Skip to content

Instantly share code, notes, and snippets.

@Adamulek123
Last active November 1, 2026 17:33
Show Gist options
  • Select an option

  • Save Adamulek123/496c72047eb4ee9685790f6dd50c54bf to your computer and use it in GitHub Desktop.

Select an option

Save Adamulek123/496c72047eb4ee9685790f6dd50c54bf to your computer and use it in GitHub Desktop.
R35 baseline-only checkpoint failure, no performance comparison

R35 partial baseline diagnosis only. Main3fd2b855c5c6b9169cdfe3733b34011c1c5ed4b7, PR00aa49e72570c31bbddf953d9afaa0b422b23959. Six stopped controls passed; first active control failed ordered-checkpoint-skipped. 23 checkpoints planned, 2 accepted, 3 in final rAF ledger. No candidate built or run; no current performance comparison. Historical Windows percentages were not revalidated. No private IDs/content/traces/resources are published here.

Follow-up source audit confirms the legacy asset URL bridge misses the newer long-lived HTML URL path. Required target iframes were absent during the geometry diagnostic. Prior passing stopped controls therefore do not establish the intended loaded-HTML workload. A matched transport/session fixture and mandatory expected-frame/load controls are required before another comparison. No exact hook state or product defect is asserted.

R35 active checkpoint failure: read-only review

The rejection is supported by the saved ledger and is not a demonstrated observer bookkeeping false alarm. It does not prove a visually presented app regression or that the candidate fails. Keep R35 incomplete and unqualified; do not sort/reorder the saved ledger or waive the order gate.

Correct counts

The route declares 23 checkpoints, not 14. The controller accepted two checkpoint visits before stopping. The final saved ledger contains three distinct witnesses. The handoff's “two of fourteen ordered checkpoints” wording is incorrect and should be corrected in any parent-facing/public summary. No frozen evidence was modified for this audit.

What the raw evidence proves

  • The fixture ledger iterates the predeclared checkpoint list in order on each rAF capture, recording each checkpoint only at its first witnessed visible intersection. It checks the mounted DOM row identity against the current list data before recording it. The controller consumes newly appended ledger entries in their saved order before its separately sorted 20 Hz visible-checkpoint fallback.
  • The ledger chronology is row index1799 at1934.6ms, then the HTML row1786 at2022.1ms, then the required intervening message1791 at2028.4ms. These last two are distinct rAF captures, not two entries accidentally reversed within one capture.
  • The HTML row witness has bounds -545 to952 relative to the viewport (height1497), whereas the subsequently witnessed message has bounds -23 to57.25. The rows retain matching current DOM identities. Thus the stricter “first witness must follow route order” contract was genuinely violated by recorded DOM geometry.
  • Observation2 has scrollTop179070 and total scrollHeight218303; observation3 has scrollTop180269 and total height220002. Those are increases of1199px and1699px while the recorded input sends in this interval are negative/upward wheel deltas. Geometry/scroll offset therefore changed beyond the commanded wheel direction. This is consistent with content measurement/reflow and anchor correction, but the data does not identify which callback caused it.
  • At observation3, the 20 Hz snapshot also sees the intervening checkpoint1791. However, the earlier rAF ledger already witnessed1786 first. The controller appends ordered-checkpoint-skipped, preserves that failure, and later can accept1791 without erasing the rejection. That behavior matches its source; it is not loss of1791 due solely to slow polling.
  • No recorded identity/resource/trace/focus failure establishes a missing resource or corrupted row. A complete document/resource flag does not independently establish stable iframe height. Checkpoint visibility based on DOM rectangles does not prove that exact intermediate geometry reached the display; no synchronized presented-frame image was saved.

Limits and disposition

The evidence supports an active-fixture geometry/order instability. It does not prove that streaming caused it, that the HTML resource itself resized, that a real client visibly jumped, or that the PR changes the behavior. It also does not establish the candidate's speed, memory cost or functional regression. Six stopped baseline controls passed, but the predeclared active gate failed before any candidate; those stopped observations cannot supply a main/candidate comparison.

An honest maintainer comment can be posted now: corrected coverage oracle passed all six stopped controls; active baseline failed ordered content traversal; no current performance comparison or qualification is available. Do not report this as a PR defect or close the PR on this evidence alone. The only demonstrated oracle false alarm so far was the separately diagnosed2px metadata margin, already addressed in R35.

Smallest useful further diagnostic

If further measurement work is authorized, make one separate baseline-only active diagnostic, preserving R35 and its rejection. Capture the existing route without changing its inputs or order gate. Around the first out-of-order witness, record every rAF's bounds/mount identity for the expected1791 and observed1786 rows, scrollTop/scrollHeight/buffer state, actual input send/ACK times, stream revision/commit times, and HTML iframe/container load/resize/height milestones. Stop at the first anomaly after saving the structural evidence and cleanup. Keep all private data local and make no timing claim.

