fix: harden against confirmed review findings

[conorbronsdon]

From an automated multi-agent review (personal-context#62); implemented + verified by Claude Code. (PR auto-recovered — the agent completed and pushed this branch but a transient API error interrupted it before it could open the PR.)

---

fix: harden unquote UTF-8 decode and urljoin CPython parity

Fixes four confirmed review findings, each reproduced against CPython
urllib.parse ground truth before and after the change, and covered by
new fixtures/regression cases (fixtures.txt regenerated from CPython).

1. unquote: reject overlong UTF-8 (security). unquote("%E0%80%AF")
   previously passed the raw invalid bytes e0 80 af through; CPython
   emits three U+FFFD. The UTF-8 decoder now enforces the lead-byte-
   specific second-byte range (RFC 3629 / Unicode Table 3-7), so
   overlongs, surrogates (ED A0..BF), and out-of-range 4-byte leads are
   rejected instead of decoding to raw/invalid bytes.

2. unquote: Unicode maximal-subpart replacement. A truncated multibyte
   escape (e.g. %E6%97, %F0%9F%98) emitted one U+FFFD per byte;
   CPython emits exactly one U+FFFD and resumes past the maximal valid
   subpart. The decoder now advances by the consumed subpart length.

3. urljoin: digit-led pseudo-scheme. urljoin("http://a/b/", "10:30.html") returned "10:30.html" (treating "10" as a scheme);
   CPython returns "http://a/b/10:30.html". urljoin is now a faithful
   port of CPython's algorithm and defers scheme detection to urlparse,
   which already requires an alpha-led scheme.

4. urljoin: CPython backward-compat + absolute-ref dot segments.
   urljoin(base, "http:g") returned "http:g"; CPython resolves a
   same-scheme reference relatively -> "http://a/b/c/g". An absolute
   reference with a differing scheme is now returned verbatim (no dot-
   segment removal), matching CPython (e.g. "http://x/../y" stays as-is).
   The RFC 3986 5.4 table row for http:g (dropped to keep 41/41) is
   restored with CPython's value, so the table is now 42/42.

Also fixes the pixi Mojo pin (>=1.0.0b3 sorts below dev nightlies, so
pixi install fails to solve) to >=1.0.0b3.dev0,<2 as a build
prerequisite. README conformance/fixture counts updated where a fix made
them false (41/41 -> 42/42, 161 -> 173 fixtures) and the urljoin scheme
note corrected to describe the new CPython-matching behavior.

Co-Authored-By: Claude noreply@anthropic.com

[conorbronsdon]

🤖 Independent Claude review: Ready to mark for review — the four target fixes are all correct and well-tested. One low-severity parity gap noted below (non-blocking).

Verified (differential-tested against CPython 3.9):

• UTF-8 hardening decoder is faithful. Traced all 8 new fixtures: per-position second-byte ranges reject overlongs/surrogates/out-of-range 4-byte leads, and maximal-subpart resumption emits exactly one U+FFFD and resumes at the first non-subpart byte. Confirmed the old decoder decoded overlong %E0%80%AF to a literal / (the real security bug) and emitted 2× U+FFFD for truncated %E6%97 — the new regression fixtures genuinely FAIL on pre-fix code, not vacuous.
• http:g → http://a/b/c/g, 10:30.html → http://a/b/10:30.html, http://x/../y returned verbatim (no dot-seg removal) all match CPython. Alpha-led scheme check correctly rejects the digit pseudo-scheme.
• _uses_relative/_uses_netloc/_uses_params lists match CPython exactly.
• pixi pin: confirmed 1.0.0b3.dev5 >= 1.0.0b3 is False, so the old pin excluded dev nightlies — the .dev0 fix is correct.

Non-blocking findings:

1. src/url/parse.mojo urljoin parses the reference as urlparse(url) (empty default scheme) rather than CPython's urlparse(url, bscheme). Since _uses_params("") is always True, ;params is always split off the reference. For a scheme in uses_relative but NOT uses_params (file, ws, wss, gopher, nntp, wais, svn, svn+ssh) with a dot-segment-plus-param reference, this diverges: urljoin("ws://a/b/", ".;x") → this lib ws://a/b/;x vs CPython ws://a/b/.;x; urljoin("file://h/a/b/", "..;p") → file://h/a/;p vs CPython file://h/a/b/..;p. Root cause is a regression from this rewrite (the old struct splitter did no params split). Realistically unreachable (dot-segments carrying ;params are not real URLs), so it does not block, but it contradicts the unqualified "byte-for-byte" docstring claim. Fix: gate the reference's params split on the resolved scheme (mirror urlparse(url, bscheme)), or skip params splitting on the ref entirely and let urlunparse recompose.
2. test/data/fixtures.txt has a duplicate http:g row (lines 140 and 172); gen_fixtures adds it in both RFC3986_ABNORMAL and urljoin_extra. Harmless, just inflates the 173 count by one.
3. Doc nit: README/docstring still call it "the canonical Section 5.4 conformance table (42/42)" while the http:g row deliberately uses CPython's value, not the RFC's strict http:g. The "deliberately does NOT do" section discloses this, but "canonical … conformance table" slightly overclaims for that row.

[conorbronsdon]

Code review (opus, static — verify CI): SHIP-WITH-NITS

All four fixes correct. _utf8_lossy rewrite implements RFC 3629 / Table 3-7 lead-byte ranges properly (hand-traced all 8 fixtures vs bytes.decode("utf-8","replace")); urljoin port to CPython's algorithm verified against http:g→http://a/b/c/g, 10:30.html, http://x/../y; digit-led scheme fix correctly delegates to urlparse.

Nits: parse.mojo:641 segments.append(seg) omits the .copy() used at :651/:662 (style); parse.mojo:63 _uses_netloc lists rtsps which isn't in CPython (pre-existing, only reformatted here). Confirm CI green before merge (no toolchain in review).