toothywalrus

arduino, week 32: the traffic light is real

Thu, 06 Aug 2026 00:00:00 +0000

The kids’ track had a plan and no built hardware. Now it has built hardware.

The traffic light works. Three LEDs, breadboarded and flashed on the real board. It’s the opening project because it is the shortest path from “wires in a bag” to “a thing that visibly does something,” and that distance is the whole game when the audience is eight.

Five programs, one circuit. Rather than a project per lesson, tutorial 02 ships five interchangeable programs that all run on the same wiring — so the build happens once and the payoff repeats. One is a seven-pattern light show. One is calm_light: a single slow drift, no events, nothing to react to. I added it as an afterthought and it’s the one I keep running.

The toolchain question is closed: arduino-cli, not PlatformIO. Written down so it stops getting relitigated every time I open the repo.

Small thing I’m pleased about: I dumped the board’s flash before reflashing it. That is a habit I acquired the expensive way and it costs thirty seconds.

Also this week: the tutorials are now generated from shared sources in both Russian and English with a kids mode, rather than maintained as parallel documents. The vendored kit docs were swapped to the starter kit actually sitting on the desk instead of the one I’d guessed at, which is the sort of error that only shows up when a child asks where a part is. Two ESP32S boards and a second components kit joined the inventory, with their pin gotchas written down before they bite. And the thermometer project moved off a manual nohup onto a proper user service, and now keeps its readings in SQLite with a history view.

Where this came from

Work between 3 and 6 August 2026 in the arduino repository — about 21 commits, plus issues #51 and #56. The repository is private.

The traffic light and the five programs were built and flashed on hardware, not simulated. The ESP32S pin notes are transcribed from datasheets and have not yet been verified on the boards themselves.

game-lab, week 32: the camera started following the ball

Thu, 06 Aug 2026 00:00:00 +0000

metamorph-ball spent the week turning from a physics sandbox into something with a shape.

The route replaced the sandbox. The arena became an actual course, with obstacles wired into both the physics world and the renderer, a run clock, a checkpoint cue and a completion panel. There is now a thing to finish, which is the difference between a demo and a game.

Two new forms. A jelly form that squeezes through a grate, plus rubber and heavy forms with pads that trigger on contact. The grate opens by dropping its collider, which is either elegant or cheating depending on how closely you look.

Juice. Squash-and-stretch and spark bursts, fired on the same cues as the audio so the hit reads as one event instead of three. The sound is synthesised rather than sampled — fewer assets, and it can be tuned per-event instead of per-file.

The change that mattered was four lines. The camera now follows the ball instead of waiting to be dragged. Before that, playing it felt like operating a diagram. Nothing else this week moved the needle as far.

Plumbing: a small persistence package for namespaced, versioned local settings, used first by a quality switch that survives a reload. The asset pipeline exists now, with one model through it end to end. And the deploy key is confined to an rrsync jail, so the thing that can write to the server can only write to one directory.

Where this came from

Work between 3 and 6 August 2026 in the game-lab repository — about 20 commits, plus the decision records on the rrsync jail, the grate collider and the camera. The repository is private.

Everything above runs in the browser build; the deploy-key jail was applied and used, not just documented.

homelab, week 32: provisioning became code, and the fleet moved out

Thu, 06 Aug 2026 00:00:00 +0000

Infrastructure week, of the unglamorous and satisfying kind.

Provisioning is Ansible now — and, more to the point, it has been applied to the actual host rather than written and left aspirational. The distinction matters: half-finished infrastructure-as-code is worse than none, because it tells you the machine is in a state it isn’t.

Monitoring exists, bound to the private network only, with nothing exposed publicly. The agents run natively rather than in containers, which was a deliberate choice and is written down as one. Two things learned in passing and recorded: how much drive health you can see depends on the enclosure the drive is in, not just the drive; and volume identity has to be normalised per platform or you end up with the same disk under two names.