This can distinguish late resource-height measurement, list anchoring/recycling and a ledger placement problem. If a controlled follow-up needs a no-stream or no-input arm, declare that new diagnostic comparison before running it; do not mix it into the frozen performance cohort. A fix, fixture correction or redesigned traversal would then need a fresh baseline proof before a candidate. No unchanged retries until one happens to pass.

Read-only source/raw inspection only. No tests, execution, build, install, app/browser/server, source/fixture edits, retries, publication or subagents were performed.

HTML transport fixture audit (read only)

Conclusion: the legacy URL adapter does not cover the current HTML URL path. It is a harness gap shared by the two source roles, not evidence of a product defect or PR regression. No runtime, install, build, server/browser, test, source edit or further delegation was performed. Only this report was written outside the worktree.

Exact source evidence

  • maintainer-r35/build-current-r35.mjs loads exactly one controlled asset bridge: its Vite load hook replaces apps/web/src/assets/assetUrls.ts. It separately replaces the external fixture, retains actual product modules and has no second long-lived-assets/session/connection override. The bridgeHits === 1 assertion proves this old module was replaced, not that every HTML consumer uses it.
  • render-r4/asset-bridge.ts supplies useAssetUrlState, useAssetUrlRefresh, useAssetUrls and exports resolveAssetUrl. It maps attachment resources to the controlled private-resource origin. It does not supply useLongLivedAssetUrl, assetEnvironment, a prepared connection or an environment RPC session.
  • Current HtmlRenderFrame.tsx imports useLongLivedAssetUrl from assets/longLivedAssetUrls.ts and calls it before setting its local loaded URL. The old bridge affects its refresh hook, but that refresh is on the user-operated fullscreen button, not its initial inline document adoption.
  • useLongLivedAssetUrl uses the real makeLongLivedAssetUrlFamily(assetEnvironment.createUrl, ...), filesystem-access hook, and usePreparedConnection. It never calls the bridged useAssetUrlState/useAssetUrls.
  • state/assets.ts constructs assetEnvironment using the production connectionAtomRuntime. createAssetEnvironmentAtoms issues the real WS_METHODS.assetsCreateUrl request through the environment registry/supervisor. createEnvironmentQueryAtomFamily also observes connection/session state. state/session.ts derives prepared connection from the supervisor's prepared stream.
  • The external R35 fixture renders RouterProvider/MessagesTimeline directly and supplies environment-local, without an explicit test registry/runtime/prepared-session initialization. No controlled success path for the new query is supplied in the reviewed builder/fixture. There may be production defaults, but those are not the controlled attachment transport.
  • assetUrlStateFromResult returns Loading if the prepared HTTP base is null, even if a query otherwise succeeds. Attachment resources are not blocked by filesystem-read scope in blockedAssetUrlState; the access hook is called, but an attachment is not proof of a permission failure.
  • HtmlRenderFrame reserves its calculated height while loaded remains null and only renders HtmlRenderDocument when a URL is adopted. HtmlRenderDocument renders an inline iframe; it does not portal that frame elsewhere. Thus a tall HTML row is not proof that a document loaded.

The relevant product sources are byte-identical in frozen main 3fd2b855c5c6b9169cdfe3733b34011c1c5ed4b7 and candidate 00aa49e72570c31bbddf953d9afaa0b422b23959:

Path Git blob in both roles
HtmlRenderFrame.tsx 93764adecfeb32aa79eca46b6fc5a65827558ab2
assets/longLivedAssetUrls.ts 2643e44751be23b96fad320e51c515ab69b00d51
state/assets.ts 08c3e3a9640759b4d99335a388edffca4cf9becb
client-runtime/state/assets.ts 7017046f8bea8294a7189e48b767051eeafcbb9a

What this establishes and does not establish

The active diagnostic explicitly found no iframe inside the selected tall HTML root across its saved captures. Combined with the source path, that is strong evidence the intended loaded-HTML workload was not exercised for that checkpoint. It does not establish the hook's exact Loading/Failure value, all queries made, all other HTML rows' frame state, or why the geometry/anchor correction occurred. Do not claim a permission failure, asset-RPC error, stable iframe height, or iframe-resize cause without further observation.

The six stopped controls passed the implemented gates. They do not establish a resource-complete current workload: mounted-resource checks enumerate existing img/iframe nodes, so an absent required frame contributes nothing and can pass vacuously. Likewise the list/root/row identity sentinel can pass an empty reserved HTML container. Those controls may describe the actual rendered partial workload but cannot support the intended real-HTML scrolling comparison. The pipeline is already unqualified; this adds a separate fixture-completeness reason. Do not retroactively invalidate a historical cohort whose source still used the old bridged path; audit its exact source separately.

Smallest defensible fixture repair design

Use one shared transport/session fixture for both current roles. Leave HtmlRenderFrame, HtmlRenderDocument, useLongLivedAssetUrl, its atom family/refresh/adoption/cache logic, filesystem-access logic and prepared-connection hooks unmodified. Do not return a fabricated Success from the whole hook or copy an older renderer.

