fix(hooks): stop unit tests reading the developer's config.toml

[David-Martel]

Problem

cargo test is red for contributors who use rtk and green in CI, which is why this hasn't
surfaced upstream. Two places in the hook decision path do file I/O, so unit tests inherit the
contributor's machine configuration.

1. get_rewritten calls Config::load() — reads $CONFIG_DIR/rtk/config.toml. With a
configured

[hooks]
exclude_commands = ["git", "ls", "grep", ...]

commands the tests expect to be rewritten are excluded, get_rewritten returns None, and the
decision comes back Defer. test_claude_rewrite_git_status — which asserts git status →
rtk git status — fails, along with 22 others spanning every host handler (claude, cursor,
gemini, droid, vibe, copilot), since they all funnel through decide_from_verdict.

2. evaluate calls check_command — reads ~/.claude/settings.json and
$PROJECT_ROOT/.claude/settings.json. The unattestable_passthrough tests pass explicit empty
exclusions but still inherit that ambient verdict, so with a matching deny rule the outcome is
Deny instead of Ask and 2 more tests fail.

Excluding git is a reasonable real-world config — it's what you'd set to keep exact git
output for a coding agent — so this is not an exotic setup.

Measured

Pristine checkout of develop, no source changes, only the config file moved aside:

condition	hooks::hook_cmd
config.toml present with [hooks] exclude_commands	23 of 108 fail
same file moved aside	108 pass, 0 fail

Fix

Split file I/O out of the decision path, mirroring the convention permissions already
established with check_command_with_rules — whose doc comment reads "Internal implementation
allowing tests to inject rules without file I/O." The seam existed; these paths just didn't
use it.

commit	change
stop unit tests reading the developer's config.toml	get_rewritten_with / decide_from_verdict_with take exclusions as parameters; hook_exclusions() becomes the single loader and is #[cfg(test)]-stubbed to empty so no test path reaches the filesystem.
stop rewrite tests reading the host's permission settings	evaluate_with_verdict takes the verdict as a parameter; evaluate loads it and delegates; the test module supplies PermissionVerdict::Default.

Production behavior is unchanged in both cases — the loading wrappers still load. The only
behavioral difference is inside #[cfg(test)].

Added test_get_rewritten_with_honors_exclusions so the exclusion behavior the other tests now
bypass stays pinned rather than silently untested.

Verification

Run with a populated config.toml and settings.json in place — the condition that used to
fail:

• unit test binary: 25 failures → 0 (2571 passed, 8 ignored)
• cargo clippy --release --all-targets -- -D warnings — clean
• cargo fmt --all -- --check — clean

Note for maintainers

Two related issues found while investigating, not addressed here:

1. tests/copilot_selfheal_test.rs (added in fix(hooks): copilot self heal dual hooks (drop camelCase entry) #3449) fails 10 of 12 on a pristine develop
   checkout on a machine with existing Copilot hook config — same class of ambient-state
   dependency, in brand-new code. Happy to extend this PR to cover it if you'd like.
2. core::tracking tests are flaky under full-suite parallelism — they share one real SQLite
   tracking DB. They pass 14/14 in isolation, both parallel and single-threaded.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[KuSh]

Hi @David-Martel,

Thanks for your PR. Could you extend it with copilot fix? The flaky DB tests are already fixes in #3206 with commit 2aef8e8