docs(normalizer): add overrides/ README with structure + examples

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

tools/import-normalizer/overrides/README.md

@@ -0,0 +1,81 @@
+# Overrides
+
+Human corrections applied **deterministically on every run**. An override **wins** over the
+automatic date parser / name matcher, so this is how you fix the residue the tool can't resolve
+on its own. Two CSV files live here; both are read by `overrides.load_overrides()`.
+
+- Missing or header-only files are fine — they just contribute zero overrides.
+- Keep these files committed to git (they're your curated corrections); the generated `out/`
+  and `review/` folders are *not* committed.
+- Matching is **exact** on the `raw` value after trimming surrounding whitespace. Copy the
+  `raw` value verbatim from the matching `review/*.csv`.
+
+## The iteration loop
+
+1. Run `python normalize.py`.
+2. Open `review/unparsed-dates.csv` and `review/unresolved-names.csv` (sorted by frequency).
+3. Add correction rows here, then re-run. Repeat until the residue is acceptable.
+
+---
+
+## `dates.csv` — fix unparseable dates
+
+Header: `raw,iso,precision`
+
+| column | meaning |
+| --- | --- |
+| `raw` | the date string exactly as written in the spreadsheet (= the `raw` column in `review/unparsed-dates.csv`). |
+| `iso` | the corrected date as `YYYY-MM-DD`. For partial dates use the 1st: month-only → `YYYY-MM-01`, year-only → `YYYY-01-01`. Leave **empty** if truly unknown. |
+| `precision` | one of `DAY`, `MONTH`, `SEASON`, `YEAR`, `RANGE`, `APPROX`, `UNKNOWN`. |
+
+### Example
+
+```csv
+raw,iso,precision
+23.Juni 58,1958-06-23,DAY
+8.März 60,1960-03-08,DAY
+Mayo 18-1929,1929-05-18,DAY
+Abril 10-929,1929-04-10,DAY
+30.April,1909-04-30,DAY
+Mai 1895,1895-05-01,MONTH
+Herbst 1913,1913-10-01,SEASON
+1945/46,1945-01-01,RANGE
+um 1920,1920-01-01,APPROX
+?,,UNKNOWN
+```
+
+Notes:
+- `23.Juni 58` / `8.März 60` — two-digit years `58`/`60` fall in the parser's ambiguous
+  `58–72` band (just past the 1873–1957 window), so they aren't auto-parsed; here you assert 1958/1960.
+- `Mayo`/`Abril` — Spanish month names (Mexican-branch letters) the parser doesn't know yet.
+- `30.April` — month+day with no year; pick the year from the letter's context.
+- Empty `iso` + `UNKNOWN` records a deliberate "unknown date" (stops it showing up as residue).
+
+---
+
+## `names.csv` — map a name string to a canonical person
+
+Header: `raw,person_id`
+
+| column | meaning |
+| --- | --- |
+| `raw` | the sender/receiver name string exactly as written (= the `raw` column in `review/unresolved-names.csv`). For a multi-name cell that was split (e.g. `"Walter und Eugenie"`), use the **individual** name part. |
+| `person_id` | the canonical id to map it to. **Must be a real id** from the `person_id` column of `out/canonical-persons.xlsx` (a register person or an already-created provisional). |
+
+### Example
+
+```csv
+raw,person_id
+A.Klucke,klucke-anna
+? Hans de Gruyter,de-gruyter-hans
+Eltern Cram,cram-john-james
+Tante Lolly,blomquist-charlotte
+```
+
+Notes:
+- Use this for partial / misspelled / illegible / aliased names that should point at a known person.
+- It maps one string → **one** person. It does **not** split a two-person cell: for genuine
+  pairs like `Ella Anita` (flagged `ambiguous_pair`), there is no split-via-override yet — leave
+  them, or add both given names to `config.EXTRA_GIVEN_NAMES` so they keep getting flagged.
+- Look up valid `person_id` values in `out/canonical-persons.xlsx`. An id that doesn't exist
+  there will create a dangling reference (no validation yet).