Remove the `rtk test` subcommand: let native `test` fall through to passthrough (0/526 vs 212/526)

Summary

rtk test is both RTK's generic test-runner wrapper and a shadow of the POSIX test utility. That collision is what #2515 fixed: rtk test -d graphify-out was joined into a shell string and run as sh -c "-d graphify-out", producing sh: 0: Illegal option -d.

#2515 resolves it with a predicate that decides, per invocation, whether the arguments look like a native test expression. It is a real improvement, but it is a heuristic, and the heuristic is the whole problem — it can only ever be approximately right.

Removing the subcommand is exact instead. RTK already passes unknown subcommands through to the real binary (rtk basename /a/b.txt → b.txt), so with Commands::Test gone, rtk test <expr> reaches /usr/bin/test and behaves identically to it.

Measurement

526 generated test expressions, exit code diffed against /usr/bin/test:

The 212 that survive #2515 are the cases its predicate cannot reach: every operand-first form (rtk test 1 -eq 1, rtk test abc = abc → sh: 1: 1: not found, exit 127) and rtk test -h FILE, which clap consumes as its help flag and answers 0 for every path. Removal fixes both, along with the !/( ambiguity that needed a follow-up commit on #2515.

What it deletes

is_native_test_expression and its unit tests, run_test,
extract_test_summary, and MAX_RUNNER_FAILURES / MAX_RUNNER_LINES in
src/cmds/rust/runner.rs all become dead once the arm is gone.

What needs a decision, not just a deletion

It drops off the integrity whitelist. Commands::Test is listed in is_operational_command (src/main.rs:2543), which gates
hooks::integrity::runtime_check() at src/main.rs:1486. The passthrough catch-all is not on that list, and the whitelist's own comment says a command missing from it "fails open (no check)". So removal silently ends runtime hook verification for this path. That should be a deliberate call rather than a side effect.

What is lost

rtk test <cmd> as a generic wrapper, documented at README.md:201 and docs/usage/FEATURES.md:563-571. Every ecosystem named in those docs already has a dedicated filter (rtk cargo test, rtk pytest, rtk go test, rtk vitest, rtk rspec, …) and rtk err <cmd> remains as the generic fallback, so the real loss is ecosystems with neither. No rewrite rule emits bare rtk test — the ("test", 90.0) entries in discover/rules.rs are subcommand savings for rtk cargo test and friends — so the hook generates no traffic that this breaks.

Notes

Filed from the review of #2515, which should merge on its own merits first; this supersedes rather than blocks it.

The agent guidance that generates rtk test invocations — the "prefix every shell command" rule and the wording that makes it false for builtins — moved to #3969, which covers the same three instruction files for every builtin, not just test. Removal here and the wording fix there are independent: read and help have no binary to pass through to, so only the wording reaches them.

Start with the Commands::Test definition and its handling in the CLI, then inspect the dead-code candidates in src/cmds/rust/runner.rs and the integrity whitelist at src/main.rs:2543. Review README.md:201 and docs/usage/FEATURES.md:563-571 for the documented generic wrapper. Done means the native test expressions match /usr/bin/test across the reported corpus, the integrity behavior is deliberately addressed, and stale code and documentation are handled.