refactor(desktop): remove dead shell CSS recipes and close check-dead-css blind spots

[UncertaintyDeterminesYou4ndMe]

Summary

Implements #1980 end to end: fixes the three check-dead-css blind spots, then deletes everything the fixed scanner reveals — the audit's 13 dead rules were still present on main (the gate printed "no dead classes or tokens found ✓" over them the whole time).

Scanner fixes (scripts/check-dead-css.mjs)

1. Class scan now includes maka-tokens.css — previously it was scanned for tokens only, so .maka-composer-toolbar and .maka-shimmer were invisible outright.
2. Exact class-name matching — the substring check let maka-shell-astryx keep .maka-shell alive, maka-chat-message-* keep .maka-chat alive, etc. Consumers now match with hyphen-aware boundaries (\b alone treats - as a boundary, so plain word-boundary matching would not have fixed the collision).
3. Sheet-internal token reads are attributed, not blanket-seeded — a var() read in a class rule counts only when the rule's classes have consumers; a read inside another token's value counts only once that token is live (fixpoint over derivation edges). This is what let the dead shell rules keep --w-rail/--w-sidebar alive.

The script also gains an isMain guard and exports its pure pieces for the new scripts/check-dead-css.test.mjs (fixtures cover both acceptance scenarios: prefix-collision and token-read-only-by-dead-rules). Wired into test:scripts.

Deletions (all zero-consumer under the fixed scanner, re-verified individually)

• maka-tokens.css: .maka-shell, .maka-shell-rail-right, the .maka-sidebar header/section/row/button family, .maka-titlebar (+strong), .maka-chat, .maka-composer-toolbar, .maka-shimmer + its @keyframes (folded into functional-motion.stories.tsx, its only consumer), and --w-rail/--w-sidebar. .maka-main, .maka-turn, .maka-code stay — they are live.
• Two prefix-collision finds the original audit missed: .maka-plan-card (plan-reminders.css; its "consumer" was maka-plan-card-menu) and .maka-skill-tab (theme-glass.css; its "consumer" was maka-skill-tabs-bar).
• --ease-in-out-strong lost its last consumer with .maka-shimmer, but it is the named movement curve of the motion governance vocabulary (out-strong = feedback, in-out-strong = movement), so it joins RESERVED_SCALE_TOKENS with that rationale instead of being deleted.

Theme build (scripts/build-astryx-theme.mjs)

@astryxdesign/cli 0.2.0 stopped emitting maka.variants.d.ts but its maka.d.ts still opens with a /// <reference> to it (masked by skipLibCheck). Post-processing now strips the dangling line; maka.d.ts regenerated (css/js byte-identical), astryx:theme --check current. Covered by scripts/build-astryx-theme.test.mjs.

Closes #1980.

Verification

• check-dead-css reports the 13 rules + 3 tokens before the deletions, and is clean at baseline 0/0 after; --check passes.
• npm run astryx:theme -- --check: artifacts current.
• npm run typecheck: clean across all workspaces.
• Script tests 19/19 (including the two new files); check-story-annotations clean.
• e2e/main/preload grepped for the deleted class names as locators: none.

Note: dev-app-runtime.test.mjs › ignores an unreadable or wrong-schema environment file fails on my machine both with and without this change (it picks up the developer's real local env file — a pre-existing test-isolation issue, unrelated).

[Astro-Han]

Approve. Two scanner issues to address before merge:

1. Comma-group selector liveness (check-dead-css.mjs:282). .a, .b { ... var(--t) } with only .b live drops the rule wholesale because of .every, so --t is misattributed as dead. No comma-group class rule exists in the token sheet today, but the fix is small: split each selector, run .every per branch, combine with .some.
2. The class scan misses packages/ui/src/styles.css (check-dead-css.mjs:323), which is imported into the product CSS. It currently holds four classes with no consumers: .maka-daily-review-append/copy/quick-run/save (the daily-review panel was simplified in July). The three listed blind spots are closed; this one is not. Add the file to the scan and delete the four, or document why the component-library stylesheet is exempt.
   Also on record, non-blocking: the fixpoint tests lack self-reference, cycle, and diamond cases; the reservation comment for --ease-in-out-strong and the motion-governance comment in maka-tokens.css disagree about the shimmer curve.

[UncertaintyDeterminesYou4ndMe]

All addressed in b0cfb7f, including the non-blocking notes:

1. Comma-group liveness — rule liveness now splits the selector on commas: .every(isClassLive) per branch, .some across branches. Covered by a new test (a comma group with one live branch keeps its token reads, both directions).
2. packages/ui/src/styles.css joins the class scan — and the four dead daily-review button classes are deleted. The wider scan also surfaced .maka-markdown-code-default/-compact, which are runtime-composed (maka-markdown-code-${props.density} in markdown-body.tsx / mermaid-diagram.tsx), so they join DYNAMIC_STYLE_HOOKS with the call sites named. Scoped to the one hand-written sheet rather than all of packages/ui/src — the only other stylesheets that could appear there are generated ones; the exemption rationale is in the UI_STYLE_FILE comment.
3. Fixpoint tests — added self-reference (no loop, no self-revival), cycle (terminates; lives or dies as one unit), and diamond (all paths to the shared base) cases.
4. Comment disagreement — the reservation comment claimed the shimmer sweep uses --ease-linear; the live text shimmer actually animates with bare linear, so the claim is dropped and the comment now restates only what the motion governance comment says (out-strong = feedback, in-out-strong = movement).

check-dead-css --check clean at 0/0, script tests 13/13, biome clean.

[UncertaintyDeterminesYou4ndMe]

The test_headless / test failures on the latest run are inherited from main, not from this PR: #2081 moved workspace-instructions.ts out of apps/desktop/src/main, and the AHE snapshot source ref in packages/headless/src/ahe-target-protocol.ts still points at the old path — main's own last two CI runs (f39210128, 5daa6d191) are red on the same cli.test.js subtest. Fix: #2094 (one-line repoint, verified 11/11 locally).

Everything this PR owns is green on the same run: typecheck (biome/knip), storybook, e2e, test_workspaces, test_runtime_host.

[Astro-Han]

Rebase to the latest main would fix it.