feat: cargo_diag backend (BuildDigest, rustc/cargo build diagnostics)

[onlyxItachi]

Summary

Adds cargo_diag, a BuildDigest backend for SAVED cargo build --message-format=json
output (NDJSON): per-crate error/warning digests answering "where do this Rust
build's diagnostics concentrate" without spending agent context on rendered
compiler output. Formats cargo-diag + alias cargo-json; suffixes
.cargo-diag.jsonl + .jsonl; new domain constant
DOMAIN_BUILD_DIAG = "build_diag" in core/metrics.py.

Design decisions

• Units are CRATES, one per distinct package_id seen across
  compiler-artifact AND compiler-message records, in first-appearance
  order — a crate that FAILED to compile emits messages but never an artifact
  and must still be a unit (the error fixture pins exactly this:
  the bin crate has artifacts == 0 but is fully addressable).
• package_id parsing verified against the real capture before writing the
  parser (it changed across cargo versions). cargo 1.94 emits the Package ID
  Spec URL: path+file:///ws/mathlib#0.1.0 — the fragment is a bare version
  when the crate is named like its directory, name@version otherwise
  (registry PURLs). The parser handles the modern URL shape (bare-version and
  name-at-version fragments, git ?rev= queries, .git suffixes), the legacy
  name version (source) string, and falls back to the raw id (addressable,
  never crashing, never inventing a name) on unknown shapes. Unit names are
  short name@version (e.g. mathlib@0.1.0).
• The honesty story of this backend: duration_us is None for EVERY
  unit. cargo's JSON stream carries no timing whatsoever, so this is a whole
  backend whose duration is honestly absent — summarize_report ranks by its
  existing file-order fallback (tested: sorted_by == "file_order", and no
  total_duration_us/coverage keys are fabricated), and get_metrics
  returns not_available_in_this_export for duration_us end-to-end. The
  usage prompt redirects build TIMING questions to ninja-log /
  clang-time-trace.
• Counts are the opposite honesty case — genuine 0.0. Like ptxas's
  Used ... line, the message stream is a complete enumeration of the
  diagnostics this build emitted, so a crate that appears in the stream with
  no messages has MEASURED zeros. Verified live before shipping the claim:
  a fully cached rebuild (fresh=true artifacts) replays cached warnings as
  compiler-message records, so the enumeration stays complete even when
  nothing recompiles (this became a probe note).
• Numeric digest, text stays out (translator, not judge): core set is
  errors / warnings / notes_helps (levels partitioned as: starts-with
  error including ICEs; warning; everything else — note/help/rustc's
  span-less failure-note). expand() carries the WHERE story numerically:
  per-level-per-file keys built from the primary span's workspace-relative
  file_name (warning@mathlib/src/lib.rs → 2.0, span-less diags in an
  explicit (no-span) bucket), per-code keys (code:E0308 → 1.0,
  code:dead_code → 1.0), plus artifacts/fresh_artifacts and the raw
  package_id. Reading diagnostic TEXT is the agent's editor/compiler job —
  the usage prompt says exactly that.
• build-finished.success: the digest schema has no report-level slot and
  core stays unbent, so the report-level outcome rides expand() as a
  build_finished_success bool on every unit — absent entirely (never
  fabricated) if the capture was truncated before the build-finished record.
• Loud on non-cargo input: a line that is not JSON (a Chrome trace is ONE
  multi-line document — its first line alone fails), not an object, or an
  object without cargo's reason key raises a named ValueError, with a
  redirect hint for the trap case (a perf-stat export is ALSO JSON-lines —
  format perf-stat-json).
• Shared files: ONE import line in server/app.py::_register_backends(), ONE
  domain constant in core/metrics.py, one PROFILER_TOOLS entry
  ("cargo_diag": "cargo" — same executable as criterion, different artifact).
  Rebased over the criterion (Add criterion backend (Rust benchmark digest, directory-ref reports) #14) and prev-green (Prove previous-green CI comparison on real two-run artifacts (CIDigest) #15) merges; no
  format/suffix collisions (criterion claims criterion/criterion-json
  with report_is_directory=True; cargo_diag is a plain file backend).

Fixture provenance

Real cargo 1.94.0 captures, not hand-fabricated. The 2-crate workspace is
committed for provenance at tests/fixtures/cargo_diag_ws/ (mathlib lib
crate with deliberate unused_variables + dead_code warnings; mathapp
bin crate depending on it — clean main.rs plus the main.rs.error-variant
that produced the error fixture):

• cargo_build_warnings_sample.cargo-diag.jsonl — fresh build of the clean
  workspace: 2 real warnings on mathlib, mathapp compiles clean,
  build-finished success=true.
• cargo_build_error_sample.cargo-diag.jsonl — post-cargo clean build after
  introducing a real E0308 (calling add(2, "three")): mathlib's warnings,
  mathapp's error + rustc's failure-note, success=false, and no
  compiler-artifact for the failed crate.

Host-identity check: cargo unavoidably embeds the absolute workspace path in
package_id/manifest_path/artifact filenames, so the capture was run from
a neutral /tmp/perfdigest-ws root — the committed NDJSON contains no
hostname, username, or home path (grep-verified); nothing was hand-edited.
Span file_names are workspace-relative (cargo run from the workspace root —
verified), which is what the per-file concentration keys are built from.

Test evidence

New tests/test_cargo_diag.py, 14 tests: registry dispatch on both formats;
crate units named from real package_ids; duration-None honesty end-to-end
(not_available_in_this_export through get_metrics, all-None in
list_kernels); genuine-0.0 for the clean crate; real error/warning counts
from both fixtures; summarize file-order fallback with no fabricated coverage
keys; expand concentration keys (per-level@file, (no-span) bucket, per-code)

• substring section filter; report-level build_finished_success via expand
  (true and false fixtures); failed-crate-has-no-artifact; list/expand index
  agreement; synthetic (clearly marked) package_id shape pinning for
  registry/git/legacy variants the path-dependency fixture cannot produce; loud
  ValueErrors on the three non-NDJSON traps; vocabulary hint for foreign terms.

Full suite: uv sync --extra dev && uv run --extra dev pytest -q →
182 passed, 16 skipped (168 post-#14/#15 baseline + 14 new;
hardware-gated skip count unchanged).

Test plan

• uv run pytest -q — 182 passed, 16 skipped (after rebase onto a82e7fd)
• Cached-rebuild warning-replay claim verified against the live cargo
  toolchain before it shipped as a probe note
• Manual smoke: platform_capabilities() (digest + capture rows) and
  suggest_profile_command('cargo_diag', ...) on this host

🤖 Generated with Claude Code