test(scheduler-core): normal-mode <-> simulation decision equivalence (task 11.6)

Chain the already-shipped recording (task 9.2) and sim-output (tasks
11.1-11.5) pipelines: record a synthetic multi-queue session through the
normal-mode driver, replay it through the simulation driver with
--sim-output, and assert the two .replay files' decisions match in full
structural content. This is the spec's central validation contract (Req 5
AC 7.1) -- if normal and simulation modes diverge, the core/driver
separation is broken.

Honestly scoped to PEND-only decisions: sim_loop.c's picker thread
(task 11.2) has no code path that can emit a live DISPATCH today, a
pre-existing gap that belongs to the not-yet-designed scheduler-allocation
spec, not something this task can close.

.kiro/specs/scheduler-core/tasks.md

@@ -345,11 +345,12 @@
   - _Depends: 11.2_
   - _Verified: 2026-08-06 — AC 16: `sim_loop.c`'s event/session read loop now selects `sim->alt_events_file` over `sim->replay_file` when the former is set (the initial snapshot, already loaded from the base file in task 11.1's startup, is untouched). AC 19: new `sim_record.c` writes a sim-run `ReplayHeader` (no cluster info — legitimate single-host branch `write_replay_header` already supports, `sched_policy` from `sim->core.job_pick_plugin`) + the loaded `SnapshotEntry`, then passes through each `ReplayEvent` and finalizes one `ReplaySession` per drain cycle, coexisting in the same loop with task 11.2's stdout default and task 11.4's `--sim-verify`. Kept the sim driver's zero-`xce_scheduler_t*` isolation intact (Req 5 AC 17) by extracting scheduler_t-free halves (`sched_record_session_ctx_new()`/`_finalize()`) out of `recording_session.c`'s existing `_begin()`/`_end()` rather than giving any sim-side file a `scheduler.h` dependency — confirmed behavior-preserving for normal-mode's existing callers (`scheduler record lifecycle test` still passes) and confirmed via `test_sim_isolation.c`'s extended 3-file (`sim.c`+`sim_loop.c`+`sim_record.c`) source-scan/struct-scan/nm(1) coverage. New `test_sim_alt_events.c` (AC 16) and `test_sim_record_output.c` (AC 19, decodes the sim-produced file via the same `decode_*` capnp typemap functions normal-mode recording round-trips against) cover both Observables. Round 1 review REJECTED solely on invalid self-reported RED-phase evidence (a `git stash` scope that hard-failed the build rather than reproducing the claimed 3 test failures) — round 1 explicitly confirmed the code itself was correct; round 2 fix produced genuine RED evidence via a direct in-place probe (hardcode the AC16 selector, no-op the AC19 write gates) showing 30/32 with real failures in exactly the two new tests, then reverted to 32/32. Round 2 review independently reproduced that same RED/GREEN cycle with its own probe and APPROVED. 32/32 scheduler suite (was 30/30), whole-project build clean.
 
-- [ ] 11.6 Deterministic replay and normal-to-simulation equivalence test
+- [x] 11.6 Deterministic replay and normal-to-simulation equivalence test
   - Record a synthetic scheduling session through the normal-mode driver, then replay the resulting file through the simulation driver, and assert the decision sequences match exactly (same jobid, workerid, allocation amounts, pend reasons)
   - Observable: the equivalence test passes with zero decision-content divergence between the recorded and replayed runs — this is the spec's central regression test
   - _Requirements: 5.7_
   - _Depends: 11.4, 9.2_
+  - _Verified: 2026-08-06 — new `test_sim_equivalence.c` (740 lines, no production code changes — this task chains two already-shipped pipelines) records a synthetic session (2 queues at different priorities, 4 jobs) through a real normal-mode `xce_sched_driver_run_normal_mode_round()` with recording active (task 9.2's `sched_record_start()`) to produce `.replay` file R, then replays R through `sched_sim_driver_run()` with `--sim-output` (tasks 11.1-11.5) to produce `.replay` file S, decodes only each file's `ReplaySession.decisions` via `decode_ReplaySession_ptr()` (deliberately never decoding `ReplayHeader`/`SnapshotEntry`/events, so the test structurally cannot assert on the timestamp/wall-clock metadata AC 7.1 explicitly excludes), and asserts R and S's decisions match in count, type, and full structural content (reusing `sim_verify.c`'s `xce_job_alloc_t`/`xce_job_preason_t` field-by-field comparators, not a jobid-only shortcut). RED phase independently reproduced by review (corrupted a decoded jobid post-decode, confirmed FAIL, reverted, confirmed PASS). Honestly scoped to PEND-only decisions: independently confirmed (by review, reading `sim_loop.c`'s `sim_picker_thread_fn()` directly, committed under task 11.2) that the simulation picker thread has no code path capable of emitting a live DISPATCH decision at all — a genuine, pre-existing structural gap predating this task (real resource-matching allocation belongs to the not-yet-designed `scheduler-allocation` spec), not a shortcut taken here; documented prominently rather than silently avoided. 33/33 scheduler suite (was 32/32), whole-project build clean. Review APPROVED, zero blocking findings.
 
 - [ ] 12. Core: Interactive REPL