Decode a base64 run only where the scanner decodes it

[azrollin]

Closes the inherited pair of GLS-PI-007 rows named in #35 (one padded, one not) where the scanner blocks and the Worker allows. The same pair differs on main before #35. The rule set is untouched; this aligns which base64-looking runs the Worker's preprocessor accepts for decoding (padding acceptance) with the scanner. Two decoding differences remain and are listed under Limitations: invalid-byte skipping and the printable-character table.

Cause

The scanner reads a run of 20 or more base64 characters with Python's b64decode, which refuses a run whose padding cannot complete the last group and then leaves the run as written. The Worker used atob(), which is more forgiving: it accepts a run with no padding and refuses surplus padding. In the failing rows the text is written backwards and contains a 23 character run of letters and a slash. The scanner leaves that run alone and still reads the reversed phrase. The Worker decoded it to noise, which destroyed the phrase.

The same difference ran the other way: a run with surplus padding that the scanner decodes was left as written here.

Change

src/preprocessor.js, decodeBase64Segments only:

• A run is decoded only where b64decode accepts it: one leftover character in the last group never decodes, two need both =, three need at least one, and padding beyond that is ignored.
• The result must be longer than four code points (it was four UTF-16 units).
• Invalid bytes are dropped, a leading BOM is kept (so the result is not printable), and a valid U+FFFD is kept.
• The printable test follows str.isprintable() through Unicode categories (other, separator) instead of a short list of control characters.

The matching entry in LIMITATIONS and the README line are updated to what remains.

Test, red on main and green on this head

tests/base64_segment_padding.test.mjs with tests/fixtures/scan_corpus_base64_padding.json. Expected values come from the scanner's own function and scan (0.6.8, commit 0a3b4686). It covers both padded and unpadded forms of the failing rows (including 2,100 characters of padding that the normalizer removes, in two kinds), runs with each amount of padding, and the result checks above.

node --test tests/base64_segment_padding.test.mjs
main (8f5b870):  pass 2, fail 6
this head:       pass 8, fail 0

Differential against the scanner

decode_base64_segments on 546 generated texts (14 payload shapes x padding states, 300 random byte strings, raw alphabet runs of 20 to 27 characters), output text and recorded segments compared: 105 differences on main, 0 on this head.

Decision and sorted finding ids, one scan per row, stock rules:

set	rows	differing on main	differing now
length rows from the #35 review	1,712	2	0
supplementary-padding rows from the #35 review	908	4 (2 decision, 2 finding-id only)	0
scanner-recorded corpus (scan_corpus_068.json)	1,307	1	1 (unchanged)
generated cue x payload set	12,614	16	16 (unchanged)

What remains different

• The printable test reads this runtime's Unicode tables where Python reads its own, so a character assigned in one and not the other can decide a segment differently.
• A valid U+FFFD is dropped along with invalid bytes when one segment holds both.
• Offsets and windows count UTF-16 units where the scanner counts code points (the list in Port the 0.6.8 engine reads: folded regex view, negation walk, record hook #35). Tracked separately.

This PR makes no claim that the hosted demo matches the scanner beyond the rows and checks listed here.

Gates on this head

Private checkout of the scanner at 0a3b4686. All exit 0 on head 48409df: disclosure gate; policy parity (14 cases, 0 mismatches); engine parity (137 cases, 0 verdict, 0 finding-set deltas); channel parity (2,488 cases, 0, 0); wide parity (1,555 pairs, 0, 0, no known entries); counterexample parity (0 deltas, 0 slower than the bound); parity_test.py (882 regexes, 0 misses; 13,024 benign checks, 0 disagreements; seed 4173371, replay with PARITY_SEED); v-flag disclosure; metadata contract. node --test tests/*.test.mjs: 164 passed, 0 failed (156 on main, plus the 8 new). python3 -m pytest on the dotless-i tests: 62 passed. CI: wide-parity and workerd-runtime green.