Reduce theme-fetch promise continuation latency during workbench startup

[wieslawsoltes]

Parent handshake issue: #289
Runtime scheduling context: #287 and #293
Workspace acceptance: #280 / #252
Related CSS work: merged #245 and #292

Problem

Unchanged Code OSS remote-workspace startup resolves the built-in light_vs.json theme through fetch(). WebScene's fetch body materialization and promise settlement are fast, but the promise microtask continuation holds the runtime for seconds while applying the theme. A protocol-writer timer that is already due cannot run until the required microtask checkpoint completes, delaying the 1.07 MB remote extension-host initialization send.

Timer arbitration cannot safely preempt this boundary: promise microtasks queued by a task must reach their checkpoint before the next task. The reusable owner is therefore the work performed by the theme-fetch continuation and its DOM/style bindings, not socket wakeup, wire throughput, response-body copying, promise settlement, or cross-source queue selection.

Exact retained evidence

Exact unchanged-Code trace inputs:

• WebScene main base 1bd8596e3dc07a3281204504acad379eb1f6bc78
• local query-gated observer commits a50fd344 and 7ad62971 (not pushed)
• AppScene main package lineage 1420e227e52579a097e7ae52a2cadb41cf073e4d
• exact Node 24.18.1
• unchanged VS Code OSS 645f29cc3176500b4b5762ba887cf2a7f0ffdf2c

The dominant fetch task was:

• URL: built-in extensions/theme-defaults/themes/light_vs.json
• response: status 200, 7,407 body bytes, 10 headers
• total drain_fetch_task: 2457.000 ms
• task selection: 0.000 ms
• ready future retrieval: 0.002 ms
• response object/body/header construction: 0.011 ms
• promise settlement: 0.001 ms
• PerformMicrotaskCheckpoint: 2424.455 ms
• recascade batch finish: 32.530 ms

The corresponding 50 ms protocol timer was overdue 2459.903 ms. Its competing-source histogram was fetch 2457.097 ms, async message 0.034 ms, window message 0.107 ms, V8 platform 0.002 ms. In the same run, protocol-send request 1789663793255 → native WebSocket.send entry 1789663796523 was 3268 ms; native send was 0.677 ms and its microtask 0.716 ms.

A 44,303-byte Seti icon-theme fetch also spent 266.624 ms in its promise microtask checkpoint and 6.368 ms finishing styles, while response preparation and settlement remained below 0.02 ms. This scaling reinforces the downstream continuation attribution.

Evidence: /private/tmp/vscode-252-fetch-phases-7ad62971.log, SHA-256 d1feb5b7970d56298ee80c2a49228dfc3060c4663dec3a2249b09e77c4d98b72.

Scope

First reduce the unchanged theme-fetch continuation into a product-neutral native/browser fixture and profile the synchronous JavaScript, DOM/CSS binding, cascade/layout, microtask, and scene-publication phases. Then optimize the smallest generic WebScene hot path. Preserve browser microtask checkpoint semantics; do not yield halfway through the microtask queue or add a Code OSS/theme special case.

Re-audit merged #292 before editing its DOM-listener/CSS paths. Keep #287/#293 scheduler candidates separate; they are proven generic fairness defects but did not qualify the product gate.

Acceptance

• Query-gated profiling attributes the light_vs.json continuation across JavaScript execution, DOM/CSS bindings, recascade/layout, microtasks, and scene publication without changing product behavior.
• A product-neutral fixture applies representative work from the exact 7,407-byte theme payload for at least 20 cold/warm cycles in Chromium and native WebScene.
• Native continuation p95 is ≤ 250 ms and max ≤ 500 ms, with response construction and settlement individually ≤ 25 ms.
• Promise ordering, run-to-completion, microtask checkpoint semantics, stylesheet/token application, computed styles, theme colors/icons, and error/rejection behavior match Chromium.
• DOM nodes, selector/index work, scene count, task/wakeup counts, V8 heap, RSS, and retained listener/style state remain bounded through navigation and teardown.
• Existing CSS/DOM contracts from Optimize structural CSS invalidation and DOM mutation checkpoints #245/Fix listener ownership and text-topology CSS work #292 and fetch, Worker/MessagePort/WebSocket/FileReader, idle-platform, and scheduler gates pass.
• Unchanged Code OSS retains the exact remote root/provider/five children/five Explorer rows and completes remote Initialized plus first-child paint within Qualify remote workspace bootstrap, Explorer contents, and watcher refresh #252's two-second bound.

[wieslawsoltes]

Query-gated packaged attribution on current main 716aaf829ebe1ea14da35249bad189367e2e6ca2 rules out native DOM/CSS binding time as the owner.

