[StevenLi-phoenix]

Created by Claude Opus-5-5: the code, tests and this description were written entirely by Claude Opus-5-5.

Fixes #1398.

What was broken

Tectonic hung at "Running BibTeX" for any entry whose alphabetic label needs
a braced special character such as {\"o}. That covers \bibliographystyle{alpha}
and biblatex style=alphabetic with backend=bibtex, whenever the special
character falls within the first three letters of the surname. With alpha.bst
one entry is enough to trigger it, e.g. author = {Kurt G{\"o}del}. biblatex's
label code calls text.prefix$ on a name only in some configurations, which is
probably why a single entry did not reproduce it in the report. Tectonic printed
nothing and never finished.

Root cause

interp_text_prefix (text.prefix$) has an inner loop that skips to the closing
brace of a special character. That loop never advanced idx, so it spun forever.
#1185 fixed the same loop in interp_text_len (text.length$). Both loops also
counted every byte of the special character as a text character. bibtex.web
counts an entire special character as one text character (incr(num_text_chars)
once, after the loop).

Changes

Three commits. The second and third are independent of the hang and can be
split into separate PRs if you prefer.

1. bibtex: fix text.prefix$ hang on special characters. Advances idx inside
   the loop and counts the special character once. G{\"o}del now gives the
   label G{\"o}d31, as bibtex.web specifies.
2. bibtex: count a special character once in text.length$. This is the same
   counting fix in interp_text_len. text.length$ on H{\"a}hnle now returns 6
   (it returned 8 before). In alpha.bst the miscount changed labels for a single
   author whose last name starts with a special character: {\.I}lhan was
   labelled {\.I} and is now {\.I}lh. This changes the existing
   control_seq_alpha expectations. The entry order and the warning order in
   the .blg also change, because alpha.bst sorts by label.
3. bibtex: fix text.prefix$ returning an empty string for fields. I found
   this while testing. PoolCursor::append_substr copied the bytes but did not
   advance the cursor, so text.prefix$ on a non-scratch string (an entry field
   or a literal) returned only the closing braces it adds. In alpha.bst an
   entry with a key field and no author got the label 99 instead of
   {\"O}st99. It has been wrong since the StringPool cursor rewrite
   (2742e7a), and 0.17.0 also returns an empty string for "Smith" #3 text.prefix$
   on a field. The fix advances the cursor, as append_str already does.

Tests

• New tests/bibtex.rs case test_text_prefix_special (fixture
  tests/bibtex/cites/text_prefix_special.*, using the existing alpha.bst) has
  three entries: G{\"o}del, G{\"{a}}rtner (nested braces inside the special
  character) and an entry labelled from key = "{\"O}sterreich".
  • On unmodified master the test hangs: timeout 70 cargo test --features external-harfbuzz --test bibtex test_text_prefix_special
    exits 124 after printing "has been running for over 60 seconds".
  • With only the idx increment (per-byte counting kept), the labels come out
    as G{\"o}31 and G{\"{a}}50, and the test fails.
  • Before commit 3 the key entry is labelled 99, and the test fails.
• control_seq_alpha.{bbl,blg} now expect {\.I}lh. The test fails before
  commit 2 and passes after it.
• New unit test pool::tests::test_write_substr. It fails before commit 3
  (the string is ! instead of Hello!) and passes after it.
• No TeX Live bibtex was available here, so I wrote the expected .bbl files
  from bibtex.web and cross-checked them with pybtex 0.26.1, which gives the
  same labels: G{\"o}d31, G{\"{a}}r50, {\"O}st99 and {\.I}lh. A run against
  real BibTeX would still be welcome.

Commands run (macOS arm64, Rust 1.97.1):

cargo test --workspace --features external-harfbuzz       # all pass
cargo test --features external-harfbuzz --test bibtex     # 25 passed
cargo test -p tectonic_engine_bibtex                      # 10 passed
cargo fmt --all -- --check                                # clean
cargo clippy -p tectonic_engine_bibtex --all-targets --all-features -- --deny warnings   # clean
cargo clippy --all --all-targets --all-features -- --deny warnings                     # clean

End to end, with the reproducers from the issue:

• \bibliographystyle{alpha} with a single Kurt G{\"o}del entry: Tectonic
  0.17.0 is still in BibTeX when killed after 120 s. The patched build finishes,
  and doc.bbl has \bibitem[G{\"o}d31]{g31}.
• biblatex backend=bibtex,style=alphabetic with H{\"a}hnle, M{\"u}ller,
  Lov{\'a}sz and G\"o\"os: Tectonic 0.17.0 is killed after 90 s. The patched
  build finishes with labelalpha values H{\"a}h10, M{\"u}l01, Lov79 and
  G\"22. The last one is unchanged by this PR. The accent in G\"o\"os is not
  braced, so text.prefix$ counts \ and " as ordinary characters, as
  bibtex.web does.
• A small .bst that writes text.length$ and #3 text.prefix$ of each
  author field gives 5 Smi, 6 H{\"a}h, 7 G{\"{a}}r and 5 {\.I}lh for
  Smith, H{\"a}hnle, G{\"{a}}rtner and {\.I}lhan. 0.17.0 gives lengths
  8, 11 and 7 for the last three, and an empty prefix for Smith.

Not addressed

The issue also asks for a diagnostic or a timeout when BibTeX does not
terminate. This PR fixes the cause of the hang only.

PoolCursor::append writes strings[end] without growing the pool. If a string
ends exactly at the end of the allocated pool, the next append panics with
"index out of bounds" (shown with a unit test that adds a 65536-byte string and
then appends one byte). add.period$ and the closing braces of text.prefix$
go through this path. The code before the StringPool rewrite reserved room
first. I left it out of this PR because it is a separate bug.