t27c suite: run the in-process phases 1a2-1a7 through the parallel pool, workers on an 8 MiB stack (#8838 follow-up) #8841

Part of #8737 (hot reload); a follow-up to #8837 and #8840.

Problem

• Phases 1a2-1a7 (parse completeness, vacuous invariants, vacuous Verilog tests, backend omissions, bare keywords, literal percent) are the last per-spec phases still run one spec at a time: about 17 s of the 167 s suite.
• t27c suite runs its read-only per-spec subprocess phases in parallel -- those phases 269 s -> 71 s, the suite 498 s -> 276 s (Closes #8829) #8837 kept them sequential on purpose, because they call the compiler in-process.

Audit

• No files, no subprocesses: none of the six writes a file or spawns a process. Each reads its spec and calls Compiler::parse_ast_accounted, compile, compile_verilog or compile_verilog_for_simulation.
• No shared mutable state in the compiler: the only global in compiler.rs is an AtomicU64 that makes unique CSE keys (UNCSE:n), and those keys never reach the output. The other globals in bootstrap/src (a lazy_static table in formula_eval.rs, atomic counters in service/catalog_gate/host, a thread_local in service.rs) are thread-safe or not on this path.
• Stack: the suite ran these phases on the main thread, 8 MiB on macOS, with no enlarged stack elsewhere in t27c. Spawned threads get 2 MiB by default.

Change

• Workers: par_in_order spawns its workers with an 8 MiB stack (std::thread::Builder::stack_size), the main thread's own.
• The six phases move to run_phase_par.
• suite_jobs.t27's comment names every parallel phase and why the rest stay sequential.

Acceptance criteria

• Same stderr, same stdout (apart from the JSON path line) and the same JSON (apart from *_elapsed_ms) with T27_SUITE_JOBS=1 and with the default.
• Phases 1a2-1a7 take at least 2x less wall time with the default.
• Hand-written foreign lines do not grow.

Boundary

specs/verified/suite_jobs.t27

User Scenarios

• Given the test suite runs with the default parallelism, When phases 1a2-1a7 (parse completeness, vacuous invariants, vacuous Verilog tests, backend omissions, bare keywords, literal percent) execute, Then they run through the parallel worker pool with 8 MiB stack per worker instead of sequentially on the main thread.
• Given the test suite runs with T27_SUITE_JOBS=1, When phases 1a2-1a7 execute, Then they produce identical stderr, stdout (apart from the JSON path line), and JSON output (apart from *_elapsed_ms) as the default parallel run.

Requirements

• FR-001: The six phases 1a2-1a7 MUST be moved from sequential execution to run_phase_par so they run in the parallel worker pool.
• FR-002: The parallel worker pool MUST spawn workers with an 8 MiB stack size using std::thread::Builder::stack_size.
• FR-003: The comment in specs/verified/suite_jobs.t27 MUST be updated to name every parallel phase and explain why the remaining phases stay sequential.
• FR-004: Running the suite with T27_SUITE_JOBS=1 and with the default parallelism MUST produce the same stderr, same stdout (apart from the JSON path line), and same JSON (apart from *_elapsed_ms).
• FR-005: Phases 1a2-1a7 MUST complete in at least 2x less wall time with the default parallelism compared to the previous sequential execution.
• FR-006: Hand-written foreign lines MUST not increase as a result of this change.