General engineering habits
46 lessons. Most-repeated first.
- CARTO basemap tiles now need an API key. Plain OSM tiles work (with attribution).
- Read the mode’s known false alarms before “fixing”.
- Image_part was removed (now shot_parts); only internal, no caller broke (correction from the human)
- Bisecting by swapping single files is confounded when the new files import each other (missing export kills the module). Instrument the event path (capture listeners for mousedown/mouseup/auxclick) instead.
- Desktop grid: put the hint in its own row under the cat (rows
auto 1fr auto auto 1fr) so cat+hint stay one block and the side column centres against it.
- Calling a
const arrow function (openSheet) before its declaration throws a TDZ error that only shows on the first run; declare UI helpers before use.
- Guide or tutorial art reads better as real renders of the product (world.photo()) than as separate SVG drawings, and costs about 16 KB per WebP.
- Flex rows with ::before separators put dots at the start of wrapped lines; use inline text joined with “ · “ instead.
- Duplicate
const names across appended code break a whole ES module (MONTH_NAMES vs MONTH_TITLES).
- User overrides queue step: never end with Next/queued items; finish all sync work in the same run (correction from the human)
- Close cheap limitations instead of only listing them (correction from the human)
- Several shading bands on a thin limb read as a second tube. Draw anatomy with the silhouette itself (wide shoulder root, tapered wrist) and use one crescent of shade.
- Bot scores swing ~1 point between identical resamples; pool 2-3 samples. Feature gains from bot complaints beat cosmetic tweaks. (correction from the human)
- Never reuse a short class name (.pop) for a new state; it silently inherited position:absolute and threw the chip over text.
- MakeColumn title-cases names, so any localization lookup on column names must be case-insensitive.
goto to a same-page hash link waits for a ‘load’ that never comes: go via about:blank first.- User wants: no asking, no excuses, keep going until queue empty, say why it stopped (correction from the human)
- User wants limits tackled not explained; final reply must have no queued items (correction from the human)
- User: strictly keep going even if bigger, until no queued or next items (correction from the human)
- A 9x8 dHash is a framing hash, not a scene hash: 2-3 % drift breaks it. For burst duplicates use aligned normalised correlation on a small thumbnail. Simulate bursts as equal-size crops at offsets, not crop-vs-full-frame (that is a zoom and gives false low scores).
- Kernel: a restarted screen driver must not trust the previous driver’s hardware cursor. A line of exactly 80 characters followed by a newline shows a blank line on an 80-column text screen.
- Async thumbnail generators that share one viewer race each other; render thumbnails sequentially (or on a separate renderer) and use static thumbnails under SwiftShader.
- QEMU: with
wait=on connect the serial socket before QMP. -d int adds about 100 µs per trap and writes hundreds of MB; never time with it.
- App: measure a Tk settings window against the work area before adding help rows; when a section outgrows the window, give it its own tab.
- Bench: “one pass instead of two” did not halve the unpack time (91 s -> 52 s and 377 s -> 235 s); measure before writing “twice as fast”.
- App: work started at a key press belongs on its own thread, and state another thread can read must be complete before it is published.
- App: a wrapped Tk label changes the window width with its text; pin the width with an empty frame as wide as the wrap length.
- Speech: a character count is not an accuracy measure: a Japanese text got shorter while its error fell; compare error rates.
- Bench: a comparison column must use the same clips on both sides; return “not measured” unless every clip is found.
- Bench: a “processor busy” figure sampled while your own code decodes includes your own load; say so in the column.
- Run: a batch of edits behind a failing first statement makes no edits; put the risky call last or in its own call.
- Words: a one-phoneme tolerance on real words still rewrites everyday speech; only exact homophones are safe.
- Batch-relative thresholds: use an upper percentile, not the median, so bad batches cannot lower the bar.
- Kernel: a lookup “by name” in a task table must skip ended tasks that still wait to be noticed.
- Bench: a diagnostic that loops over recordings must skip an empty one before it concatenates.
- Bench: busy loops started by a benchmark must end by themselves.
- (inti r4) A positional mutant sample is redrawn whenever the list of places changes, so earlier scores and the “held-out” label do not carry over. Count the overlap with the earlier samples instead of guessing why a score moved.
- (inti r4) Start every detached job with the job environment; without it QEMU is not found and the run says nothing about the tree. Mark such a run invalid, do not count it.
- (kilat r1) rustup’s toolchain list changed its default marker; parse the active toolchain from
rustup show active-toolchain instead.
- (inti r4) A usage figure read in the middle of an accounting window is not a measurement; keep the last complete window and read that.
- (inti r4) Two writers on one serial line interleave under host load. Fix the sharing (one writer), not the one symptom that showed.
- (kilat r1) A fuzz harness must prove it can catch a planted crash and a planted hang before its “no failures” means anything.
- (inti r4) Reusing a job name overwrites the earlier log; give reruns new names so the record of a failed run survives.
- (mlbb-blame r6-7) Remember a list of seen hashes, not one, or the second-last item gets offered again.
- PDF text: spans joined without a separator put datum letters next to each other; join with a space before parsing GD&T frames.
- OCP: call
STEPControl_Controller.Init_s() before setting write.step.unit, or the unit setting is ignored.