[JacobStephens2]

Summary

The warm suite still missed #174's targets (249s wall, 650s user + sys), and the Python fakes' startup used about 352s of that CPU. They're now one std-only Rust executable. It acts as gh or claude depending on the name it's called by, and keeps the same contract: the same state and record JSON files, argv handling and output.

 tests/
-├── fakes/gh              # Python
-├── fakes/claude          # Python
+├── fakes.rs              # entry: dispatch on argv[0]; also a Cargo test target (clippy, rustfmt, JSON tests)
+├── fakes/gh.rs           # port of the fake gh, same contract
+├── fakes/claude.rs       # port of the fake claude, same contract
+├── fakes/json.rs         # std-only JSON, laid out as Python's json module does
 └── support/
+    ├── fakes.rs          # builds the fakes with rustc once, cached in target/tmp by source hash, symlinks them into bin/
     └── mod.rs            # Scenario installs them via support::fakes

The fakes aren't a [[bin]], so neither the crate nor a dist release ships them. No scenario changed. tests/release.rs only changed how its fixture installs the fakes, and it now uses mod support.

Evidence

All measurements are on a 4-core machine that wasn't idle. The full table with load is in the Result comment on #174.

• Before: warm cargo test took 249s wall and 444 + 205 = 650s user + sys. Python fake self-CPU was 333s for gh and 20s for claude.
  After: warm cargo test took 107s / 129s / 113s wall and 188s / 202s / 206s user + sys. All 25 test binaries passed. The last two runs had a failing python3 stub first on PATH, and nothing called it.
• Cold cargo test --no-run took 108s (target ≤ 300s). Building the fakes adds about 1.2s once, at the first test that needs them.
• cargo fmt --check and cargo clippy --all-targets -- -D warnings are clean.

Merge Danger

Door: two-way

Test-only. Reverting brings the Python scripts back.

Blast Radius: tests

Running the suite now needs rustc on PATH (or $RUSTC) instead of python3. Stale target/tmp/fakes-* builds pile up as the fakes change, a few MB each.

Unaddressed findings

Standards

• Hidden global state (static LOCK, state_path() from env): skipped. It mirrors how the fake lets go of the lock while a hook script runs, and threading a lock through every command adds noise to a test fake.
• Primitive Obsession (untyped Json state): skipped. Untyped state passes unknown keys and key order through untouched, which the state-file contract relies on.
• Old fakes-<hash> and failed .partial builds never cleaned up: skipped. They're small, cargo clean removes them, and pruning them could race with parallel test processes.

Spec

• tests/fakes.rs is an extra test target (not asked for): kept. It's the only way the fakes get clippy and rustfmt.
• release.rs pulls in all of support: kept. Only fixture setup changed, no assertions.
• A hand-written 437-line JSON module: kept. A fake built with plain rustc can't use serde_json, and the module has its own tests.
• pr view --jq .<object field> prints JSON, not a Python repr: skipped. Nothing selects a non-string field.
• gh fake … rejects extra arguments; gh fake pr on an unknown head says "no pull requests found" instead of a traceback: skipped. Both still fail (exit 1 or 2), and no caller passes extras.
• run list treats a run missing branch/workflow/event as not matching (Python crashed); claude no longer turns \r\n into \n in stdin and the plugin snapshot: skipped. No test or script records such a run or uses CRLF.
• The merge scratch dir is named tmp-merge-<pid>, not tmpXXXX: skipped. It's removed before any error, and nothing lists that directory.

Closes #207

🤖 Generated with Claude Code