AGENTS: document cargo-mutants as a periodic package-owner tool (#298)

## Summary

Add a rust-style entry covering `cargo-mutants` as a package-owner
audit, not a CI gate. Lands next to the `cargo-careful` paragraph so the
two behavior-validation tools sit together. Source fragment edited in
`agents-md/sections/06-rust-style.md`; `AGENTS.md` and `CLAUDE.md`
regenerated via `nix run .#agents-md -- --write`.

## Trial results

Ran on `code-tokenizer` (37 tests, all-pure tokenizer, no async, no
FFI):

```sh
nix shell nixpkgs#cargo-mutants -c cargo mutants --package code-tokenizer --in-place --no-shuffle
```

31 mutants tested in 2 min: 19 caught, 3 timeouts (also caught), 6
unviable (compile errors), **3 missed**. The three survivors are real
gaps:

- `packages/code-tokenizer/src/lib.rs:113` `self.position += 1` → `*= 1`
(EOF early-return).
- `packages/code-tokenizer/src/lib.rs:144` `self.position += 1` → `*= 1`
(normal token-boundary return).
- `packages/code-tokenizer/src/lib.rs:140` `current_offset +=
ch.len_utf8()` → `*=` (separator branch).

`packages/code-tokenizer/src/tests/helpers.rs` only collects
`token.text`, so `Token::position`, `offset_from`, and `offset_to` are
never asserted. Tracked separately in #297 so this PR stays focused on
the trial decision.

## Recommendation

Keep cargo-mutants a periodic package-owner tool rather than a CI gate:

- Runtime scales linearly with mutant count; even 31 mutants on a tiny
pure crate ran 2 min, the same workload across the workspace would
dominate CI.
- Equivalent mutants need human judgment: the survivor on line 113 is
reachable but inert for any caller that ignores `Token::position`, which
is a real call-site question, not a binary pass/fail.
- A surviving mutant is a prompt to look, not a regression to block. CI
gating would push the team toward writing assertions just to kill
mutants instead of asking whether the behavior is worth pinning.

## Test plan

- [x] `nix run .#lint`
- [x] `nix run .#agents-md -- --write` produces no further diff
- [x] cargo-mutants trial captured above

Closes #252
Refs #297

Author: wyattgill9

AGENTS.md

@@ -304,6 +304,17 @@ against a debug-assertion standard library and surfaces some unsafe-precondition
 and stdlib-invariant breakage, but it does not model aliasing, uninitialized
 reads, or data races, so it complements Miri rather than replacing it.
 
+When auditing a crate with deterministic, fast tests, run
+[`cargo-mutants`](https://mutants.rs/) with
+`nix shell nixpkgs#cargo-mutants -c cargo mutants --package <name>` to surface
+behavior that coverage cannot prove protected. Let the default copy-to-`target`
+mode hold; `--in-place` is faster but leaves the source tree dirty on interrupt
+or panic, so reserve it for disposable checkouts. Treat surviving mutants as
+candidates for tighter assertions, equivalent-mutant write-offs, or
+unreachable-by-test code, and keep cargo-mutants a package-owner tool rather
+than a CI gate: equivalent mutants need human judgment, runtime scales with
+mutant count, and a survivor is a prompt to look, not a regression to block.
+
 ## Python style
 
 Default repo-owned Python apps to uv: `pyproject.toml`, committed `uv.lock`,

CLAUDE.md

@@ -304,6 +304,17 @@ against a debug-assertion standard library and surfaces some unsafe-precondition
 and stdlib-invariant breakage, but it does not model aliasing, uninitialized
 reads, or data races, so it complements Miri rather than replacing it.
 
+When auditing a crate with deterministic, fast tests, run
+[`cargo-mutants`](https://mutants.rs/) with
+`nix shell nixpkgs#cargo-mutants -c cargo mutants --package <name>` to surface
+behavior that coverage cannot prove protected. Let the default copy-to-`target`
+mode hold; `--in-place` is faster but leaves the source tree dirty on interrupt
+or panic, so reserve it for disposable checkouts. Treat surviving mutants as
+candidates for tighter assertions, equivalent-mutant write-offs, or
+unreachable-by-test code, and keep cargo-mutants a package-owner tool rather
+than a CI gate: equivalent mutants need human judgment, runtime scales with
+mutant count, and a survivor is a prompt to look, not a regression to block.
+
 ## Python style
 
 Default repo-owned Python apps to uv: `pyproject.toml`, committed `uv.lock`,

agents-md/sections/06-rust-style.md

@@ -33,3 +33,14 @@ against a debug-assertion standard library and surfaces some unsafe-precondition
 and stdlib-invariant breakage, but it does not model aliasing, uninitialized
 reads, or data races, so it complements Miri rather than replacing it.
 
+When auditing a crate with deterministic, fast tests, run
+[`cargo-mutants`](https://mutants.rs/) with
+`nix shell nixpkgs#cargo-mutants -c cargo mutants --package <name>` to surface
+behavior that coverage cannot prove protected. Let the default copy-to-`target`
+mode hold; `--in-place` is faster but leaves the source tree dirty on interrupt
+or panic, so reserve it for disposable checkouts. Treat surviving mutants as
+candidates for tighter assertions, equivalent-mutant write-offs, or
+unreachable-by-test code, and keep cargo-mutants a package-owner tool rather
+than a CI gate: equivalent mutants need human judgment, runtime scales with
+mutant count, and a survivor is a prompt to look, not a regression to block.
+