fix(profile): 58 test targets CI never ran, 110 failing, and a syscall the profiler could not name

[noahgift]

Test triage of a binary surface area, per the standing instruction: bubble the tests up, fix what can be fixed, delete what cannot.

aprender-profile appears in ci.yml zero times and carries 58 integration test targets. They all compile. Running them:

	targets	passed	failed	ignored
before	61	1966	110	48
after	61	2079	0	48

Two root causes

109 failures, one cause

`CARGO_BIN_EXE_renacer` is unset

[lib] name = "renacer" but the binary target is aprender-profile, so that env var never existed. 126 cargo_bin("renacer") call sites across 12 files, dead since the APR-MONO rename. Identical defect to the cgp one in #2496, different crate. Binary name taken from cargo metadata, not assumed.

1 failure, hiding a real defect in the profiler

test_realtime_anomaly_detects_slow_syscall compiles a program that sleeps 50 ms and asserts an anomaly is reported. It wasn't. Two findings:

a. clock_nanosleep (x86_64 syscall 230) had no name. The table jumped (228, clock_gettime) → (231, exit_group), so the tracer printed syscall_230 and the 50 ms sleep was invisible as a named syscall. Added 229 clock_getres and 230 clock_nanosleep.

b. The test could never have passed for the right reason. The detector builds a per-syscall baseline and flags >3σ from it — and clock_nanosleep occurs exactly once, so there is no baseline to deviate from. The green runs were unrelated microsecond jitter on write (8–19 μs against a 10 μs baseline) happening to exceed 3σ. Its own comment already read "increased for test stability" — which is what tuning a test that asserts the wrong thing looks like.

Rewritten so the anomaly has a baseline: 200 one-byte writes, then one 4 MiB write. Same syscall, so the detector can see it. Stable across three consecutive runs. It now also asserts the fixture compiled — a gcc failure would otherwise make the test assert about nothing.

The falsifier, because nothing guarded this

Removing the two syscall entries again left all 107 syscall tests green. Every existing test asserts a property — determinism, non-empty — and a property test over 0..500 cannot catch a missing entry, since "unknown" is a valid answer across most of that range. Only naming numbers can.

mod table_completeness names nine ABI-fixed syscalls, asserts the 228–231 clock range has no holes, and carries a non-vacuity arm: 9_999 must still be "unknown", or a permissive lookup would satisfy the rest.

Mutation: restore the gap → the two name assertions go RED, the non-vacuity arm stays green.

Gated

cargo test -p aprender-profile --lib wired into guard-runner-labels (in gate.needs) so the crate cannot go dark again. --lib only: the integration targets shell out to the built binary and take minutes.

Two measurements I got wrong and corrected

• I first read "21 targets ran, 58 files exist" as 37 files never running. Wrong — cargo stops at the first failing target. cargo metadata confirms all 58 exist; --no-fail-fast gave the real 61/110.
• I tested a stale binary once and had to resolve the path from cargo --message-format=json — the worktree target/debug/ is not where builds land here.