The fleet configuration moved to its own repo. It had grown past being a subdirectory of the edge config — different lifecycle, different blast radius, different reasons to change. Splitting it was overdue.

Routing changes: the Minecraft server now accepts the Java edition alongside Bedrock, and a new subdomain went up behind Basic auth, because “not finished” and “not on the public internet” should be the same state. Also removed a redundant proxy header that had been quietly duplicating what the edge already sets.

Where this came from

Work between 3 and 6 August 2026 in the homelab-infra repository — about 11 commits, plus the decision records on native agents, volume identity and drive visibility, and the split-out of the fleet repo. Issue #14.

Specific hostnames, addresses and topology are deliberately absent. This entry describes what changed, not where anything lives.

mc-server, week 32: Line 1 opens, traps everyone, reopens

Thu, 06 Aug 2026 00:00:00 +0000

The week’s theme was transport. The subway is the first thing on the server that has to work for other people rather than just exist, and that changes what counts as done.

The stack decision came first. Server-side, on Paper’s own minecarts, rather than anything that needs players to install a client mod — the whole point is that someone can join and ride it. And explicitly transport first: the settlement-simulation ideas are a separate thing that can wait, which is now written down specifically so I stop bundling them in.

Then Metro, then Line 1. Surveyed the server-side plugin options (and corrected a wrong claim about TrainCarts I’d already committed to a decision record — that one is on me), adopted Metro, and built Spawn Central to East Terminal as a test line. Getting it onto the web map took route_points, and then a caveat: the map’s depth test makes an underground line look wrong, so the line is drawn above the terrain on purpose. Accuracy traded for legibility, deliberately.

The station ate me. The walls generated at a station edge are sensible for something arriving on rails and a sealed box for something arriving on foot. I walked in and could not get out. Fixed, and recorded as a general lesson — if I generate geometry from config again, a human may approach it from an angle the config never considered. East Terminal also turned out to need a surface exit before it was a station rather than a hole in the ground.

The bake-off has no result. With Metro working, I installed TrainCarts alongside it, because multi-passenger riding is what I actually want and I don’t know which plugin gives it to me. Two bugs so far and both are mine: a spawner sign outside the rail’s range, and signs placed where they could never link to the rail. Nothing decided yet.

Away from the metro: the settlement sim got SQLite persistence and a real tick-loop scheduler instead of whatever it had before, with the persistence choice recorded as stdlib-only and no ORM until something demands one. NPCs will be FancyNpcs rather than Citizens. And the Bedrock bridge turns out to go stale between restarts, so there’s now a refresh script that runs at 4am and skips it if anyone is actually playing.

Where this came from

Work between 3 and 6 August 2026 in the mc-server repository — about 23 commits, plus decision records 0011–0017 written in the same window. The repository is private; this is a summary of what’s recorded there.

The TrainCarts comparison is reported as unfinished because it is. Nothing above describes a configuration that wasn’t actually applied to the server.

moneybtn2, week 32: three merges and a numbering collision

Thu, 06 Aug 2026 00:00:00 +0000

A merge-heavy week: three branches that had been open for a while all landed.

What shipped. Candle storage and the charts that read from it; event overlays on those charts; and a detail popup in the ops view. Each of those had been sitting on its own branch long enough that merging them was most of the work.

One thing worth writing down came out of the chart work: the two data sources feeding the chart are independent of each other. That sounds obvious stated plainly, and it was not obvious from the code — it’s exactly the sort of fact that gets rediscovered painfully in six months, so it’s a decision record now.

And the comedy. Two branches both created decision record 0089. Fine, renumber: 0107 and 0108. Except 0089 had also been claimed on main while those branches were open, so the third one became 0109 — followed by a commit to index it and fix a stale reference to its old number that had survived the first two renames.

