Read and splice nuget.config through formats::nuget in hosted, vendored and restore #594

[agent] Filed by the scheduled architecture audit routine (ecosystems and formats). Register: E11.

Kind: refactor. Source: review §3.7 #3, Part 3.4 and Part 5.4; register E11 (the NuGet half of E10).

Problem

nuget.config is read and spliced by six private code paths with four different XML rules. Only one of them is a real tokenizer.

Readers of the <packageSources> keys:

• formats::nuget::parse_config is a bounded tokenizer. It skips comments, CDATA and PIs, and only records <add> directly under configuration/packageSources. Upstream restore and VEX use it.
• Hosted nuget_package_source_keys runs two regexes over the raw text, so a commented-out <add> counts (Hosted NuGet mapping reads commented-out package sources #561).
• Vendored parse_config_source_keys uses find("<packageSources") plus a substring attr_value over blank_comments output.

Splice anchors and removers:

• Hosted insert_nuget_source and nuget_mapping_open_end take the first regex match of <packageSources> / <packageSourceMapping>, including one inside a comment (Hosted NuGet splices the Socket source (and mapping) into a commented-out <packageSources> / <packageSourceMapping> block, so every restore fails NU1100 while scan reports success and its in-run VEX attests not_affected #585). nuget_after_last_clear carries its own comment masker.
• Hosted restore remove_source builds an `<add … key=…/>` regex and uses another private [`attr_value`](https://github.com/SocketDev/socket-patch/blob/203e092bdb91f573df9d0120eae2c2004a50d358/crates/socket-patch-core/src/patch/redirect/upstream/nuget.rs#L39-L48).``
• Vendored build_config_edit anchors on the comment-blanked text, and excise_source_mapping matches an exact 4-space byte pattern.

Config file names are spelled four times: redirect::NUGET_CONFIG_FILE_NAMES,`` vendor::nuget_config::CONFIG_NAMES, `hosted/memory/roots.rs` and `formats/registry.rs`. They haven't drifted yet.

The code paths themselves have drifted. The vendored writer is comment-safe; the hosted writer and reader are not. The reader that upstream restore uses (parse_config) and the one the hosted rewriter uses disagree about the same file.

Symptoms

• Hosted NuGet mapping reads commented-out package sources #561: hosted mapping reads commented-out sources (reader).
• Hosted NuGet splices the Socket source (and mapping) into a commented-out <packageSources> / <packageSourceMapping> block, so every restore fails NU1100 while scan reports success and its in-run VEX attests not_affected #585: hosted splices the Socket source and mapping into a commented-out <packageSources> / <packageSourceMapping> (anchors). Restore then fails NU1100 while scan reports success.
• Related writer-semantics bugs in the same code: Hosted/vendored NuGet mapping isn't exclusive when nuget.config already maps the exact package id to another source, so restore races the Socket feed against nuget.org (NU1403 with a lock, silently unpatched without one) #462 (mapping not exclusive when the config already maps the id elsewhere) and NuGet nuget.config authored by vendored/hosted mode maps '*' only to nuget.org, cutting off sources inherited from parent or user-level configs (NU1101) #354 (* fanned only to nuget.org).

Impact

Every hosted NuGet bug about comments, self-closing sections or attribute spelling has to be fixed up to three times, and the reader each writer trusts is not the one restore and VEX trust. Size: about 250 production lines of regex and substring scanning.

Proposed change

1. Extend formats::nuget with a span view: parse_config_spans(text) -> Option<NugetConfigSpans> records the byte offsets that every writer needs (the <configuration> open-tag end, the <packageSources> open/close or self-closing span, the <packageSourceMapping> ditto, the last <clear/> end in each, and each <add>/<packageSource> element span with its key). It is computed by the same tokenizer as parse_config, so comments and CDATA are never anchors.
2. Hosted add_nuget_source and rewrite_nuget take the keys and anchors from it. Delete nuget_package_source_keys, the NUGET_PACKAGE_SOURCES_REGION_RE / NUGET_ADD_KEY_RE statics, the regexes in insert_nuget_source / nuget_mapping_open_end / add_nuget_source, and nuget_after_last_clear's comment masker.
3. Hosted restore remove_source and vendored excise_source_mapping / parse_config_source_keys use the element spans. Delete both private attr_values and blank_comments (if no other caller remains).
4. One formats::nuget::CONFIG_FILE_NAMES; delete the other three lists.

This can land as two PRs if it's too big: (a) the span view plus hosted (closes #561 and #585), (b) vendored and restore.

Size and scope

formats/nuget/mod.rs, patch/redirect/mod.rs (NuGet section only), patch/redirect/upstream/nuget.rs, vendor/nuget_feed.rs, vendor/nuget_config.rs, hosted/memory/roots.rs. About +200 / −300 production lines. Out of scope: Maven/Gradle XML (rest of E10) and the packages.lock.json walks (#593).

Acceptance criteria

• One NuGet config tokenizer (formats::nuget). No regex or substring scan over nuget.config text remains in redirect/, hosted/ or vendor/.
• Existing hosted, vendored and restore NuGet tests and goldens stay green unchanged (cargo test -p socket-patch-core --lib nuget, the redirect equivalence suites, e2e_nuget*).
• New table test: one set of configs (commented <add>, commented <packageSources> before the real one, commented <packageSourceMapping>, self-closing sections, single-quoted and key = "…" attributes, CRLF) run through the hosted rewriter, the vendored writer and parse_config. The Socket source and mapping land in the live sections, and the keys each writer sees equal parse_config's.
• The Hosted NuGet mapping reads commented-out package sources #561 and Hosted NuGet splices the Socket source (and mapping) into a commented-out <packageSources> / <packageSourceMapping> block, so every restore fails NU1100 while scan reports success and its in-run VEX attests not_affected #585 repros are regression tests.

Dependencies

Supersedes the narrower fix in #561 (either can land first; if #561 lands first, this deletes its remaining regex). Touches rewrite_nuget like #593, so sequence the two. Blocks the Maven/Gradle half of E10.