Tracking: hosted rewriters behind one HostedRewriter trait with a per-dependency outcome

[mikolalysenko]

[agent] Filed by the scheduled architecture audit routine (ecosystems and formats). Register: E31.

Kind: tracking. Source: review §2.1 and Part 3.7 recommendation 2; register E31.

Problem

Verified on main 05ecc6e. Hosted mode has no rewriter abstraction. Each rewriter is a free function fn(&BTreeMap<String,String>, &[DepOverride], &mut RewriteResult), and each one reports which patches it pinned by writing to its own field of a shared struct.

• RewriteResult now holds 27 per-ecosystem uuid sets (20 at the review): for example confirmed_cargo_uuids, refused_pdm_uuids, yarn_berry_uuids, vlt_foreign_uuids, gradle_uuids, confirmed_sbt_uuids and refused_sbt_uuids. There are four meanings, spelled ad hoc per ecosystem:
  • owned: this rewriter decides the uuid (yarn_berry_, python_lock_, hatch_, gradle_);
  • confirmed: the pin landed;
  • refused: the rewriter declined it;
  • foreign: vlt_foreign_, plus bundled_skipped_.
• merge_group_delta destructures and extends every field by hand.
• engine::confirm is 225 lines. It is an ordered chain of 25 ProbeStep outcomes keyed on purl prefixes, pdm_drives(files), vlt_drives, binary_bun, sbt::maven_confirmation and those sets. Precedence is encoded only by the order of the ifs (vlt before binary bun, pdm before hatch, sbt before Gradle).
• The CLI reads four of the sets directly (scan/hosted.rs#L1215, #L1245-L1246, #L1339). vex/discover/gradle.rs and upstream/gradle.rs also reach into the Gradle sets.
• rewriter_groups is a hand-wired Vec<Box<dyn Fn>> of 11 groups. Each rewriter filters its own deps with string compares (ecosystem == "…": 12 in redirect/mod.rs, 15 in hosted/engine.rs).

Adding an ecosystem to hosted mode still means editing these parallel tables:

• formats::registry::REGISTRY;
• upstream::format_of and restore_pass;
• engine::file_ecosystem;
• memory::select::classify;
• roots::UNSUPPORTED_MARKERS;
• rewriter_groups;
• the RewriteResult fields, merge_group_delta and confirm().

The sbt support added in #690 shows the cost: two new sets, new fields in merge_group_delta, and a three-way sbt/Gradle/pom precedence block inside confirm().

Symptoms

• Precedence bugs between rewriters that share an ecosystem are decided by the order of the ifs in confirm(). Examples are the pdm/hatch ordering comment and vlt beside bun.lockb, both fixed after the fact.
• E73 (Decide whether a hosted patch is pinned through lockfile discovery alone #1058) shows that "is this pinned" is answered in several places. Decide whether a hosted patch is pinned through lockfile discovery alone #1058 adds a discovery check after confirm() but keeps confirm() itself.

Impact: every hosted ecosystem change touches the same three hot spots, which serializes concurrent hosted PRs on mod.rs and engine.rs.

Target design

pub enum DepStatus { Confirmed, Refused, Foreign, Skipped }
pub struct Outcome { files, binary_files, edits, warnings, per_dep: BTreeMap<String /*uuid*/, (Rewriter, DepStatus)> }
pub trait HostedRewriter {
    fn id(&self) -> Rewriter;
    fn ecosystem(&self) -> Ecosystem;
    fn drives(&self, files: &FileSet) -> bool;      // pdm/vlt/uv/sbt "driver" rules
    fn rewrite(&self, files: &FileSet, deps: &[&DepOverride]) -> Outcome;
}

• confirm() becomes: if the driving rewriter for the candidate's ecosystem reported a status, use it, otherwise use the needle probe.
• The CLI reads Outcome accessors, never per-ecosystem fields.

Checklist (each item one PR)

• 1. Replace RewriteResult's 27 per-ecosystem uuid sets with one report map #1076: replace the 27 uuid sets with one reports map keyed by (Rewriter, Report), behind accessors. Mechanical, with no behavior change; merge_group_delta becomes one loop.
• 2. Move the driver rules (pdm_drives, vlt_drives, binary_bun, sbt::maven_confirmation) into one driving_rewriter(ecosystem, files) table, and reduce confirm() to "driver's report, else needles". Behavior-preserving; guarded by the existing confirm tests and group_equivalence_tests.
• 3. Introduce trait HostedRewriter and port the rewriters group by group (JS, Python, then one per remaining ecosystem). rewriter_groups becomes a registry. This lands after the Move patch/redirect/mod.rs's inline test modules into sibling test files #1011 / Tracking: split patch/redirect/mod.rs into per-ecosystem modules and sibling test files #1010 file split, so each port touches one file.
• 4. Derive engine::file_ecosystem, memory::select::classify and upstream::format_of from the registry, deleting the hand tables.
• 5. Delete the per-dep ecosystem == "…" filters: the registry hands each rewriter its own deps.

Dependencies

• Step 1 can start now. It conflicts textually with Decide whether a hosted patch is pinned through lockfile discovery alone #1058 (it edits confirm()), so land Decide whether a hosted patch is pinned through lockfile discovery alone #1058 first or rebase on it.
• Steps 3–5 come after Move patch/redirect/mod.rs's inline test modules into sibling test files #1011 and the per-ecosystem moves in Tracking: split patch/redirect/mod.rs into per-ecosystem modules and sibling test files #1010.
• This blocks E32 (one hosted pipeline) and simplifies E73.

[mikolalysenko]

[agent] Triage: priority:p3 (cross-cutting refactor tracking issue). Not a duplicate; no open PR covers it. #1076 is a slice of it.

---

Generated by Claude Code