diff --git a/RECURRING_BUGS_REPORT.md b/RECURRING_BUGS_REPORT.md
new file mode 100644
index 0000000..7fb3385
--- /dev/null
+++ b/RECURRING_BUGS_REPORT.md
@@ -0,0 +1,176 @@
+# Recurring-Bug Report — data-ingestors (2026-05-18 → 2026-06-18)
+
+> Scope: bug classes that were fixed, then **reappeared** in a different file / use-case / layer
+> within the last month. Built from `git log` on `develop` + the Cursor Bugbot threads.
+> The point isn't the individual fixes — it's that the same *class* keeps coming back because
+> the fix was applied at one site instead of at the shared seam.
+
+**TL;DR — six bug classes account for ~50 of the month's fixes.** Four are effectively closed;
+**two are still live** (Class 1 in 7 validators, Class 6 latent). Each recurs because the
+invariant lives in prose, not in one enforced place.
+
+---
+
+## Class 1 — Preflight validator reads data differently than the ingestor  ⚠️ STILL LIVE
+
+**Invariant that keeps breaking:** a validator must read the CSV/JSON with the *exact* same
+rules `CSVIngestor`/`JSONIngestor` use, or validation and ingestion disagree
+(validate-pass → ingest-fail, or valid data rejected).
+
+The canonical contract: `keep_default_na=False` + `coercion.build_csv_na_values(schema)`,
+header `.str.strip()`, case/whitespace-insensitive column resolution, `on_bad_lines="error"`,
+string-dtype pin for string columns, **fail-closed** on read errors.
+
+| When | PR(s) | Instance |
+|---|---|---|
+| 06-08 | #166 | type-convert + strip headers on **every** chunk, not just the first |
+| 06-09 | #191 (bugbot #190) | duplicate headers / ragged rows / whitespace-header dtype pin |
+| 06-09 | #204 | align per-record JSON validation with CSV |
+| 06-15 | #242 | single source of truth for NA policy (`build_csv_na_values`) |
+| 06-15 | #262 (#261) | strip whitespace from label values (silent class duplication) |
+| 06-15 | #252 | **LabelDiversityValidator — six consecutive bugbot rounds**: fail-closed on read error; partial-read skew; NA rules; stripped-vs-full schema; whitespace headers; pandas-native header parse |
+| 06-17 | #292 (#289) | fail fast on missing schema columns |
+
+**Still unfixed (this week's audit), all using bare `pd.read_csv(..., on_bad_lines="warn")`:**
+- `BaseValidator._load_data` — `validators/base.py:147` (shared fallback; also swallows read errors → `None` → benign skip)
+- `BIOLabelValidator` — case-only column resolve, no NA policy, no header strip
+- `TimeFormatValidator` :129 · `TimeOrderedValidator` :78 · `TimeBeforeTodayValidator` :81 · `NumericColumnsValidator` :137 · `TimeToEventValidator` :216
+
+Two concrete sub-bugs still present in 6 validators:
+- `on_bad_lines="warn"` (vs ingestor `"error"`) — a ragged row is silently dropped at validation but **errors at ingest**. This *is* the #190 bug, never propagated past `DataValidator`.
+- `BaseValidator._load_data` swallows parse errors → "no data" skip while ingest hard-fails. This *is* #252 finding-1, never propagated to the base.
+
+**Why it recurs:** every validator re-implements the read. **Permanent fix:** one shared
+`read_for_validation(path, schema)` helper; route all validators through it; one parity test
+vs `CSVIngestor`.
+
+---
+
+## Class 2 — NULL / NA / empty-cell semantics  ✅ mostly closed (via Class 1 consolidation)
+
+**Invariant:** null-like tokens (`""`, `NA`, `null`, `None`, `NaN`, JSON `null`) must
+round-trip to SQL `NULL` identically across CSV, JSON, every column type, and both the
+validator and the ingestor — never become the literal string `"nan"`/`"None"` and never be
+counted as a type violation.
+
+| When | PR(s) | Instance |
+|---|---|---|
+| 06-06 | #150 | stop counting missing values as "non-numeric" |
+| 06-08 | #167 | VARCHAR/CHAR/TEXT must tolerate NULL |
+| 06-08 | #168 | extend NULL-tolerance to DATE/DATETIME/TIMESTAMP/TIME |
+| 06-08 | #170 | JSON `null` tolerate as missing, not a type error |
+| 06-08 | #172 | VARCHAR/CHAR empty cells → `None`, not literal `"nan"` |
+| 06-08 | #176 | JSON `""` → NULL in `process_record` |
+| 06-08 | — | null-like values must round-trip end-to-end as SQL NULL |
+| 06-09 | #205 (#195) | tolerate nulls in time-series feature columns |
+| 06-15 | #242 (#236/#237) | NA policy + int64 range → single source of truth |
+
+**Why it recurred:** the NA set was redefined per-type and per-layer until `coercion.NA_SENTINELS`
+became the single source (#242). Now centralized — but only the layers that *call* it benefit,
+which is why Class 1's stragglers still diverge.
+
+---
+
+## Class 3 — Silent half-ingest / wrong success accounting  ✅ closed
+
+**Invariant:** if any record is dropped, any file transfer fails, or backend registration fails,
+the run must **not** report success (exit non-zero, surface counts) — a green run must mean a
+complete dataset.
