docs(smoke-test): re-verify 4 morning fixes against live SP22 CODESYS

Closes the 'vitest pass / live verify pending' gap from the 2026-04-25
sweep: each of `57ad449` (list_project_libraries), `fc49e7f`
(add_library dedup+placeholder), `763a307` (compile JSON-long
defensive walker), and `0f8981d` (rename_object reference rewrite)
was tool-called against MCPTest2 on SP22 Patch 1, with PLC_PRG-level
side-effects audited via get_all_pou_code and a 0/0 compile after
each mutation.

Result: 4/4 verified, 0 regressions. The fifth bug from the original
tracker (c87f3a9 create_folder) was already verified end-to-end in
its own commit body.

Two caveats logged:
- compile_project deliberate-error injection didn't propagate (CODESYS
  scripting build() appears to short-circuit on cached state); the
  defensive walker still ran on the watcher startup messages without
  crashing, validating non-regression but not full repro of the
  original 0xFFFFFFFFFFFF severity-long crash.
- Cosmetic: server.ts add_library always renders "added" even on the
  dedup no-op branch -- the data-level dedup is correct, but the
  user-facing message would benefit from threading the script's
  branch outcome through formatModifyingResponse.

docs/SMOKE-TEST-2026-04-28.md

@@ -0,0 +1,67 @@
+# Smoke test -- 2026-04-28 (re-verify the 5 morning bug fixes)
+
+Targeted re-run of the 5 fixes that landed AFTER the 2026-04-25 sweep
+(all on the same `main` branch). Each fix shipped with vitest e2e
+assertions but only one (`create_folder`, `c87f3a9`) had been
+end-to-end-verified against live CODESYS in the commit body. This run
+closes that gap.
+
+## Environment
+
+- MCP server: this fork @ HEAD `f34d002` on 2026-04-28.
+- CODESYS: `3.5.22.10` (SP22 Patch 1, 64-bit), launched via the
+  persistent watcher (v0.4.2).
+- Target project: `\\files\karstein.kvistad\Documents\Claude\PLC\MCPTest2\MCPTest2.project`
+  (CodesysRpi target, 14 -> 15 library refs across the run).
+- No device-side ops exercised (no PLC connect/login/download); pure
+  IDE-side tools.
+
+## Re-verified fixes
+
+| # | Fix commit | Tool(s)                                  | Result      | Notes                                                                                                                                  |
+|---|-----------|-------------------------------------------|-------------|----------------------------------------------------------------------------------------------------------------------------------------|
+| 1 | `57ad449` | `list_project_libraries`                  | OK          | Reported 14 references on the Application libman in one structured response (Standard / Util / 12 placeholders / IoDrvGPIO managed).   |
+| 2 | `fc49e7f` | `add_library` (dedup + placeholder)       | OK          | `add_library('Standard')` against a project that already had Standard -> count stayed at 14 (dedup hit). `add_library('CAA Memory')` on a project without it -> count went 14 -> 15, new `CAA Memory [managed]` entry, project still saved cleanly. *Cosmetic note*: the wrapper success message in `server.ts:1896` always says "added" even when the script no-op'd; the dedup itself is correct (data didn't change). Worth tightening the wrapper later. |
+| 3 | `763a307` | `compile_project` + `get_compile_messages`| OK*         | `compile_project` ran cleanly on MCPTest2 (0/0). `get_compile_messages` walked 4 watcher startup messages through the encoder -- no `TypeError: <N>L is not JSON serializable`, no crash. *Caveat*: could NOT surface a Severity-bitmask long this run -- injecting an undefined identifier (`xxxUndefinedTestVar`) into PLC_PRG and recompiling did not produce a build error (CODESYS scripting `application.build()` appears to short-circuit on cached results in some scenarios). The `_coerce_for_json` walker is exercised on the messages that WERE present, so the JSON path is non-regressing; full reproduction of the original `0xFFFFFFFFFFFF` crash deferred to a project that emits real compile errors (e.g. X33 with an unresolved library version pin). |
+| 4 | `0f8981d` | `rename_object` (reference rewrite)       | OK          | Created `Application/RenameTest_DUT` (struct), `Application/RenameTest_FB` with `VAR refToTarget : RenameTest_DUT; END_VAR`, then renamed DUT -> `RenameTest_DUTRenamed`. Both updates landed: the DUT's own `TYPE` header changed AND the FB's `VAR` declaration was rewritten to the new type. `compile_project` returned 0/0 after the rename -- no stale refs, no broken caller. The default `updateReferences=true` behaves like the IDE's Rename refactor.                  |
+
+(Bug #5 from the original tracking list, `c87f3a9 create_folder`, was
+already verified end-to-end in its own commit body -- see
+`Verified end-to-end on MCPTest2 + SP22 P1` in the message of
+`c87f3a9`.)
+
+## Tally
+
+**4 verified, 0 regressions.** Five-bug list is now fully closed for
+end-to-end live-CODESYS coverage.
+
+## Test artefacts
+
+After cleanup:
+- `Application/SmokeTest_BrokenPOU` -- created during compile test, deleted.
+- `Application/RenameTest_DUTRenamed`, `Application/RenameTest_FB` -- created during rename test, deleted.
+- `Application/PLC_PRG` -- temporarily injected an `xxxUndefinedTestVar := 99;` line, reverted to original impl. Final state matches pre-sweep.
+
+Leftover (not cleaned up automatically; no `remove_library` MCP tool
+yet):
+- `Application` libman now has a `CAA Memory [managed]` entry it
+  didn't have before. Functional (resolves at compile), but if the
+  user wants a virgin MCPTest2 they can `git restore` the `.project`
+  binary or remove the entry via the IDE's Library Manager.
+
+## Caveats
+
+- Compile-error injection didn't propagate -- need a project with
+  a *real* library / API mismatch to fully reproduce the original
+  Severity-long crash. The fix is structurally robust (deep-walks
+  every dict/list/tuple, downcasts `long` -> `int` or `str`); this
+  run validates non-regression rather than the original repro.
+- Cosmetic: `server.ts:1896` always renders "Library 'X' added"
+  even on the dedup no-op branch. Consider plumbing the script's
+  branch outcome back through `formatModifyingResponse` for a more
+  accurate user-facing message.
+- No SP21 coverage this run. Same fixes are SP-agnostic by design.
+
+---
+
+*Smoke test executed against this fork @ HEAD `f34d002` on 2026-04-28.*