Correct seven errors in the v2 card

An independent verification pass over the published artefacts found the
previous revision of this card wrong in seven places.

* The transcript-overlap row compared two different languages as though it
were one language's before and after. dga went 19.1% -> 0.2%; the 1.5%
figure is ach, which is a regression from 0.6%. Both are now shown
explicitly, including the regression.
* The warning against reading only one v1 split claimed it reintroduces
speaker leakage. It does not -- v2 moves whole speakers, so any subset of
one split stays disjoint from the others. The real damage is losing 88% of
the evaluation set, silently.
* Streaming was described as fetching almost nothing. There is no predicate
pushdown and audio is inline, so it transfers the whole labelled half:
measured, Shona's 1,749-row test set costs 552 MB and ~8 minutes.
* The unlabeled split was given as 816 GB in one place and 637 GB in
another. It is 637 GB.
* sample_rate and channels were described as measured per shard. They vary
per row -- one amh row group holds 14 stereo and 6 mono clips -- and the
corpus includes 44.1 kHz, which went unmentioned.
* The gender table did not say it counts only labelled rows, that it is
reference-word weighted (per-utterance is much looser), or that the raw
field needs case-folding.
* The 12-hour and 2,000-utterance eval bounds were called caps. They are
targets with slack: 14 of 38 eval splits exceed 2,000 utterances.

data/ASR_v2/README.md

@@ -101,7 +101,9 @@ asserted from the code that produced them.
 |---|---|---|
 | Languages whose eval speakers also appear in train | 14 of 19 | **0 of 19** |
 | Worst eval speaker leakage, any language | 100.0 % | **0.0 %** |
-| Worst train/eval transcript overlap | 19.1 % (`dga`) | **1.5 %** (`ach`) |
+| Corpus-wide eval utterances with an exact twin in train | 1.09 % | **0.094 %** |
+| Leakiest language in v1, `dga` | 19.1 % | **0.2 %** |
+| Leakiest language in v2, `ach` | 0.6 % | 1.5 % |
 | Languages with gender labels | 12 of 19 | 12 of 19 |
 
 **Speaker-disjointness is structural, not statistical.** The unit of assignment
@@ -109,9 +111,12 @@ is the speaker, so an utterance cannot cross a split boundary without its
 speaker. It is not a threshold that was met; it is a property the construction
 cannot violate.
 
-**Sizes are 80/10/10 by duration, with a 12-hour cap per eval split.** The cap
-binds on the six largest languages, which therefore evaluate on roughly 5.4 %
-per side rather than 10 %; the surplus goes to train. Corpus-wide this comes to
+**Sizes are 80/10/10 by duration, with eval splits targeted at ~12 hours.**
+That target binds on the six largest languages, which therefore evaluate on
+roughly 5.4 % per side rather than 10 %; the surplus goes to train. It is a
+target, not a hard ceiling -- five eval splits land between 12.0 and 12.2 h,
+and the nominal 2,000-utterance bound carries 35 % slack, so 14 of 38 eval
+splits exceed 2,000 utterances (largest: `tir` validation at 2,686). Corpus-wide this comes to
 84.5 %/7.7 %/7.7 % by duration. Per-language figures are in the table below.
 
 **What v2 does not change.** No quality filtering, no re-encoding, no
@@ -245,18 +250,24 @@ for v1_split in ("train", "validation", "test"):        # all three -- see below
 ```
 
 `pd.read_csv` on an `hf://` path needs `huggingface_hub` installed (it is, if
-`datasets` is). Streaming is recommended: an eval split is a few hundred rows
-out of tens of thousands, so you fetch almost nothing. Materialising instead
-(`streaming=False`) downloads the language's entire labelled half first --
-1.6 GB for `ach`, 12.7 GB for `sid`.
+`datasets` is). Pass `dtype=str` -- some ids begin with an underscore.
+
+**Budget the transfer.** Streaming does not save bandwidth here: there is no
+predicate pushdown and the audio bytes are inline in the same row groups, so
+filtering still pulls the language's whole labelled half. Measured: Shona's v2
+test set is 1,749 rows but costs 552 MB and about 8 minutes. Reconstructing all
+three splits of one language ranges from 1.6 GB (`ach`) to 12.7 GB (`sid`).
+Streaming's advantage is bounded memory, not bounded download.
 
 ### Two mistakes that fail silently
 
 **Read all three v1 labelled splits.** Utterances move between them. Acholi's
 v2 `test` set of 515 rows is drawn 397 from v1 *train*, 58 from v1 validation
-and 60 from v1 test. Reading only v1's `test` split returns 60 rows instead of
-515 -- and every one of those 60 speakers is in v1 train, so you reintroduce
-exactly the leakage v2 removes. No error is raised.
+and 60 from v1 test. Reading only v1's `test` split returns **60 rows instead
+of 515** -- 88 % of your evaluation set silently missing, with no error. (It
+does *not* reintroduce speaker leakage: v2 moves whole speakers, so any subset
+of one v2 split stays disjoint from the others. The damage is that you are
+measuring on a twelfth of the data, and on a non-random twelfth.)
 
 **Never read the `unlabeled` split.** It carries no transcripts, takes no part
 in the re-split, and is by far the largest thing in the repository (637 GB
@@ -299,7 +310,7 @@ The six v1 columns, unchanged, plus derived fields:
 | `audio` | `struct<bytes, path>` — **byte-identical** to v1 |
 | `v1_split` | which v1 split this utterance came from |
 | `duration_s`, `n_bytes` | derived; the corpus is 128 kbit/s CBR MP3 |
-| `sample_rate`, `channels` | measured per shard; **heterogeneous** (see limitations) |
+| `sample_rate`, `channels` | measured **per row**; heterogeneous (see limitations) |
 | `provider`, `licence` | the licence travels with the row |
 
 The stray `__index_level_0__` column present in 36 of the v1 shards is dropped,
@@ -311,13 +322,15 @@ and every v2 file uses 100-row row groups.
 
 - **No gender labels at all** for `aka`, `dag`, `dga`, `ewe`, `ful`, `kpo`, `mlg`.
   The gender goal is undefined there, not satisfied, and is reported as `n/a`.
-- **Heterogeneous audio format.** The corpus mixes 16 kHz mono and 48 kHz
-  stereo MP3 between languages. v1 documents neither; v2 records the measured
-  values per row but does **not** resample, because a silent normalisation
-  cannot be verified against the original.
+- **Heterogeneous audio format**, and it varies *within* a shard, not just
+  between languages: 16 kHz mono (`ach`, `lug`), 44.1 kHz stereo (`aka`, `dag`,
+  `dga`, `ewe`), 48 kHz mono (`ful`, `lin`, `sna`, `tir`, `wal`, most `amh`)
+  and 48 kHz stereo (some `amh`). One `amh` test row group holds 14 stereo and
+  6 mono clips. v1 documents none of this. Nothing is resampled -- a silent
+  normalisation cannot be verified against the original.
 - **This is a re-split, not a clean-up.** No quality filtering has been
   applied. Utterance-level defects present in v1 are present here.
-- **The `unlabeled` split is out of scope** (~816 GB, no transcripts). Take it
+- **The `unlabeled` split is out of scope** (637 GB, no transcripts). Take it
   from the v1 config, which is unchanged.
 - **`speaker_id` is trusted, not verified acoustically** unless the manifest
   says otherwise. If an id turns out to cover more than one voice, the