+
+| When | PR(s) | Instance |
+|---|---|---|
+| 05-19 | #100 (#99) | surface file_transfer failures in the ingestion summary |
+| 06-09 | #187 | fail loud when backend dataset registration fails |
+| 06-11 | #223 | count batch API-send failures as failed records, exit non-zero |
+| 06-12 | #230 | propagate hard ingest failures in all modalities; send only inserted records |
+| 06-15 | #243 (#234/#235) | dropped records fail the run; JSON read-layer fails fast |
+| 06-15 | #265 (#260) | defer table creation until validation passes |
+| 06-17 | #279 | regression test: validator-rejected ingest creates no table |
+
+**Why it recurred:** "success" was implicit (no exception thrown) rather than asserted against
+the record ledger. Each modality/layer had its own escape hatch. Largely closed by
+`IngestionSummary.has_failures` + the deferred-table-creation guard.
+
+---
+
+## Class 4 — Schema-type vocabulary diverges across the 3 layers  ✅ closed, fragile
+
+**Invariant:** the DDL layer (`Database._get_sqlalchemy_type`), the preflight
+(`DataValidator.type_validators`), and the cast (`CSVIngestor._validate_csv`) must accept the
+**same** SQL-type vocabulary. When one knows a type the others don't, valid schemas are
+rejected with "Unknown data type" — or worse, a typo hint suggests a wrong-but-close type.
+
+| When | PR(s) | Instance |
+|---|---|---|
+| 06-09 | #192 | accept `NUMERIC` (alias for `DECIMAL`) |
+| 06-09 | #203 | VARCHAR/CHAR/TEXT accept numeric scalars |
+| 06-15 | #249 | `CHAR(N)` must map to a SQLAlchemy type |
+| 06-15 | #241 | report real MySQL column types from `get_table_schema` |
+| 06-15/17 | #264 / #293 | TINYINT/SMALLINT/MEDIUMINT/BLOB/LONGBLOB + "Did you mean …?" (bugbot caught BLOB→BOOL) |
+
+**Why it recurs:** three hand-maintained type maps. **Permanent fix (not yet done):** derive all
+three from one type registry, or a test that asserts the three key-sets are identical.
+
+---
+
+## Class 5 — semantic_segmentation `mask_id` cross-layer indirection  ✅ closed
+
+**Invariant:** `mask_id` is a per-row pointer to a mask **file**, not a DB column; it must reach
+`file_transfer` but never the insert — yet must be stored when the schema explicitly declares it.
+
+| When | PR(s) | Instance |
+|---|---|---|
+| 06-09 | #207 (#196) | honor `_mask` suffix in file pairing |
+| 06-09 | #212 | preserve `mask_id` through `process_record` (+ pop before insert) |
+| 06-17 | #280 (P5) | move `mask_id` off the DB-bound record (lend-then-strip in `map_file_transfer`) |
+| 06-17 | #282 | e2e semseg case + drop the #136 xfail (superseded #277/#285) |
+
+**Why it recurred:** a value that's "data for one layer, metadata for the next" with no single
+owner. #280 finally gave it an owner (the transfer lends it from the raw record). Watch for
+regressions whenever the record/transfer boundary moves.
+
+---
+
+## Class 6 — Test stub/mock signature drift  ⚠️ LATENT
+
+**Invariant:** when a production function gains a parameter, every test stub standing in for it
+must gain it too — otherwise the stub `TypeError`s (swallowed by a `try/except`) and the test
+passes **without exercising the real path** (false green).
+
+| When | PR(s) | Instance |
+|---|---|---|
+| 06-17 | #283 | drop dead `max_retries` param (stubs/ctors lagged) |
+| 06-18 | #305 (bugbot #304) | `map_file_transfer` stubs missing `source_record` (and earlier `cfg`) — 3 NLP-profile tests passed without running file transfer |
+
+**Why it recurs:** `map_file_transfer` grew `cfg=` then `source_record=` across P4/P5; stubs are
+hand-written lambdas with no link to the real signature. This week's audit found all current
+stubs now correct — but the pattern reappears every time that signature changes. **Mitigation:**
+prefer `autospec=True` / a shared fake bound to the real signature, or a single
+`fake_map_file_transfer` helper.
+
+---
+
+## Adjacent recurring theme — Docs/examples drift from the validator contract  ✅ ongoing
+
+Not a code bug, but the same shape: shipped YAML/README examples that the validators reject.
+#131 A-series, #197/#208 (text_classification wrong schema block), #206 (#198/#199 sample
+defaults), #200 (timestamp VARCHAR→TIMESTAMP), #224 (token_classification reserved `filename`),
+#290/#296 (tsf README). **Mitigation:** the `test_template_equivalence` / e2e parity suites are
+the right home; every shipped example should be asserted ingestable.
+
+---
+
+## Recommendation — convert "fix-in-a-loop" into "fix-at-the-seam"
+
+| Class | Status | One-time fix that ends the loop |
+|---|---|---|
+| 1. Validator read divergence | ⚠️ live (7 sites) | one `read_for_validation(path, schema)` + parity test vs CSVIngestor |
+| 2. NULL/NA semantics | ✅ | done (`coercion.NA_SENTINELS`) — finishes when Class 1 routes through it |
+| 3. Silent half-ingest | ✅ | `IngestionSummary.has_failures` + deferred table creation |
+| 4. Type vocabulary | ✅ fragile | derive 3 maps from 1 registry, or key-set equality test |
+| 5. mask_id indirection | ✅ | single owner (transfer lends from raw record) |
+| 6. Stub signature drift | ⚠️ latent | `autospec=True` / shared fake for `map_file_transfer` |
+
+**Highest leverage:** Class 1. It is the same bug as Classes 2 and 4 viewed from the read layer,
+it is the one that keeps generating bugbot rounds, and it is the only one still actively live in
+production code (7 validators). Consolidating the read kills the most recurrence for the least change.