Exact inputs:

• observer commits ff2ec63f, a31c1c52 (local only)
• AppScene package lineage 1420e227e52579a097e7ae52a2cadb41cf073e4d
• unchanged VS Code OSS 645f29cc3176500b4b5762ba887cf2a7f0ffdf2c
• exact Node 24.18.1
• executable SHA-256 30079829e8ab5300669875835fdd8f98bd0e6a775705389185ad52f2e7ad8d05
• WebScene SDK archive SHA-256 4ff5edd68fb49b921f0a286e240a9ff1747baf0aef02882467afbfe221fe0d3b
• log /private/tmp/vscode-297-binding-profile-a31c1c52.log, SHA-256 b063428e7ed432addeba4f71f0acca242de4458d642cf152577f186559880a70

light_vs.json:

• total fetch task: 2477.417 ms
• PerformMicrotaskCheckpoint: 2441.311 ms
• style batch finish: 36.087 ms
• measured native binding work inside the checkpoint: only 0.843 ms total:
  • canvas backing store: 2 calls / 0.003 ms
  • DOM property: 285 calls / 0.271 ms
  • DOM mutation: 93 calls / 0.258 ms
  • DOM query: 3 calls / 0.012 ms
  • style: 120 calls / 0.299 ms

The Seti icon theme showed a distinct native DOM mutation cost (78 calls / 117.037 ms), but its full microtask was 245.751 ms and this is not the 2.44-second light_vs.json owner.

The exact root/provider/five children/five rendered rows remained functionally correct. The two-second gate still failed because provider calls were delayed by the same long checkpoint. Peak measured process-tree RSS was 1,165,936 KiB. The app, guardian, server, extension host, agent host, watcher, and Copilot descendants all terminated, and both observed port 58606 and reserved port 57829 were free.

A third local observer commit, 800912de, adds a one-shot URL-matched V8 CPU sampler. Its reduced fetch control passed and emitted a profile without altering promise ordering. The same package is prepared for one exclusive light_vs.json sample after the active release build is terminal; no runtime behavior fix has been made yet.

[wieslawsoltes]

Qualified fix and unchanged-product evidence

The reusable owner was repeated full publication of mutable stylesheet text from Code OSS theme construction. CSSStyleSheet.insertRule/deleteRule rebuilt and published the entire owning <style> source after every mutation, so the required promise checkpoint spent seconds doing repeated work before the already-due extension-host protocol timer could run.

The focused candidate now keeps the live CSSOM wrapper and synchronous observable semantics, stages the final serialized stylesheet source per owner, and flushes once at an observable style/layout read or checkpoint. It does not interleave timers inside a required microtask checkpoint or special-case Code OSS.

Exact current branch

• current main: 723dd48faffdcea9488503722a6712b6d3c916fb
• candidate: 4e1b376130caee10b2a53c4d64ec53371ea96ae8
• branch: fix/batched-stylesheet-publication-297
• unchanged Code OSS: 645f29cc3176500b4b5762ba887cf2a7f0ffdf2c
• exact Node: 24.18.1
• local query-gated compact observer: vscode-demo 1503022 (not part of the WebScene change)

The candidate was first qualified on exact main b69f931d after CSS metadata PR #313. Main then advanced through parser/resource-interception PR #310 to 723dd48f; its overlaps in webscene_v8_runtime.cpp, DOM core, native test registration, and the component profile were audited and rebased cleanly. Exact-current direct gates still pass:

• 20 × 256-rule mutation p95 6.484 ms, max 8.588 ms;
• one final scene publication, bounded layout work, synchronous sync → microtask → timer order;
• V8 heap 665,496 → 665,496 bytes;
• stylesheet CSSOM identity/parse errors/computed style/matches/layout/disabled/media/disconnect/reconnect/MutationObserver: pass;
• Monaco view-line text topology: pass;
• cascade scaling and 31-navigation teardown gates: pass, heap/listeners flat.

The product-neutral CSSOM contract passes 5/5 subtests in native WebScene and Chrome 153. It covers every observable staged write, error behavior, enable/media changes, checkpoint ordering, topology, and twenty product-shaped cycles.

Unchanged packaged Code OSS

Exact b69f931d package, candidate app SHA-256 a4c1ffe8fae4c91e0b0740ccbbbbc6718eb1d0b6398ed2d9da0dccfd887ab452:

• exact remote root/provider retained;
• five resolved children, five Explorer model children, five rendered rows;
• workspace observer status: passed;
• root resolve: 97.582 ms;
• first Explorer child: 1,487.051 ms;
• remote Ready → Initialized: 759.817 ms;
• 1.07 MB init-data send request → native WebSocket.send entry: 374.573 ms;
• log SHA-256 38944ab3f9408a5390e08a1b834c2bd52886954af2513e0c41f7ec498a389868;
• parsed summary SHA-256 6187e6a95690375dfa79510c84f530472566ea676717f150a7965665a25c5547.

The exact-264 no-candidate control took 5,215.651 ms Ready → Initialized, 3,309.197 ms from init send to native WebSocket entry, 5,277.997 ms for root resolve, and 8,365.675 ms to first Explorer paint. This confirms the stylesheet candidate removes the multi-second product boundary.

Resource attribution and non-claims

The package remains above the 750 MiB release RSS budget: exact-b69 candidate peak/retained process-tree RSS was 1,656,896/1,653,344 KiB. The same-package exact-264 no-candidate control was already 1,640,736/1,633,872 KiB; the candidate delta on the paired exact-264 run was only +0.15% peak and +0.25% retained while reaching the same phase 7.48 seconds sooner. AppScene #42 owns this pre-existing whole-package memory blocker with the per-process A/B evidence.

This fix does not claim WebScene #299's Monaco rendered-text defect. The earlier #299 marker smoke remains separate. It also does not close the broader #252 workspace/provider surface; this trace qualifies only the startup timing boundary exercised here. Every app, guardian, server, extension-host, agent-host, watcher, and Copilot descendant was terminated and each observed port was released after the bounded runs.