refactor(tui): remove the unused half of pi-test-support (plan D.5) - #63

Summary

Plan phase D.5. pi-test-support has seven modules that no test refers to. They are the mock inference server, the SSE builders, the ACP stdio clients, the headless runner and the counting server: test kit of the Rust agent runtime, which is gone. The workspace builds with its tests without them. They go, with the two helpers that only they called, and with the dependencies that only they needed.

 tui/crates/codegen/pi-test-support/
 ├── Cargo.toml                    # 11 dependencies go
 ├── README.md                     # written again
 └── src/
-    ├── mock_server.rs            # 2,259 lines
-    ├── sse.rs                    # 1,342
-    ├── inference_override.rs     #   678
-    ├── acp_client.rs             #   561
-    ├── headless.rs               #   361
-    ├── scripted.rs               #   174
-    ├── counting_server.rs        #    68
     ├── env.rs                    # grok_binary and git_workdir go; EnvGuard and env_parse stay
     ├── lib.rs                    # the re-exports of the removed modules go
     ├── process.rs                # stays: TestProcess
     ├── resources.rs              # stays: ResourceSnapshot
     └── sandbox.rs                # stays: TestSandbox

Two commits: the code, then the CHANGELOG.md entry. pi-test-support is a dev-dependency of pi-shell, pi-pager and pi-tools (publish = false).

Evidence

• Before: 14 files in the crate. After: 7 files. 5,646 lines go and 27 come. 47 test functions go with the modules (31 in mock_server, 13 in sse, 3 in headless).
• Before: pi-shell 838 passed and pi-pager 5,651 passed, 12 skipped (the references of the manifest).
  After: tui_baseline.py test (Linux, WSL2, nextest): pi-shell 838 passed, 0 failed, 1 skipped. pi-pager 5,651 passed, 0 failed, 12 skipped (includes python_agent_e2e). The 23 tests that remain in pi-test-support: 23 passed.
• tui_baseline.py check on the branch: warnings 0, consumer build exit 0, cargo check --workspace --tests --locked exit 0, graph 970 packages (ceiling 970), residue leader 243/243 and x.ai 297/297 (unchanged: the ceilings stay).
• A per-package build can differ from a workspace build, because cargo unifies features: tokio-util had the feature compat from this crate. cargo check --locked -p pi-tools --tests, and the suite builds of pi-shell and pi-pager, all pass.
• tui_notices.py check: 984 packages ship, 984 entries, 0 missing, 0 stale. Cargo.lock loses 11 lines and no package.
• This PR and refactor(tui): remove 16 dead modules, the MCP loaders and the folder-trust gate (plan D.8, decision 10) #62 together: a test merge of both branches (only CHANGELOG.md conflicts) has the same tui/Cargo.lock as cargo writes by itself (equal file hash), and tui_baseline.py check on it is all green: warnings 0, graph 970, leader 239/239, x.ai 282/282, cargo check --workspace --tests --locked exit 0.
• ruff check . and ruff format --check . pass. No Python file changes.

Merge Danger

Door: two-way. One revert brings the files back.

Blast Radius: tests only. No crate that ships depends on pi-test-support, so zypi does not change.

CHANGELOG.md conflicts with #60, #61 and #62: all add a section under [Unreleased]. Keep all of them. There is no other conflict. This PR and #62 both change tui/Cargo.lock, in different places. The result of both merged is the lock that cargo writes by itself (same file hash), so the order does not matter.