test: build the native crate's suites unoptimised

[MarcusKainth]

What this changes, and why

[profile.test] sets opt-level 2 for the emulator's fixture and trace suites. The native crate got it too, and its test binaries are the largest compile cost in build-tests: 876 CPU seconds of the workspace's own units at opt-level 2, against 200 at opt-level 0 (cargo --timings from the measurement run 36121287631). The native crate builds SQL text and its suites wait on the server, so this gives it opt-level 0 through [profile.test.package.clickdoom-native]. Every other crate keeps opt-level 2, refemu included.

Evidence

The generated SQL is byte-identical

A scratch test, not committed, hashing everything the crate generates:

let (a, b) = tick::resident_statements("nat");            // xxh64 of each
tick::run_statement("nat", &[Input { tic: 7, source: 1, keys: 3, mouse: (1, -2) }]);
tick::demo_statement("nat", 1, 40);                         // sql, body, settings
load::plan + sql::level_statements("nat","E1M1","demo3") + sql::render_statements("nat","SKY1");
sql::parity::first_divergence("nat") + sql::parity::field_summary("nat");

Run under each profile, with the level confirmed from cargo test -v (-C opt-level=2, =1, and no flag for 0):

$ bash cmphash.sh; echo "exit=$?"
hash-o2        9 lines 84e10ba884417051
hash-o1        9 lines 84e10ba884417051
hash-o0        9 lines 84e10ba884417051
identical across opt-level 2, 1, 0
exit=0

stage1 2179678 ddbc3c3eca51187b
stage2 203311 1040c854b1969203
load+level+render e5990d653f124421
parity 90f73aa513156ae4

Rerun on a Linux runner in the throwaway run 36123894500. cargo test -v showed opt-level=2 for the first build and no flag for the second:

opt-level 2: 0ad7d300a3f5c0a9b57e373676c8ecbaaeb2a274bec5255a56fa921f985122cc  -
stage1 2179678 ddbc3c3eca51187b
stage2 203311 1040c854b1969203
opt-level 0: 0ad7d300a3f5c0a9b57e373676c8ecbaaeb2a274bec5255a56fa921f985122cc  -
stage1 2179678 ddbc3c3eca51187b
stage2 203311 1040c854b1969203
identical at opt-level 2 and 0

Compile time, from this PR's CI run 36123100020

build-tests restored the same rust-cache key main uses (full match: true), so only the workspace crates compiled:

	before: run 36116301376 (PR #580)	this PR
Finished test profile	6 m 17 s	2 m 52 s
Archive the suites (with the ROM suites archive)	448 s	235 s

cargo --timings in the measurement run: the native crate's test units took 876 CPU s at opt-level 2, 607 at opt-level 1 and 200 at opt-level 0.

Suite times

The emulator group doesn't regress. These are test-seconds summed from each run's nextest PASS lines. Main's run and #584's are on this PR's base commit, and #580's is on an earlier commit. Together they show the spread from one runner to the next:

emulator	36116301376 (#580)	36122388569 (main)	36122982586 (#584)	this PR
sqlcpu_live the_sql_cpu_passes_its_own_suite	135.9	101.3	130.1	134.2
demo3_parity	29.0	22.2	29.0	29.0
run_live	15.9	12.0	15.8	15.6
group sum	250.5	197.2	243.8	251.1

refemu's own suites, which this PR can't touch, moved from 36.8 to 48.5 test-seconds between main's run and this one. That is the runner-to-runner spread.

native-rest's live suites match run 36116301376 within a few seconds each (group sum 653.1 there, 685.7 here). The pure-Rust statement tests get slower at opt-level 0: every_cut_builds_a_statement_the_server_could_run goes from 3.2 to 5.0 s to 15.3 s, and no_cut_statement_holds_a_comparison_operand from 1.6 to 2.8 s to 8.2 s. About 15 test-seconds in all, in a group that runs two at a time. The simulation groups' sums fall inside the spread between runs (sim-f: 908.8 main, 969.7 here).

Every group passed and ci-passed succeeded.

Invariants

None. Only the test profile of one crate changes. opt-level doesn't change what safe Rust computes, and the hashes above show the SQL text is unchanged.

Spec impact

• None. No contract in SPEC.md is touched

Checks

• make gates: not run locally. CI's lint job ran make lint and every test group passed
• No AI attribution trailers in the commits

Written mostly by Claude Opus 5.5.