[JacobStephens2]

Summary

This makes the test suite cheaper to build and run, so an agent session can run the full cargo test in one foreground Bash call.

 Cargo.toml
+  [profile.dev] debug = "line-tables-only"; dependencies debug = false   # #202
+  comment: why aws-lc and the image stack stay (axoasset pins them)      # #203
 tests/support/mod.rs
-  THIRDSHIFT_POLL_MS=10
+  THIRDSHIFT_POLL_MS=100                                                  # #205
 tests/*.rs   loops over --email positions, signals, labels, spellings, key sources
-  one Run per case
+  one Run per behavior; the cases move to unit tests                     # #204, #206
+    src/args.rs (new seam: argument parsing), src/interrupt.rs, src/run.rs (Budgets),
+    src/resend_key.rs, src/spec_run.rs, src/config.rs, scripts sourced by tests/release.rs
 tests/fakes/
-  gh, claude            (Python, ~0.5s to start under load)
+  fakes.rs gh.rs claude.rs json.rs   (one std-only Rust executable, same contract;
+                                      built with rustc on first use, cached by source hash)  # #207

The review found a few things, which this PR also fixes:

• The fakes now share one git runner and one reader for their FAKE_* paths.
• gh fake … is now handled by fake_command.
• A failed Run whose Run notification also fails to send is tested end to end again. Collapsing its loop had dropped that case.

Evidence

Machine: 4 cores. None of these runs was on an idle machine.

Measurement	Target	Baseline (55bcf6e)	After (#207)
Warm full cargo test, wall	≤ 180s	268s	107s / 129s / 113s
Warm full cargo test, user + sys	≤ 350s	479 + 222 = 701s	57 + 131 = 188s
Cold cargo test --no-run, wall	≤ 300s	127s	108s

I ran the warm suite again at this PR's head, after the review fixes: 127s wall, with 75s user + 156s sys = 231s CPU. All 25 test binaries pass.

The #207 runs also passed with a python3 stub first on PATH that fails if called, so nothing calls Python any more. The per-ticket measurements are in the comments on #174.

Merge Danger

Door: two-way

The only production-side changes are:

• the dev profile
• environment_key split out of from_environment in src/resend_key.rs, with the same behavior
• scripts/release*.sh now run main only when executed, not when sourced, and is_plain_version is extracted

Blast Radius: tests

The shipped binary is unchanged: release builds keep their profile, and the TLS stack and axoupdater are untouched. The integration tests now need rustc on PATH (or $RUSTC) rather than python3. Dependency frames in dev backtraces lose file and line.

Follow-up

• aws-lc and the image stack stay. axoasset's remote feature pins reqwest's default-tls (aws-lc) and image. Removing them needs either an upstream axoasset change or a replacement updater. Both are out of scope here. The details are in the Make the test suite faster #174 comment on Stop building aws-lc-sys, and axoupdater's image stack if a flag allows #203 and in Cargo.toml.

Unaddressed findings

Standards

• Primitive Obsession: the gh fake's state is untyped Json reached by string keys. Kept on purpose: it keeps the Python fake's JSON contract, which the harness writes and reads.
• Global static LOCK in tests/fakes/gh.rs, released by run_hook. Kept: it mirrors how the Python fake let go of the lock while a hook ran, and the doc comment explains why.
• environment_key exists only as a test seam. Kept: it is small and well named.

Spec

• "Configure reqwest/rustls so aws-lc-sys is no longer built at all": not done. It can't be done with a feature flag (see Follow-up). The spec allows skipping a lever once the targets are met.
• The Repair and budget-cycle end-to-end tests still run every round. Each is already a single Run: the cap only runs out after the real rounds, so those rounds are the behavior under test. The counting also has unit tests in src/run.rs.
• email.always with no key, or with a key only in the Credentials, has no end-to-end test. Not re-added: main.rs turns email.always into the same NotificationAsk as --email, and config.rs's unit tests cover that. From there the key lookup is the same path, which is tested end to end.
• SIGINT to the Spec run alone has no end-to-end test. Not re-added: a unit test in src/interrupt.rs shows SIGINT sets the interrupt flag. The Spec run then always sends SIGTERM to the Ticket's Run, and that path is tested end to end.
• New unit-test modules besides argument parsing, plus the sourceable release scripts. Kept: this is lever 2's "the rest can be checked below the binary". Argument parsing was the only seam that had no tests before.
• The Rust fakes, about 1,860 lines including a JSON module, are larger than the Python fakes. Kept: they cut the suite's CPU time by about two thirds. Each fake stays one file with its contract documented at the top.
• The timings were not measured on an idle machine. Every target is met even under load, so an idle machine would only do better.

Closes #174

Tickets

• Record the test suite baseline and make debug builds cheaper #202 landed with Use line-table debuginfo for dev builds #209
• Stop building aws-lc-sys, and axoupdater's image stack if a flag allows #203 landed with Say why aws-lc and axoupdater's image stack stay #213
• The test harness polls CI about every 100ms, not every 10ms #204 landed with Poll CI every 100ms in the test harness #216
• Unit tests for argument parsing; one Run per --email end-to-end test #205 landed with Unit tests for argument parsing; one Run per --email end-to-end test #214
• Collapse the looping end-to-end scenarios to one Run per behavior #206 landed with Collapse the looping end-to-end scenarios to one Run per behavior #217
• Measure the test suite against #174's targets; replace the Python fakes only if still short #207 landed with Replace the Python fake gh and claude with a Rust helper #218

🤖 Generated with Claude Code