Measuring Hermes Bytecode Delta Updates: One Added Module Cost 29x More Than One Changed Line
I measured OTA delta sizes against a 2.4MB Hermes bytecode bundle. Changing one line cost 2,104 bytes; adding one module cost 61,197. The widely repeated advice about stable module IDs turned out to matter least.
The 429s Kept Coming After Recovery: Measuring Retry Jitter Across 2,000 Clients
My backend came back but the 429s did not stop. I rebuilt the scenario in a 2,000-client discrete simulator and measured what plain exponential backoff, delay caps, 10% jitter, and honoring Retry-After cost in seconds — then reran everything with server capacity that degrades under load.
Cutting MCP Tools Didn't Make Anything Lighter — 2,377 Bytes of Definitions vs 110,298 Bytes of Response
I wrote a minimal MCP server for a wallpaper catalog and measured the byte cost of tool definitions against the byte cost of responses. Here is which side actually matters, and the real reason to merge tools.
Deciding overlay text legibility at ingest time instead of on device — four metrics measured side by side
Moving the question of whether text stays readable over a wallpaper out of the device and into the content pipeline. Four candidate metrics measured across 240 images, including what downscaled judging actually computes.
A JSON File 19% Smaller That Ships Only 3% Smaller — Measuring Catalog Payloads Against gzip
Five payload formats measured on a 20,000-item wallpaper catalog. Key shortening saves just 3.2% after gzip, while a columnar layout cuts 26.8%. Real numbers for Brotli, paging, and JSON.parse time.
Exactly 20 Characters, Still Rejected: Putting Character Counting Behind One Boundary
Emoji and combining marks make your client and your server disagree about length. I measured four counting methods on Node v22 and found 75.6% of naive truncations split a grapheme and 36.1% produced replacement characters over UTF-8. Here is the shared-module design that fixed it.
When 1.10.0 Gets Locked Out: Measuring Four Version Comparison Approaches for Forced Updates
A remote-config gate for minimum supported version, rebuilt after measuring four version comparison approaches. Includes the cases where localeCompare reports equality, the fail-open boundary, and the incident caused by gating on a build still in review.
When Two-Character Queries Silently Return Nothing: Measuring Japanese Search Indexes in an Expo App
SQLite FTS5's trigram tokenizer returns zero rows for Japanese queries shorter than three characters, without raising anything. I benchmarked linear scan, a bigram inverted index, and FTS5 over a 20,000-item catalog to find the real threshold.
What Renovate may bump in an Expo app, and what it must never touch
Turning on automated dependency updates in a Rork-generated app also hands Renovate the 123 packages Expo SDK 57 pins. Measured on 2026-07-30, six of them sit a full major version behind npm latest. Here is how to generate the ignore list from the SDK instead of maintaining it by hand.
Counting what prebuild --clean will erase before you upgrade to Expo SDK 57
A raw diff between two generated ios/ trees showed 649 changed lines; only 3 were real edits. How to count what prebuild --clean erases, and move it into a config plugin.
When to Raise Your Minimum iOS Version — Count Leftover Branches, Not User Percentages
Judging a minimum OS bump by usage share produces the same answer every year, so the decision never happens. Here is the annotation convention, the sweep script that counts how many iOS and Android branches each candidate floor would retire, and what to watch for 30 days after.
What Filtering Out Prerelease OS Devices Cost Me — Separating Telemetry Tracks Instead of Dropping Them
Filtering prerelease OS devices out of telemetry left me with no warning signs on GA day. Here is the rebuild: an os_track dimension, a remote GA baseline, split alert thresholds, and a release gate that separates blockers from fixes.