Three commits of pure bookkeeping to land one document about something unrelated. Sequentially-numbered append-only records are a good convention until you have parallel branches, and I don’t have a real fix yet. Allocating the number at merge time rather than at write time would work, but it means the filename changes late, which is its own annoyance.

Also: the deploy script does not do two of the steps required to actually deploy, and now says so out loud in the docs rather than living in my head.

Where this came from

Work between 3 and 6 August 2026 in the moneybtn2 repository: pull requests #122, #126 and #127, and decision records 0107–0109. The repository is private.

The two undocumented deploy steps are described as documented, not as automated — they are still manual.

sdr, week 32: a bench baseline, taken before the antenna

Thu, 06 Aug 2026 00:00:00 +0000

Small week here, but it’s the start of something I want to do properly.

The receiver got measured with no antenna attached. This sounds like a non-experiment and it’s the opposite: it’s the only way to get a baseline. Once an antenna is on there, every subsequent number is the radio and the antenna and the environment, and there is no way to separate them retroactively. So: characterise the bare device first, with tooling that can re-run the same measurements and verify them rather than a one-off session and a screenshot.

That verification tooling is the actual deliverable. A number I can’t reproduce next month isn’t a measurement, it’s an anecdote.

Also written up this week: the satellite and drone reception projects, as a plan rather than a result. Nothing received yet.

The write-up of the bench characterisation is going to be this blog’s first real article, rather than a logbook line — there are numbers and a method and it deserves more than a paragraph.

Where this came from

Three commits in the sdr repository, 5–6 August 2026: the initial no-antenna bench characterisation with its verification tooling, and the satellite/drone project write-up. The repository is private.

The no-antenna measurements were run. The satellite and drone work is described as planned, because nothing has been received yet.

stockflow, week 32: born Monday, mostly cancelled Thursday

Thu, 06 Aug 2026 00:00:00 +0000

New repo on Monday. By Thursday most of it was argued away.

The idea was a tool to manage stock-footage contributions across platforms: analytics on what earns, then automation on top of that.

The problem is that there is no contributor earnings API. Not rate-limited, not partner-gated, not undocumented-but-reachable — absent. The numbers I wanted to analyse are visible to a human in a dashboard and nowhere else. Everything I’d sketched on top of that assumption goes with it.

What survives is smaller and less interesting to describe: embedded metadata and delivery. That part is real work with a real result, it just isn’t the product I opened the repo to build.

The honest version: this is four days spent to avoid four weeks, and I do believe that’s the right trade. It still reads as a loss in a week’s log, and I’d rather write it down that way than dress it up. The repo stays, at a tenth of its intended scope.

Where this came from

Work between 3 and 6 August 2026 in the stockflow repository — the initial scoping commit, the platform data-access research, and the build-vs-buy case against the planned scope. Decision records 0001–0005. The repository is private.

The “no earnings API” conclusion comes from reading platform documentation, not from attempting and failing to authenticate against one. If a private or partner-only endpoint exists, I did not find it and did not have access to test for it.

Week 32: a subway that ate its passengers, and a project I talked myself out of

Thu, 06 Aug 2026 00:00:00 +0000

Short week — Monday to Thursday, because this is the first one of these and I wanted to see what it looked like before the week was over. Eleven repos touched. The two things worth reading about are both failures.

This is the index: a paragraph per project, and a link where there’s more.

The metro opened, trapped everyone, reopened

The Minecraft server got a subway — server-side on Paper minecarts, so nobody has to install anything to ride it. I picked a plugin, built Line 1 as a test, got it rendering on the web map, then walked into a station and could not get out. The walls a station generates are perfectly sensible for something arriving on rails and a sealed box for something arriving on foot.

Fixed, and written down as a lesson with teeth: if I generate geometry from config again, a human may approach it from an angle the config never considered.

→ mc-server, week 32 — the plugin choice, the web-map caveat, and a bake-off with no verdict yet.

StockFlow, scoped down to almost nothing

New repo on Monday. By Thursday I’d written the case against most of it.

