Engineering Evidence Vault / 2026-06-14
A failure report rendered as an operating surface: the bug, the mistaken assumption, the diagnostics, the fix, and the systemic rule that came out of it.
Executive narrative
The short version: a real bug was fixed, but the valuable part was preserving the reasoning so the next engineer does not have to rediscover it.
The failure story
Why this matters
The bug was fixed. That alone is useful, but it is not the real asset.
The durable asset is the preserved decision path: what failed, what looked plausible, what evidence changed the investigation, why the first repair was incomplete, and which system rule prevents the same failure from returning.
That is workflow memory. The system does not merely record that a fix happened. It preserves enough context for a future operator to understand the tradeoff.
Direct finding
The report's central finding is specific: the first hardening pass correctly reduced image payload by moving public UI to optimized derivatives, but it overcorrected by removing broader preload behavior without replacing it with a controlled warmup path.
The result was a site that was technically lighter while still feeling cold at the exact moments that mattered: near-visible cards and card-to-modal transitions.
Modal theory demo
Each card is intentionally structured as the report's theory in miniature: a compressed state that opens into a full-screen evidence record.
Technical evidence
Open the panels when you want the engineering record: hypotheses, failed fixes, discarded theories, diagnostics, root cause, validation, and interview extraction.
Payload optimization was implemented. Perceived-performance warmup was removed. The validator blocked the old bad preload cache but did not require the new good warmup layer.
The investigation tested whether original full-size rasters were still loading, whether more blocking preloads were needed, whether the warm cache was not filling, and whether the public deployment lacked the final code. Each hypothesis was checked against live HTML, asset inventory, modal image state, and deploy receipts.
The first hardening pass generated optimized derivatives and removed JavaScript preload cache. It solved heavy delivery but missed readiness for card and modal interactions.
Restoring the old broad cache was rejected because it risked preloading too many images and original source rasters. Adding multiple blocking head preloads was also rejected because it could damage first paint on mobile.
The browser evaluator did not reliably expose the page-owned warm cache global. The investigation shifted to observed page asset inventory, live HTML checks, and modal image readiness.
Live HTML contained the warmup code and cache-versioned optimized image references. Optimized assets returned HTTP 200. The final lesson became a publish standard: one hero preload, optimized derivatives, staged warmup, interaction prewarm, and validator enforcement.
Before / After
Broad preload behavior could make later images feel available, but risked eager archive loading and original-raster waste.
Optimized derivatives reduced payload, but removing the broader warm behavior made near-future interactions feel cold.
One blocking hero preload, lazy cards, optimized derivatives, staged idle warmup, interaction prewarm, and validator gates.
Pattern extracted
Workflow memory connection
The report records the state of the problem, the state of the investigation, the state of the fix, and the state of the new rule.
The system does not hide the false starts. It keeps the discarded theories because they explain why the final solution is shaped the way it is.
Future agents, engineers, and interview preparation assets can reuse the same evidence without starting from zero.
Interview value
Source trail
Failure report: Engineering Evidence Vault record, June 14, 2026.
Resolved deploy: WebMnem Liquid Glass `v0.9.124`.
Warmup checkpoint: 20260614T073345Z-webmnem-staged-image-warmup
Case-study checkpoint: 20260614T074740Z-webmnem-image-preload-failure-case-study
