chore: comment and test cleanup against the new rules

[ipfizz]

Applies the rewritten comment and test rules from .admin/CONTRIBUTING.md across the tree. No behaviour change.

Comments

docs: correct the comments that describe behaviour the code no longer has fixes ~24 sites where the prose was wrong, not just verbose. The Windows backend landed after most api/ headers were written, so a dozen files described a two-platform world while importing a windows*Backend at the top of the same file. Also: exec-wrapper said "both backends" on line 4 and "all three" on line 24; window-controls promised "up to three things" and listed two; native-image listed getAspectRatio as deferred 112 lines above its implementation; auto-updater promised a detached install helper that does not exist; build-windows carried a dangling No from a copy-paste; (D30) should have been (D030) and was invisible to any decision sweep.

docs: cut narration comments across src then does the volume pass.

	before	after
comment lines	6,887	5,255
share of src	23%	18%
blocks > 4 lines	477	359
files opening with a header essay	184 / 189	158 / 189

Net −1,607 lines across 173 files.

Verification: every file was transpiled with comments stripped and code-hashed before and after. All 189 hashes match, so the diff is provably comment-only. The verifier was itself checked both ways — a comment-only edit reports zero drift, a one-line code change is caught.

Kept deliberately: the GTK clipboard deadlock rule, the Carbon EventHotKeyID struct-by-value derivation, every "never close a JSCallback inside its own invocation" note, the setReleasedWhenClosed:NO use-after-free note, ordering rules, security arguments, all 92 D0xx citations, and every C prototype above an FFI table — FFIType.u64 does not tell you which pointer is the out-param.

Tests

Removes change-detector tests: 12 Linux *-ffi.test.ts files (114 it() cases) that read a FFIType literal out of the source and assert it equals the same literal five lines away. Replaced by one linux-ffi-abi.test.ts keeping only assertions that encode a real C gotcha — gboolean is i32 not bool, a getter returns pointer not cstring so NULL is guardable, a variadic pinned to one arity, the finite DBUS_CALL_TIMEOUT_MS that is the four-hour-CI-hang scar.

Also: four platform-guard files merged into one table-driven platform-guards.test.ts; cocoa-msgsend-variants.test.ts collapsed from 24 near-identical cases to one; entry-barrel.test.ts rewritten to cover strictly more — it now loops IMPLEMENTED_MODULES and asserts every claimed module is really exported, which nothing checked before and which src/electron.ts depends on.

Two guards got stronger, not just smaller: the 14 Linux loader guards used to execute zero assertions on the Linux runner via an early return, and now report honestly as skips.

Deleted as dead: src/main/platform/windows/win32-window.ts (266 lines, imported by nothing in src/, kept alive only by its own test) and the unused has/clear on the Cocoa class and selector caches.

	before	after
darwin	1615 pass / 79 skip / 0 fail	1502 pass / 74 skip / 0 fail
test files	239	223

Net −1,514 lines. The darwin budget is ratcheted to the measured number (1400/90); linux and win32 keep the loose floor until their real counts can be read off a green CI run, since a guessed bound would false-fail.

bun run validate EXIT=0 locally on macOS at every commit.