[Raikaru]

Baseline-settings pass, follow-up to #5. Per-function optimizer toggles are matching tools, not original source. For each first-party function whose toggles differed from the rest of its unit, this pass removed the toggle and looked for plain C that matches under the unit's own settings. Repeated per-function push/toggle/pop scaffolding is replaced by one commented unit setting near the top of the file. Matching is unchanged at every commit: 6860/6860 first-party MATCH, 8101 C functions linked, both retail SHA-1s OK. Recovery quality stays 100% named/typed/documented.

First-party optimizer-toggle lines	main	This PR
opt_loop_invariants	1,068	411
opt_propagation	1,046	375
opt_rebuildconditionals	276	46
opt_common_subs	254	96
opt_lifetimes	80	58
opt_dead_assignments	35	13
opt_pulloutconstants	23	19
opt_scalarize	21	4
Total (files with any)	2,803 (315)	1,022 (294)

• Converted functions. Some needed no change at all once the toggle was gone. Others needed small honest source changes, such as a named snapshot of a loop-invariant value, a separate block-local counter instead of a reused local, a typed table pointer, or a switch in place of an if chain. No new helpers, volatile, asm or dummy locals.
• Kept toggles. Each one has a one-line measured: comment saying what was tried and the residual without it. Per-function probe records are in the slice TSVs (not committed).
• Untouched. schedule, tailcall, no_branch_likely and VU0 settings are unchanged. Six redundant optimization_level 2 restores inside repeated scaffolding were removed, and the files still match.
• Lint. All five pre-existing unbalanced #pragma push scopes (P001) are fixed, so decomp_lint.py --errors-only is now clean across first-party code (it reported 5 errors on main). H003 warnings dropped in every slice.
• Docs. docs/STYLE.md now says unit settings are stated once and per-function toggles need a measurement. It also notes that an optimization_level directive can reset the toggles.

Verification

Each commit was gated on whole-tree tools/verify.py (6860/6860), tools/build.py (8101 linked, both SHA-1s), no new decomp_lint.py errors in changed files, and the unit tests. Every slice also re-ran tools/recovery_quality.py on its files.

[decomp-dev]

Report for SLUS_217.82 (108806e - 6358f06)

No changes