At the existing external fixture/bootstrap boundary, supply an isolated atom registry and a test connection runtime whose environment-registry/supervisor/session represent only environment-local and the controlled loopback resource origin. Route the production assetsCreateUrl RPC method to a strict typed adapter returning schema-valid {relativeUrl, expiresAt} for exactly the frozen attachment map. Preserve query atom creation and actual long-lived-family subscription behavior. Expose production prepared/session/auth streams through this runtime, so the existing hooks derive their ordinary values rather than being replaced. If needed, a strict controlled HTTP client can answer the session-state request; it must reject all unexpected endpoints. Use an explicitly authenticated fixture scope appropriate to the resource, while keeping real permission-resolution code intact.

The narrow interception point is the connection runtime/layer/transport used by the external fixture, not the presentation hook. Implement it using actual service interfaces and existing test fixtures where available. This is a new controlled environment, so typecheck it and prove method/response/schema/connection/permissions counters in an excluded baseline smoke first. A fake runtime that silently returns empty data for arbitrary calls is unacceptable. If a small typed runtime cannot be supplied without pervasive mocking, use one isolated real T3 asset server/session instead and disclose the broader setup boundary; do not bypass the hook to save implementation effort.

Count attempted RPC/HTTP requests before creating responses, keep cache hits distinguishable from fresh mints, and fail unknown resources/methods rather than manufacturing fallback. No live state, provider or paid call is needed. Keep the old bridge for its remaining consumers or remove it only after auditing their paths; both roles must receive exactly the same fixture, source/config hashes and cache states.

Non-vacuous rendering oracles

Before candidate timing, derive the intended HTML checkpoint set from frozen row/resource metadata and opaque local identities. Observe that each intended visible HTML row constructs a production iframe with its expected controlled URL; record per-frame navigation/load completion and the corresponding successful response or cache reuse. For opaque sandboxed frames, do not require contentDocument access or change sandbox: use iframe load events plus CDP/Playwright frame-navigation/loading observations and actual network outcomes. Preserve the real onLoad and height-message behavior.

Require explicit presence/loaded evidence for the intended HTML documents before declaring the workload resource-complete. Define the exact observation deadline/phase before timing, separating first-mounted loading from settled loaded content rather than assuming every iframe must already be loaded on its first visible rAF. Retain first/cold and repeat/warm states; do not prewarm all documents and relabel that as cold. Count every missing/loading/failed expected frame even when zero frame elements exist. Check expected URL identities and schema-valid query results, not just aggregate iframe count.

Add negative fixture controls: disable the URL adapter, return an unknown/mismatched attachment, withhold a frame, fail a document response, and never dispatch load. Each must fail workload completeness. Positive smoke must show actual successful production frames and their height/lifecycle behavior on main before any candidate. None of these controls should change checkpoint ordering, native focus, input, resource, trace, memory or numeric performance gates.

This repair may alter the active reflow sequence simply by restoring missing document content. Therefore classify the new baseline after the repair; do not tune anchoring or waive its ordered-checkpoint rule based on the incomplete R35 workload. A new cohort is required, with all previous evidence preserved and no historical sample pooling. No further benchmark is authorized by this read-only audit itself.

{
"plannedCheckpoints": 23,
"acceptedCheckpointOrdinals": [
0,
1
],
"finalLedgerVisitCount": 3,
"firstFailure": "ordered-checkpoint-skipped",
"earlyGeometry": [
{
"sequence": 0,
"revision": 0,
"scrollTop": 179562,
"scrollHeight": 218545,
"firstIndex": 1799,
"lastIndex": 1811,
"mountedDomMismatches": 0
},
{
"sequence": 1,
"revision": 1,
"scrollTop": 179562,
"scrollHeight": 218545,
"firstIndex": 1799,
"lastIndex": 1811,
"mountedDomMismatches": 0
},
{
"sequence": 2,
"revision": 1,
"scrollTop": 179070,
"scrollHeight": 218303,
"firstIndex": 1794,
"lastIndex": 1807,
"mountedDomMismatches": 0
},
{
"sequence": 3,
"revision": 2,
"scrollTop": 180269,
"scrollHeight": 220002,
"firstIndex": 1787,
"lastIndex": 1799,
"mountedDomMismatches": 0
},
{
"sequence": 4,
"revision": 3,
"scrollTop": 180186,
"scrollHeight": 220002,
"firstIndex": 1787,
"lastIndex": 1797,
"mountedDomMismatches": 0
}
],
"scope": "Rejected active baseline diagnostic only, no candidate or performance comparison"
}
{
"active-checkpoint-review.md": "358975fc381d2cd490badad9627807798cbbe1623ef1d0cad696c43206fdcd8c",
"html-transport-fixture-review.md": "e63dfb0daaffb5972fc0076c17d354dc002b5ebfadd05d899338c8770b76baf9",
"r35-parent-findings.json": "e2b257e22e401606f2ac6862e81ddffaf25313443b023bec90ab04d634e165ee",
"README.md": "538589d6170c5abe9638dfcad9c58e20c7f8f8e69a6f410ae1cea4126e29cdf5"
}
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment