Reconstructed text inserts spurious whitespace around quotes, colons, and abbreviation periods

Part of #5.

Comparing the corpus's `text` layer against the raw Brown plaintext ([`brown_nolines.txt`](http://www.sls.hawaii.edu/bley-vroman/brown_nolines.txt)) and token boundaries against the tagged Brown corpus ([`browntag_nolines.txt`](http://www.sls.hawaii.edu/bley-vroman/browntag_nolines.txt)), several kinds of punctuation that should sit flush against neighbouring text have a spurious space inserted around them in this corpus's reconstructed `text`.

Verified by walking each sentence's `tokens` spans and checking whether adjacent tokens are actually contiguous in `text` (i.e. whether a gap was inserted), cross-checked against `browntag_nolines.txt` to confirm the true Brown tokenization.

### Quotation marks

Brown tokenizes `` ` ` `` (opening) and `''` (closing) as their own tokens, but they sit directly against the quoted word with no space in the real text. In this corpus a space is very often inserted on the inside of the quote:

- `data/press_reportage/br-a01.yaml`: `" no evidence "` — raw Brown: `"no evidence"`

Measured across all of `data/`:
- **5,819 / 7,592** opening quotes (pos `` ` ` ``) have a spurious space after them.
- **4,649 / 5,043** closing quotes (pos `''`) have a spurious space before them.

### Numeric ratios / times with a colon

Brown tokenizes things like `11:30` or `1:512` as a single `CD` token. This corpus splits them into separate tokens with a space inserted around the colon:

- `data/press_reportage/br-a24.yaml`: `11: 30` — raw Brown: `11:30` (single `CD` token `11:30_CD`)
- `data/learned/br-j09.yaml`: `1: 512`

Measured: **189 / 189** digit-colon-digit sequences found in the corpus have a surrounding gap — i.e. this is essentially universal, not an occasional glitch.

### Periods inside abbreviations

Brown tokenizes abbreviations like `a.m.`, `i.e.`, `mm.` as single tokens (e.g. `a.m._RB`). This corpus splits them, inserting a space after the internal period:

- `data/press_reportage/br-a24.yaml`: `11: 30 a. m. yesterday` — raw Brown: `11:30 a.m. yesterday` (`a.m._RB`)
- `data/belles_lettres/br-g02.yaml`: `(i. e., societal)` — raw Brown: `i.e.` as one token
- `data/learned/br-j06.yaml`: `6 mm. i.d. Pyrex tubing`
- `data/press_editorial/br-b09.yaml`: `70 m. p. h.`

Roughly 80 such occurrences found via a simple pattern scan (`[a-z]\. [a-z]\.`); this likely undercounts abbreviations with only one internal period or capitalized forms (e.g. `U. S.`).

### Suggested fix

These all look like the same root cause: whatever detokenizer produced the `text` layer treats `` ` ` ``, `''`, and the `.`/`:` inside certain Brown single-token abbreviations/numbers as if they were always separate words needing a surrounding space, rather than consulting Brown's actual token boundaries (or a small punctuation-attachment rule set) when rejoining tokens into text.

Start by locating the detokenizer that reconstructs the text layer, then compare its output with data/press_reportage/br-a01.yaml and the raw Brown files brown_nolines.txt and browntag_nolines.txt. Verify quote marks, digit-colon-digit sequences, and abbreviation periods against Brown token boundaries across data/. Done means reconstructed text no longer inserts gaps where the raw corpus has contiguous text.

Written by the indexing model from the issue text.