Fix SCHEMATIC silently shipping uninstrumented binaries

[byeongjee]

Summary

Fixes a latent SCHEMATIC bug that silently shipped uninstrumented binaries, and makes the benchmark result readable from NVM without a device-debug build.

The bug

ckpt bench/intermittent schematic --device-debug collected the execution trace with device_debug=False hardcoded (bench/schematic.py), then applied it to a -DDEVICE_DEBUG build. That define expands BENCH_INIT/BENCH_EXIT differently, renaming blocks in main, so TraceLoader dropped every function trace:

TraceLoader: BB 'do.end3' not found in main, skipping trace
TraceLoader: loaded 0 function traces, 39 loop traces for main
SCHEMATIC: no function traces for main — traces are required

solveFunction returned false, the driver ignored it, opt exited 0, and the pipeline linked and flashed an uninstrumented binary that died at the first outage (region_violation). Latent since the bash→Python rewrite; it only bites with --device-debug.

Minimal repro (aes, cap board):

trace collected	build	outcome
no-debug	device-debug	0 function traces → 0 region boundaries
device-debug	device-debug	1 function trace → 10 region boundaries

Changes

• Collect SCHEMATIC traces with the caller's device-debug setting — thread device_debug into collect_trace().
• Make SCHEMATIC trace-loading failures fatal — a missing or mismatched trace means the pipeline is wrong, so report_fatal_error instead of skipping the function. Infeasible solves still return false; the toolchain reports those as an infeasible result.
• Drop the mutatedIR out-parameter — it existed only to report PreservedAnalyses accurately after a failed solve; alloca hoisting runs before any failure, so returning PreservedAnalyses::none() unconditionally says the same thing without the flag.
• Commit the benchmark result to NVM in non-debug builds too — __nvm_result is now defined unconditionally and written by bench_commit_result(int), one extra FRAM word at BENCH_EXIT. Intermittent runs use --device-debug off (its counters and UART distort the energy behavior), and could not report correctness before. bench_commit_done and the identical debug_exit_commit collapse into bench_commit_result.

Verification

ckpt intermittent, trace 1, cap board, --device-debug off — 8/8 ok, every result matching the uninstrumented baseline (aes=107, crc=39423):

algo	bench	status	result	recovery	exec (us)	boundaries
milp	aes	ok	107	0	793,045	5
milp	crc	ok	39423	0	253,957	2
rockclimb	aes	ok	107	5	60,516,443	17
rockclimb	crc	ok	39423	4	51,363,351	8
schematic	aes	ok	107	10	126,451,035	10
schematic	crc	ok	39423	2	26,454,973	3
schematicO3	aes	ok	107	0	1,100,873	3
schematicO3	crc	ok	39423	0	364,818	3

uv run pytest tests/ -m "milp or rockclimb or schematic or unit" → 249 passed, 3 failed. All three fail identically on main: test_schematic_o3_dijkstra_loop_budget_uses_rare_inner_branch, and two rockclimb tests missing the adjcallstackdown/adjcallstackup energy keys.

Not addressed

• schematicO3/aes hits region_violation with --device-debug at cap board. Not an algorithm defect: the debug runtime's UART and counters cost energy the model does not account for, and schematicO3 has the thinnest margin (3 boundaries). Lowering the configured capacity to 75% makes it pass with debug on, and it passes at 100% with debug off.
• SCHEMATIC boundary placement is non-monotonic in configured capacity (38122→3 boundaries, 19061→2, 9530→3, 4765→6). No failure observed, but unexplained.