chore(benchmark): archive the unreferenced SWE-bench Rust runner

[Harry19081]

Problem

The 2026-08-16 SWE-bench "Benchmark (Beta)" archival parked the whole frontend
surface — panel, run builder, tab renderer, store atoms, benchmarkApi client,
E2E helpers and specs — but deliberately left the Rust side in place. Its own
.archive/README.md entry recorded the consequence:

src-tauri/src/benchmark/ — the Rust runner/commands still compile and are
registered; they are now unreferenced from the frontend and can be removed in
a backend-only PR

Since then src-tauri/src/benchmark/ (3,542 LOC) has compiled into every build
and registered 13 benchmark_* Tauri commands that nothing can call. A sweep of
the 1,019 entries in src-tauri/src/commands/handler_list.inc against every
non-Rust source file in the repository confirms it: none of the 13 appears in
src/, tests/, or src-tauri/capabilities/, and no Rust caller reaches the
module either — the only inbound edge outside the module was the exit-time
evaluator cleanup in app/lifecycle.rs, which exists solely to reap processes
the (now unreachable) runner would have spawned.

This is the backend-only PR that entry called for.

Solution

git mv src-tauri/src/benchmark/ .archive/src-tauri/src/benchmark/, matching
the existing archival convention: paths mirror their original location, nothing
under .archive/ is compiled or bundled, and restoring is a reverse git mv.

Unwired in place:

• src-tauri/src/commands/handler_list.inc — dropped the 13
  benchmark::benchmark_* registrations and their section comment
• src-tauri/src/lib.rs — dropped pub mod benchmark;
• src-tauri/src/app/builder.rs — dropped benchmark from the use crate::{…}
  scope that handler_list.inc resolves against
• src-tauri/src/app/lifecycle.rs — dropped the use crate::benchmark; import
  and the exit-time benchmark::terminate_running_evaluators_sync() call

Resulting invariant: the app's IPC surface no longer advertises a benchmark_*
command, and no code path can spawn a SWE-bench evaluator subprocess, so the
exit-time reaper it guarded has nothing left to reap.

.archive/README.md's SWE-bench section is updated so the entry that asked for
this removal now records that it happened, together with the in-place edits
needed to reverse it.

Not the Housekeeper token benchmark (housekeeper_token_benchmark,
rpc.validation.housekeeperTokenBenchmark) — a different feature, untouched and
still registered.

Potential risks

• Persisted user state: none. The runner wrote its run/batch history under
  the app data dir; those files are simply no longer read. Nothing is deleted
  from disk by this change, so restoring the module restores access to them.
• Wire/IPC compatibility: removing 13 commands changes the invoke surface. A
  frontend build that still called them would now fail at runtime instead of
  succeeding — but the callers were archived in the same repository three weeks
  ago, so no shipped or in-tree frontend references them. Verified by an
  exhaustive scan, not by grep of a single directory.
• Exit-time behavior: terminate_running_evaluators_sync() is gone from the
  shutdown path. If any user still has an orphaned evaluator process from a
  build predating this change, it will no longer be reaped at app exit. Those
  processes are Docker/CLI children of a feature that has had no UI entry point
  since 2026-08-16, so the practical exposure is nil.
• Unverified paths: the archived module is no longer type-checked. Any future
  restore must re-run cargo check rather than assume it still compiles.
• Rollback: reverse the git mv and revert the four in-place edits (all
  recorded in .archive/README.md), or revert this commit wholesale.

Verification

• cargo check --workspace --all-targets — passes, 44 crates, no warnings
  introduced.
• cargo clippy --workspace --all-targets -- -D warnings — passes (exit 0).
• Registry sweep: parsed all 1,019 entries of handler_list.inc and searched
  every .ts/.tsx/.js/.jsx/.mjs/.cjs/.json/.html/.md/.rs/.toml/.yml/.yaml file
  in the repository (9,690 files, excluding node_modules, target,
  .archive) for each command name. The 13 benchmark_* commands have zero
  hits outside handler_list.inc and the module itself.
• Checked for indirect invocation: the repository's only non-literal invoke()
  call sites (useEmbeddedWebview, diagnostics/rustBridge) both receive
  command names as constants, so no command is reachable by a constructed
  string. No benchmark_* entry appears in src-tauri/capabilities/default.json.
• Not run: no UI change, so no screenshots. pnpm typecheck / vitest /
  Playwright were not run — this PR changes no TypeScript, and the frontend
  callers were removed in a prior PR.
• Pre-commit hook trailer is absent. This was committed from a git worktree
  where .husky/_/husky.sh does not exist (gitignored, so git worktree add
  never creates it) and node_modules is not installed. The hook's
  "Pre-commit hook ran." tamper-evidence trailer is therefore always missing in
  a worktree and means nothing here; the checks it would have run (cargo clippy) were run manually and are reported above. No TypeScript is touched,
  so lint-staged/tsc had nothing to check.