docs(import): record normalizer completion + dry-run results in worklog

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

docs/import-migration/WORKLOG.md

@@ -4,6 +4,46 @@ Running log of each working session. **Resume here.** Newest entry on top.
 
 ---
 
+## 2026-05-25 (session 4) — Built the normalizer (subagent-driven, all 17 tasks)
+
+**Did:** Executed the plan subagent-driven (implementer + spec review + code-quality review per
+task). The tool `tools/import-normalizer/` is **complete and passing (57 tests)**. Final
+opus review: **READY** — determinism verified on the real corpus (two runs → identical cell
+matrices + byte-identical review files), zero silent drops.
+
+**Per-task code review caught & fixed real issues** (all in the committed code): leading
+qualifiers `nach/vor/…` now → APPROX; English month-first matcher hardened to structurally
+not shadow `Mai 1895`; person-id collision de-dup suffixes *all* members; `split_receivers`
+returns `[]` for a `geb.`-only cell; boolean cells no longer coerced to `1/0`; duplicate-index
+flags every occurrence; provisional ids never steal a register id; CSV-injection defanged.
+
+**REAL DRY-RUN** (`python normalize.py` over the actual archive — outputs are gitignored):
+- documents_emitted **7,582** (+225 empty +93 blank-index +42 x-suffix = 7,942 rows read, 0 dropped)
+- register_persons **163**, provisional_persons **942**
+- dates: DAY 6,509 / MONTH 36 / RANGE 36 / APPROX 28 / YEAR 17 / SEASON 1 / UNKNOWN 955
+- **unknown_date_rate 9.2%** (of dated rows; target ≤5% pre-override, ≤0.5% after overrides)
+- duplicate_index 85, index_file_mismatches 550, ambiguous_receivers 303
+
+**⚠️ Concurrency incident:** a parallel Claude session committed reader-dashboard work to this
+branch and hard-reset it mid-execution, deleting the Task 15 files and orphaning a commit.
+Recovered via reflog (`reset --hard 366b4848` + `checkout 401160e3 -- <task15 files>`); no code
+lost. Casualty: my *during-execution* edits to the plan/spec docs (02/03) for Tasks 5–14 were
+discarded — **the committed code + tests are the source of truth**, not the plan doc, which now
+reflects the pre-execution + persona-review version.
+
+**Next steps (iterative refinement — the overrides loop, as designed):**
+1. Shave the 9.2% UNKNOWN cheaply: add **Spanish month names** (Enero…Diciembre) and the
+   `Mon DD-YYYY` dash form to `config.MONTHS`/the parser (Mexican-branch correspondence);
+   revisit the 58–72 two-digit-year band (real `…58/59/60` dates = 1958–1960, just past the
+   1873–1957 window — decide whether to extend the upper bound in `config`).
+2. `?` (99×) is genuinely "date unknown" — leave UNKNOWN or add a convention.
+3. Populate `overrides/dates.csv` + `overrides/names.csv` from the review CSVs and re-run.
+4. README note: a leading `'`/`!` in a `review/*.csv` `raw` cell may be a CSV-defang artifact —
+   match against the true source value when writing overrides.
+5. Phase 2 (separate spec): wire the canonical contract into the Java `MassImportService`.
+
+---
+
 ## 2026-05-25 (session 3) — Implementation plan + persona review
 
 **Did:**