enhancement
Summary: Add a dense-leg-only retrieval regression pin for the table cell-value failure mode that ADR 0002 was meant to fix.

Current behavior:
The Retrieval Text composer now augments table chunks with raw cell text, so the implementation-side bug from #43 is fixed. However, there is no hermetic regression test that isolates the dense leg and proves a pinned cell-value query retrieves its gold table chunk without help from BM25. The benchmark history already showed why this matters: hybrid metrics can stay flat or move for unrelated reasons, and BM25 can mask whether the dense payload change actually shipped.

Desired behavior:
The test suite should contain one deterministic regression pin for the table cell-value case. Given a pinned cash-balance-style table query against a stable fixture and a deterministic embedding stub, the dense retrieval leg alone should rank the gold table chunk in the top-5 when BM25 is disabled or absent. The test should fail if raw table cells stop participating in Retrieval Text composition.

Key interfaces:

• The Retrieval Text composer for table chunks — its externally visible contract is that raw table cell text participates in the dense embedding payload ahead of optional summary/question augmentation.
• The retrieval test harness for dense or hybrid retrieval — it should support a dense-only assertion path so BM25 cannot hide regressions in the embedding payload.
• The benchmark-style chunk fixture used by retrieval tests — it needs one stable gold table chunk and one pinned cell-value query that exercises literal table-cell matching.

Acceptance criteria:

• A hermetic test covers one pinned table cell-value query against a deterministic retrieval fixture.
• The assertion is on the dense leg only: BM25 is disabled, absent, or otherwise prevented from contributing to the ranking.
• The gold table chunk ranks within dense top-5 for the pinned query.
• The test would fail if raw table_text stopped contributing to table Retrieval Text.
• No network access or live provider keys are required.

Out of scope:

• Changing Retrieval Text composition again — this issue is verification, not another policy change.
• BM25 token-source tuning, reranker behavior, or hybrid fusion changes.
• The content_with_weight writer hazard tracked separately in enrichment: augment-on-write contract for content_with_weight (latent ordering hazard) #44.