refactor(tests): adopt libtest-mimic and custom snapshot_test crate (#339)

## Why

Before this PR, both `e2e_snapshots` and `plan_snapshots` used `harness
= false` with a hand-rolled `fn main()` that:

- Parsed one positional arg as a substring filter (`args().nth(1)`), and
nothing else — no `--list`, no `--exact`, no `--ignored`, no
`--test-threads`.
- Iterated fixtures sequentially in a single thread, regardless of
machine parallelism.
- Printed fixture names to stdout as progress, mixed into `insta`'s own
stderr diff output, with no structured "passed / failed / filtered out"
summary.

Two problems fall out of that:

- **AI-agent friction.** Agents (Claude Code, Cursor, etc.) drive tests
through `cargo test` and expect the built-in harness contract: filter a
subset with `--exact`, enumerate with `--list`, read the standard "N
passed; M failed" summary to know what happened. Custom output and
missing flags force the agent to grep through ad-hoc progress prints or
re-run the full suite just to isolate one fixture. Moving to
`libtest-mimic` removes that friction — the harness behaves exactly like
a normal Rust test binary.

- **No parallelism.** `plan_snapshots` in particular is CPU-bound and
embarrassingly parallel per fixture, but the old harness ran it
single-threaded. `libtest-mimic` runs fixtures in parallel by default
(via `available_parallelism()`), using the hardware the CI already pays
for. (E2E tests spawn PTY children and send signals; on Linux we pin
them back to single-threaded — see the "Platform-specific fixes" section
below.)

## Summary

Two-part migration of the e2e and plan snapshot test harnesses:

### 1. Adopt `libtest-mimic` for proper `cargo test` integration

Migrating to `libtest-mimic` gives:

- Proper `cargo test` flag support: `--filter`, `--list`, `--exact`,
`--ignored`, `--test-threads`
- Standard test runner output formatting (`running N tests` / `test X
... ok` / `test result: ok. ... passed; ... failed`)
- Parallel fixture execution by default

### 2. Replace `insta` with a custom `snapshot_test` crate

Created `crates/snapshot_test` with two methods that return `Result<(),
String>` — containing a unified diff — so failures integrate cleanly
with `libtest-mimic::Failed`:

- `check_snapshot(name, actual)` — compares raw text
- `check_json_snapshot(name, comment, value)` — serializes to pretty
JSON, prepends `// {comment}` header, writes to `.jsonc`

Key behaviors:
- Drops the `insta` dependency and its transitive deps
- Unified diff appears directly in the `failures:` section of test
output (see
[`crates/snapshot_test/README.md`](crates/snapshot_test/README.md) for
why `insta` didn't fit)
- Plan JSON snapshots now use `.jsonc` extension with `// run ...`
command headers (e.g. `// run build --recursive`); e2e snapshots keep
`.snap`
- Existing `insta` YAML headers stripped from all 310 snapshot files
- `UPDATE_SNAPSHOTS=1` env var accepts new output in-place (replaces
`INSTA_UPDATE`)
- Snapshot directories excluded from `oxfmt` via `.oxfmtrc.json` to
prevent trailing-comma insertion in `.jsonc` files

### Platform-specific fixes

- **Windows**: CRLF line endings are normalized to LF when reading
snapshot files
- **Linux**: e2e tests forced to single-threaded (`test_threads = 1`),
matching the pre-PR custom-harness behavior. `libtest-mimic` defaults to
`available_parallelism()`, which caused PTY/signal-routing races in the
`ctrl-c` test. macOS and Windows keep parallel execution.

## Test plan

- [x] `cargo test -p vite_task_plan --test plan_snapshots` — 45 tests
pass
- [x] `cargo test -p vite_task_bin --test e2e_snapshots` — 36 tests pass
- [x] `cargo test -- --list` lists all fixture names
- [x] `cargo test -- --exact <name>` filters correctly
- [x] Tampered `.snap`/`.jsonc` file produces a unified diff in the
failure message, pointing to the `.snap.new` / `.jsonc.new` file
- [x] `UPDATE_SNAPSHOTS=1` accepts new output in-place and clears `.new`
files
- [x] CI stability: 5 consecutive green runs across Linux, macOS (arm64
+ x86_64), Windows, and musl

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: branchseer

.oxfmtrc.json

@@ -1,4 +1,8 @@
 {
   "singleQuote": true,
-  "ignorePatterns": ["crates/fspy_detours_sys/detours", "crates/vite_task_graph/run-config.ts"]
+  "ignorePatterns": [
+    "crates/fspy_detours_sys/detours",
+    "crates/vite_task_graph/run-config.ts",
+    "**/fixtures/*/snapshots"
+  ]
 }

CLAUDE.md

@@ -47,7 +47,7 @@ cargo test                                              # All tests
 cargo test -p vite_task_bin --test e2e_snapshots        # E2E snapshot tests
 cargo test -p vite_task_plan --test plan_snapshots      # Plan snapshot tests
 cargo test --test e2e_snapshots -- stdin                # Filter by test name
-INSTA_UPDATE=always cargo test                          # Update snapshots
+UPDATE_SNAPSHOTS=1 cargo test                           # Update snapshots
 ```
 
 Integration tests (e2e, plan, fspy) require `pnpm install` in `packages/tools` first. You don't need `pnpm install` in test fixture directories.

CONTRIBUTING.md

@@ -45,7 +45,7 @@ cargo test                                              # All tests
 cargo test -p vite_task_bin --test e2e_snapshots        # E2E snapshot tests
 cargo test -p vite_task_plan --test plan_snapshots      # Plan snapshot tests
 cargo test --test e2e_snapshots -- stdin                # Filter by test name
-INSTA_UPDATE=always cargo test                          # Update snapshots
+UPDATE_SNAPSHOTS=1 cargo test                           # Update snapshots
 ```
 
 Integration tests (e2e, plan, fspy) require `pnpm install` in `packages/tools` first. You don't need `pnpm install` in test fixture directories.