fix(consolidation): apply sanitize_text to the _DedupDecision merge write path

[iRonin]

Follows up #2544.

@nicoloboschi on that issue:

the _DedupDecision merge path has no output-side scrub at all (merged_text = decision.text → straight to SQL). That's a real asymmetry worth a small dedicated fix (apply the existing character-safety sanitize_text to all three consolidation write paths)... a clean standalone issue I'd support.

This is that fix, scoped to the merge path.

Character-safety only

sanitize_llm_output is an alias of sanitize_text (llm_wrapper.py), which strips ASCII control characters and lone UTF-16 surrogates — the things that break json.loads and the Postgres UTF-8 encode. This is not the declined content sanitizer; no content filtering is added here.

The asymmetry

_CreateAction and _UpdateAction each scrub text through a sanitize_llm_output field validator. _DedupDecision has only _normalize_action, so the merge path handed the LLM's synthesized text to the fold UPDATE with just .strip() applied.

-        best_id=best_id, merged_text=decision.text.strip() or best_text, should_merge=True, best_text=best_text
+    merged_text = (sanitize_llm_output(decision.text) or "").strip() or best_text

The best_text fallback is preserved: a decision whose text sanitizes away to empty still folds into the twin's existing text rather than blanking the row.

Test

Adds test_dedup_llm_merge_sanitizes_text_before_write next to the existing test_dedup_llm_merge_folds_into_twin, which is left untouched. It injects a NUL into the merge decision and asserts the fold UPDATE argument is clean.

Verification

• The new test is red without the production line and green with it. The failure is on assert "\x00" not in args[1], and the other 41 tests in the file pass either way — so it detects this gap and nothing else.
• tests/test_consolidation_dedup.py — 42 passed.
• tests/test_consolidation*.py — 114 passed, vs 113 on unmodified main; the delta is exactly this test. Both runs report the same 83 setup errors, which are environmental: this checkout lacks the optional sentence-transformers extra.
• tests/test_llm_wrapper.py — 47 passed.
• ruff check clean, ruff format --check clean, ty check hindsight_api/ reports no diagnostics in consolidator.py.

[iRonin]

Heads-up on the red check, since it doesn't come from this PR.

verify-generated-files — the out-of-sync file is skills/hindsight-docs/references/openapi.json (+16/−2). It was last regenerated in 218e6d3 (2026-07-30); 1ce308f (#3109, 2026-07-31) then changed the bank/document response models (adds last_write_at, reworks last_document_at) without a regen, and the diff shape matches those fields. It reproduces on any PR against main and is unrelated to this change, which only touches a function body in consolidator.py plus a test. I've deliberately kept the regeneration out of this PR rather than mix an unrelated generated file into a two-file fix — ./scripts/generate-docs-skill.sh on main should clear it, and I'm happy to leave that to you.

On test evidence — test-api is gated on detect-changes.outputs.has_secrets == 'true', so it can't run on a fork PR; the green checks here don't exercise the new test. The verification section in the description is the evidence I have: the test is red without the production line (failing on assert "\x00" not in args[1]) and green with it, and the consolidation suite goes 113 → 114, so the delta is exactly this test. A maintainer run with secrets would be welcome.

[nicoloboschi]

Approve. Correct, minimal fix at the single _dedup_adjudicate chokepoint — covers both reconcile-create and reconcile-update fold paths. Fallback (or best_text) and .strip() semantics preserved; test asserts the char is removed, not just that the text was dropped. Character-safety only, consistent with the declined-content-sanitizer decision on #2544/#2547.