There is no contributor earnings API — not rate-limited, not partner-gated, absent. The numbers I wanted to analyse are visible to a human in a dashboard and nowhere else, and everything I’d sketched on top of that assumption goes with it. Four days spent to avoid four weeks, which is the right trade and still doesn’t feel like a win.

→ stockflow, week 32 — what survives, and what I actually checked before concluding that.

The traffic light is real

The kids’ electronics track has its first finished project on actual hardware. Three LEDs, then five interchangeable programs on the same circuit so the wiring gets built once and the payoff repeats. One is a seven-pattern light show. One is deliberately calm — a single slow drift, nothing to react to — and it’s the one I keep running.

→ arduino, week 32 — the toolchain decision, and why the tutorials are generated rather than written twice.

Four lines that changed how a game feels

metamorph-ball gained a jelly form that squeezes through a grate, squash-and-stretch and spark bursts fired on the same cues as its audio, and synthesised sound instead of samples. None of that is the change that mattered. The camera now follows the ball instead of waiting to be dragged, and before that, playing it felt like operating a diagram.

→ game-lab, week 32 — the route, the forms, and a deploy key that can only write to one directory.

The homelab grew a control plane

Host provisioning is Ansible now and has actually been applied, not just written — a distinction that matters, because half-finished infrastructure-as-code is worse than none. It tells you the machine is in a state it isn’t. Monitoring went up, bound to the private network only, and the fleet configuration moved out into its own repo.

→ homelab, week 32 — including two things learned about drives that I’d have preferred to learn earlier.

A baseline taken before the antenna

The SDR receiver got characterised with nothing connected to it. That sounds like a non-experiment and it’s the opposite — once an antenna is attached, every number is the radio and the antenna and the room, with no way to separate them afterwards.

→ sdr, week 32 — and the full write-up is going to be this blog’s first proper article.

Everything else, briefly

ToothyGram stopped presenting itself as an official Telegram product, which it never should have. Otherwise: adopted a new upstream base, mapped where a fork can attach to it, and proved the cheapest attachment point costs exactly one modified upstream file. There’s a divergence log now, and a sync runbook, because the alternative is rediscovering all of it in six months.

The money project landed three long-lived branches — candle storage, chart event overlays, an ops detail popup. → moneybtn2, week 32.

Two repos were born: this blog, and the content-studio one that is deliberately not this blog. Both are still a README and an argument. The conventions repo gained one rule, which is that a repo must have a description at creation time — because nothing populates that field automatically and you notice eleven blank ones at once, months later.

Nothing happened on the academy project. One commit, recording a test result on the Mac and listing the iPad checks still to run. It’s on the list and it didn’t get attention, because the metro and the traffic light did.

The bit that made me laugh

Two branches on the money project both created decision record 0089. Renumbering to 0107 and 0108 fixed it — except 0089 had also been claimed on main in the meantime, so the survivor became 0109, followed by a commit to index it and fix a stale reference that had survived the first two renames.

Three commits of bookkeeping to land one document about something else entirely. Sequentially-numbered append-only records are a good convention right up to the point where you have parallel branches, and I don’t have a fix beyond noticing early.

Where this came from

Assembled from work between Monday 3 and Thursday 6 August 2026: roughly 90 commits across eleven repositories, plus the decision records written in that window. The repositories are private, so there is nothing to link to here — what’s above summarises things that are recorded, not recollected.

Load-bearing sources: mc-server 0015–0017 and its walkable-tunnel and web-map notes; stockflow’s build-vs-buy case and platform data-access research; arduino’s toolchain decision and tutorial 02; game-lab’s metamorph-ball iteration C; homelab-infra 0015–0018; moneybtn2 0107–0109.

Nothing above describes something that wasn’t actually run or committed. The TrainCarts bake-off has no result yet and is reported as unfinished; the satellite and drone SDR work is a plan, not a result.