[Mohamed Mansour (mohamedmansour)]

Every client mount of a component that uses Link CSS paid for its own copy of the stylesheet: a new <link>, a guard <style> that hides the host, a load wait, a re-read of link.sheet.cssRules, and a separately parsed CSSStyleSheet adopted by that root alone. Router soft navigations, opened lazy regions and any client-created component therefore paid N parses, N guards and N load waits for N instances. Template-owned links already cache their sheets per template; component catalog links had no cross-instance cache.

Approach

Component Link stylesheets are now shared per Document and resolved href, and promoted on their second use. What users see does not change.

• First use stays native. The first newly attached Shadow root to mount a stylesheet keeps its native <link> and only records whether that load is authorized under the existing fix: prevent dynamic component stylesheet flashes #454 rule (unredirected, not served by a service worker, successful). A component used once pays no CSSOM serialization, second parse or duplicate contents.
• Second use promotes once. The next root builds one constructable sheet from the CSSOM of any loaded, authorized link for that href and caches it. Every later root adopts the cached sheet synchronously, before its template is appended, with no link, guard or load wait. This covers a later root in the same pass, a reopened lazy region, and a cold load that settles.
• No visible switch. A guarded root that is still hidden may switch to the shared sheet before it is revealed. A visible root never switches.
• One cascade per component. Adopted sheets cascade after tree styles, so roots whose cascade includes other author styles promote on first use instead: templates with their own <style> or <link>, server-rendered or reused roots, and Light descendants. All instances of such a component adopt the same sheet, and a template-link reset that clears the root can no longer drop a native catalog link.
• Authorization is remembered per Document and href, failures included. Memory-cache hits add no resource timing entry, so a repeated check could not pass.
• Less per-root bookkeeping. One TargetStyleState record per target replaces five WeakMaps, and resource IDs stay plain strings until a second ID arrives. When TemplateElement attached the Shadow root in the current mount, adoption writes adoptedStyleSheets without reading it. A read makes Blink build and retain a wrapper for the root's lifetime. The public installComponentStyles still reads, so sheets an author adopted are kept.
• Smaller wins: URL rewriting is skipped when the CSS has no url(, image-set( or escape, and resolved hrefs are cached per base URL.

Cold loads, unpromotable stylesheets and failures keep the existing native-link, guard and error handling.

Performance

Edge 156 with 4x CPU throttling, measured with local Playwright harnesses that bundle the framework's real installComponentStyles and prepareComponentStyleLinks. Memory is performance.measureUserAgentSpecificMemory() after GC, minus a no-CSS page with the same roots. "This PR" is the path component mounts take.

N instances of one card component (~3 KB of CSS), time to styled and CSS memory, medians of 10 runs:

Rows	main	This PR	One shared sheet (ideal)
20	107 ms, 170 KB	18 ms, 92 KB	8 ms, 7 KB
200	409 ms, 562 KB	32 ms, 117 KB	17 ms, 18 KB
1,000	1,738 ms, 2,227 KB	88 ms, 193 KB	69 ms, 30 KB
Reopen 1,000	473 ms	71 ms	65 ms

200 component types (~1 MB of CSS), 1,199 Shadow roots (a 1,000-item list plus one instance of each other type), medians of 5 runs. No component's rules leak into another's shadow root.

Variant	Time to styled	CSS memory
main	3,099 ms	3,105 KB
This PR	261 ms	602 KB
Native <link> per root	254 ms	613 KB
One sheet per stylesheet (ideal)	226 ms	551 KB

Commerce app, Link CSS. Two 30-sample runs with the server ports swapped, averaged to cancel a measured 5-17 ms per-port bias. Mean +/- standard error:

Metric	main	This PR	Change
Soft nav home -> product: soft FCP and LCP	272.5 ms	257.5 ms	-15.0 +/- 5.2
Soft nav home -> product: main-thread busy	311.4 ms	296.0 ms	-15.4 +/- 3.2
Soft nav product -> product: soft FCP	164.0 ms	160.0 ms	-3.9 +/- 2.8
Full loads of home and product: FCP, LCP, hydration			each within 1 SE

The same crossover with Module CSS shows no change on full loads. The soft-nav gain comes from no longer promoting four single-use components into four separate sheets: all 49 roots now keep native links.

Review notes

• Bundle size: +1.0 KB gzip (+0.9 KB brotli) in the shared framework chunk, about 4% of commerce's main chunk.
• CSS loaded by the server-rendered page gets no gain. Commerce's repeated product cards stay on native links, as on main: their only timing entry predates the mount, so the fix: prevent dynamic component stylesheet flashes #454 rule cannot authorize them. I did not relax that rule.
• Stable responses are assumed. Sharing assumes a stylesheet URL returns the same CSS for the Document's lifetime, which holds for compiler-emitted assets. Shared sheets and authorization results live as long as the Document's style catalog.
• First-use roots remain slightly above native at scale (261 vs 254 ms), within run-to-run variance.
• Direct callers of the public installComponentStyles keep the reading path: 100 ms / 294 KB for 1,000 rows and 266 ms / 747 KB at scale.
• link-styles.ts grows from 1,245 to 1,561 lines.

Tests

• New component-link-sharing e2e fixture (7 tests): one native link then one shared sheet across a mount pass, sharing after the preload is released, second use after a lone first mount, promotion from a later root after the first is gone, guarded cold mounts, unpromotable stylesheets, and eager promotion for components with template styles. The sharing tests fail on main, where every instance gets its own link and sheet.
• component-assets e2e: a single lazy panel now pins {adopted: 0, disabled: false, links: 1} (first use stays native), and a new reopen case pins {adopted: 1, disabled: true, links: 1}.
• Unit tests for sharing, eager promotion, write-only adoption on fresh roots, preserving existing adopted sheets, remembered failed authorization, base URL changes, template author-style detection and the URL-rewrite fast path.
• cargo xtask check passes. Framework e2e: 391 passed. component-assets: 5 of 5. Router: 42 passed (4 pre-existing fixme).

specs/hydration.md and the framework README describe the new Link behavior.

[mddinbox]

Target: aab3350...dbed0e8 | Coverage: complete; CI: PR Checks, Test, Build (Linux/macOS/Windows), Lint, Docs, CodeQL — success
Challenge: 1 confirmed, 0 revised | Compatibility: no API change; memory retention only
Action: COMMENT | The inline finding identifies one medium-severity lifecycle retention issue.