# loading-performance delivery audit

final source head: `bdea5b417e97714189545b558771160513024ba0`.
upstream pr: https://github.com/pingdotgg/t3code/pull/18594.
readiness: blocked by upstream conflicts. no merge, rebase or force-push was performed during recovery. publication uses maria's existing fork branch. no new test files were created.

## measurements retained

the nine-scenario matrix was reused, not rerun after service recovery. its fixed baseline is `83fbbc18f8718b0254e696f02b8e381386fd6c7b`; measured candidate is `0b455cc93ae9aca3aed83b05b039e102e38648f7`; phase is `integration0b455b`.

112 retained measurements across nine pairs contain zero failures. all nine message-latency medians and telemetry-inclusive payload medians decrease. first-visit telemetry increases 63.6%; remote-refresh telemetry increases 2.8%. the two earlier-history samples compare different-sized pages and are descriptive.

production web builds and source servers used a transparent http/websocket proxy, matched stored transcripts of 6,000 and 200 items, and 0/80ms added rtt without a bandwidth cap. the latency metric ends at a painted visible message and excludes pairing and provider execution. sequential shared-host samples and their iqr do not establish a randomized experiment or confidence interval. the full connection-mode scope is not empirically timed.

the html's embedded report and payload arrays equal the preserved json artifacts. the delivery revision adds only a visible commit-provenance note and narrow-screen wrapping; measurements are unchanged. desktop728/light and mobile360/dark previews were inspected. latency, total payload and telemetry controls each render nine finite-valued scenarios. mobile document width equals360 with no horizontal overflow. there were no browser errors in graph previews.

uploaded html equals the local document byte-for-byte, sha256 `82077328ab30e30679f21a99b8c0a0dfcb629769f35b837d30174c43a1bd2a7c`.

https://uploads-production-47e4.up.railway.app/files/99d4d7d0-36d6-4968-b429-5b7a07ab1afe/integration-report.html

## checks retained and completed

measured0b455: 91 direct transport checks passed, including canonical histories, compact restoration, history pages, cache/auth isolation, br/gzip/identity, vary and cursor rejection. saved blacksmith results contain911 passing tests in26 existing files, web/server/client-runtime/contracts typechecks, targeted lint with0 errors/248 warnings and a production web build.

review547: web typecheck,159 tests in7 existing files, lint5 files with0 errors/100 warnings, and production build passed on blacksmith. the initial wrong-cwd test invocation exited before tests; its corrected run supersedes it. actual browser indexeddb verified small save/load, oversized replacement deletion, other-environment isolation and subsequent small save/load.

finalbdea: web typecheck,100 tests in3 existing files, lint3 files with0 errors/6 warnings and production build52.13s passed on a dedicated blacksmith worker. the amended existing streaming check observes newly appended visible text before highlighting runs and retained code identity. no test/typecheck/lint/build ran locally. the task-only ssh adapter retained the installed helper's source synchronization and bypassed its failing multiplexed transport.

downloaded final build archive sha256 matches the remote result: `1c5df26efc40417666d2c0a915f976d6ce2c86f77926137a42e43a8d5461487e`. the actual client loads `assets/index-BafQaojr.js`.

two independent gpt-6.1-sol reviewers source-approved this exact finalbdea head. they reassessed the later suffix fix, complete scoped diff and nearby consumers. their source approvals do not imply verified native clients or a merge-ready pr.

## actual runtime and evidence

measured0b455 actual codex replies passed forced socket close/reconnect, exactly-once replay, normal acknowledgment/composer clearing,36 incremental plain-response updates, completion and refresh persistence. terminal printf, palette filtering, thread details, custom snooze cancellation, sidebar scope/navigation, file tree/source preview and edit-from-here cancellation were exercised. the observed diff hunk-count warning uses unchanged baseline normalization and a byte-identical parser.

review547 actual codex reply rendered160 requested code lines exactly once, cleared the composer and survived refresh. it arrived as one highlighted update; that run does not prove live incremental highlighted-line retention. direct empty-thread navigation, reload and browser back passed. a provider-status timeout warning mounted the composer before the model-picker shortcut, leaving pre-mount timing unverified.

finalbdea actual codex reply rendered600 ordered comment lines, cleared the composer after acknowledgment and remained exactly once with the same provider message id after refresh. both parent and independent child browser observers saw one bulk update. live incremental highlighting retention and the pending suffix remain unverified through that provider path; the amended existing component test verifies those render states directly. parent inspected the full-size screenshot. its short recording is a19-second original-speed excerpt of arrival/completion, excluding provider thinking time; it is not a loading benchmark.

matching before/after loading images and recordings use1024×800 transcript/control crops from1280×800. these recordings are descriptive, separate from the benchmark matrix: before7.83s, after6.11s, with5.86s candidate long-task time under recording load. original-speed loading and the baseline's actual final-frame hold are disclosed in the pr. the exported media were inspected, uploaded with file-upload, and played through plus seeking.

the two pr comparison images load at natural1024×800 and the final reply image at1280×800. its four github video attachments now render native players, correcting the earlier image-markdown error. each played to the end without media errors and passed seeking/resume. the final19.75-second uploaded reply clip was also played through and sought in the separate user-facing uploaded player. terminal/reply evidence shows acknowledged reply, cleared composer and actual terminal output.

## limits and publication

required-wizard gating is unverified through the browser: both baseline and candidate source servers omit optional bootstrap creation flags required by the fresh-workspace helper. the synchronous guard is source-reviewed, but the runtime never reached that required-wizard decision. native desktop/mobile, native ime, non-codex turns and real relay/tunnel/ssh timings remain unverified.

seven earlier review threads were answered and resolved; the pre-existing generic boolean traits omission was explained and left out of this loading task. the later pending-code visibility finding was fixed inbdea, covered by the amended existing test, answered with attribution and resolved. no review bot was invoked.

fresh final github snapshot reports headbdea, base `58d253231d83f669a2cdfefa5fe479751d22c582`, zero unresolved review threads and no truncated review/check pages. macroscope correctness passed; coderabbit remains pending. github's rendered page shows eight expected checks awaiting conflict resolution. conflicts are reported in ThreadTerminalDrawer.tsx and pnpm-lock.yaml. github reports conflicting/dirty, so the pr is not green. maria's explicit ban on merge/rebase/force-push remains in effect. no conflicting upstream changes were integrated.

pr diff against its current merge base:112 files,4,273 insertions/1,758 deletions. the final suffix follow-up adds only18 lines and removes7 across3 existing files. repository working tree is clean.

all task-owned local servers/proxies were stopped, including4900/4901/4904 and the later172463 server. both task blacksmith workers are stopped. sandbox state and completed artifacts remain available; unrelated server3773/pid160367 was preserved. task browser tabs were closed at handoff, the pr is linked to this thread and watching is disabled. queued t3 messages were not changed.

audit result: graph provenance, saved matrix and transport results, scoped final verification, source reviews, publication and evidence are delivered. merge readiness remains blocked by conflicts and the pending bot check. all runtime limits above remain explicit; no claim that every connection or client was verified is made.