[gmh5225]

Follow-up to #1618: what still fails or drifts on dev after #1610-#1614 and #1604. This branch was rebuilt on dev after #1604 merged. #1604 independently fixed the stale HLT test, the Make array/alignment icons, the call-type menu entries, the search during a pending jump and the MSVC worker compiler, so my versions of those were dropped.

• Heartbeats of another project (Reuse analysis replicas for bounded source preparation #1612). The worker's heartbeat thread builds its message under the state lock and sends it afterwards, so a heartbeat built just before a reply that opens another project can arrive after that reply. The GUI applied it to the new project: its discovery_complete released Reuse analysis replicas for bounded source preparation #1612's large-image replica gate early, and its function count and generation briefly replaced the new project's. Heartbeats now apply only when their project_id is the one the session's replies established. That interleaving cannot be produced deterministically through the fixture worker, so it has no dedicated test; the normal path is covered by the GUI suite, which depends on heartbeat updates throughout.
• Compiler for generated C in GCC builds. 082b2e3 (in Fix source contracts and cross-platform validation failures #1604) moved the worker tests that compile the engine's C to Clang for cl.exe builds. GCC's C front end lacks the same builtins (__builtin_bit_cast), so a GUI-project build with GCC still failed NeverDWorkerCallSiteTypes on a floating-point variadic argument. GCC builds now prefer an installed Clang too, keeping GCC only when there is none.
• Deterministic search-during-jump test. c96ff8d (in Fix source contracts and cross-platform validation failures #1604) starts a search after a pending jump settles; searchResultsClearOnFailureCloseAndProjectSwitch reached that only through timing. searchDuringAJumpStartsWhereTheJumpLands navigates and presses Next before the listing answers. The listing replies first, so the test fails deterministically when the search starts from the cursor the jump is leaving (it failed twice that way before the fix).
• Formatting. 4328089 and 2da037c left whole-file clang-format 22.1.2 differences in four files, which fails scripts/check_clang_format.py for any later change that touches them. Formatting only.
• MCP comment. The server's warm:true open no longer mirrors the workbench, which Reuse analysis replicas for bounded source preparation #1612 stopped warming. The comment now gives the server's own reason, which 0a2b35d measured.

Local verification (Linux x64, Release) on this branch: see the comment below. The PR stays open until its Linux, Windows and macOS checks finish.

[gmh5225]

Local verification on 643af8c (Linux x64, Release; engine with clang 23, GUI project with GCC 15 and Qt 6.8.3):

• NeverDSessionCAPITests: 197 passed
• NeverDHighControlFlowTests: 430 passed
• NeverDOriginalBinaryUndefinedIndependenceTests: 114 passed (after the formatting-only change)
• Engine-build worker tests (CallSiteTypes, StackFrame, SourceVariableTypes, SourceVariableEdits, CLIStackFrame, DiscoveryReady, DiscoveryReadyLegacy, CLIInstructionFlow): 8/8 passed
• GUI-project CTest (all 54, including every NeverDGui* and NeverDWorker* registered there): 54/54 passed against the freshly built engine. NeverDWorkerCallSiteTypes now compiles the engine's C with the installed Clang in this GCC build.
• searchDuringAJumpStartsWhereTheJumpLands and searchResultsClearOnFailureCloseAndProjectSwitch pass repeatedly.
• scripts/check_clang_format.py with clang-format 22.1.2 and scripts/check_gui_i18n.py (structure and access keys) pass.

Merged on local verification at the maintainer's request, without waiting for the hosted checks.