deepgrain-foundry/docs/agent-log.md

# Agent log

What agents have done, newest first. Add an entry for every piece of work: the date, which
note or request it was for, what changed, the files touched, the tests run, and anything left
open or not verified. Keep each entry short; the git history has the details.

## 2026-10-24: t9 Script to run the local server for testing (Claude Code, Codex)

- New `tools/serve.py`: builds, serves `dist/` on 127.0.0.1 (port 8000, or the next free one up
  to 8010), sends `Cache-Control: no-store`, and rebuilds when `src/` or `build.py` changes; a
  failed build is printed and the old page keeps being served.
- `README.md`, `AGENTS.md`, `build.py` docstring, `docs/PROJECT_MAP.md`: point at it.
- Checked by hand: page served with no-store; touching a source file rebuilt it; with the owner's
  old server on 8000 it moved to 8001; Ctrl+C stops it. Headless suites and map check pass.

## 2026-10-24: t11 Welcome and first milestone shown again on an older world (Claude Code, Codex)

- Cause: nothing cleared the discovery cards (`#cards`) when a world was entered, so a new
  world's Welcome and milestone cards stayed up after loading an older world.
- `src/ui/session.js`: `enterGame` clears the cards, floaters, particles and the status line;
  a new world still shows its Welcome card afterwards.
- `test/play.py`: a new world shows Welcome; loading a world from the list shows no cards.
- Tests: `run_all.py --browser` passes (153 `play.py` checks). `t_smooth`'s half-resolution
  check flaked (passed and failed on reruns, with and without this change; the machine was busy).

## 2026-10-24: the to-do list moved from NOTES.md to a kandown board (Claude Code)

Request: initialize kandown (https://kandown.dev, 0.52.0) for the project, move all the notes
into it, and document how it works.

- Practised first in a scratch folder: `init --no-open` starts no daemon and makes no sample tasks
  when `tasks/` exists; `create` prints only the id and names the file `<id>_<slug>.md`; a
  multi-line `report: |` block is read (checked in kandown's parser) and `doctor` accepts it;
  `list --json` doesn't include `report`; project instructions in `.kandown/kandown_work.md`
  appear in `kandown work` (~1,100 tokens).
- A mistake on the way: `kandown init --help` ran `init` in the vault (it ignores `--help`) and
  started a daemon there. Stopped it, and removed the vault's earlier kandown test files as the
  owner asked. `docs/kandown.md` warns about it.
- `kandown init --no-open` in the repo. `.kandown/kandown.json`: default columns, Review's
  instructions say only the owner moves work to Done, tags and priority shown. `.gitignore`:
  `.kandown/kandown.html` (6.5 MB, regenerated).
- 14 notes became tasks t1-t14 (Todo: 6 tasks; Backlog: 5 tagged `discussion`, 3 tagged
  `idea`), each with the owner's note word for word under `## Context` (checked by script: 14 of
  14). `kandown doctor` is clean.
- `.kandown/kandown_work.md` holds the board rules (from AGENTS.md's old NOTES section);
  `docs/kandown.md` explains how kandown works here. `AGENTS.md` (session start, a new "Working
  from the task board" section linking both), `CLAUDE.md`, `README.md`, `docs/PROJECT_MAP.md`,
  `docs/DESIGN.md` (decision). `NOTES.md` removed.
- `kandown list` and `kandown doctor` checked inside Codex's sandbox: they work.
- Open: the web board (`kandown`) wasn't opened here; the owner uses it.

## 2026-10-24: several worlds, with a Worlds list (Claude Code, with Codex)

Note: "support multiple saves". The owner chose a world list.

- `src/ui/saves.js`: each world has an id and its own `<id>:main` / `<id>:auto` saves, plus an
  index record `worlds` (name, seed, size, created, saved, playtime). A save writes the world
  and its index entry in one transaction. `GAME.worldId` is the open world.
- Migration: an old single `main` / `auto` save is copied to "World 1", every byte compared,
  and only then are the old keys deleted (in the same transaction that writes the index). If
  anything fails the old save is kept and a toast says so.
- Title screen: Continue loads the most recently saved world; a Worlds button opens a list
  with Load, Rename (trimmed, at most 40 characters, escaped when shown) and Delete (with a
  confirmation). New and imported worlds get new ids. Settings' "Delete saves" became "Delete
  all worlds" with a confirmation. `shell.html`, `session.js`, `loop.js`, `settings.js`,
  `state.js`, `style.css`, `docs/DESIGN.md`, `docs/PROJECT_MAP.md`.
- `test/play.py` (149 checks): migration of an old save, two worlds listed, load switches,
  rename survives a reload, HTML in a name shows as text, long names cut to 40, delete and
  cancel, Continue picks the newest, import adds a world. All 11 headless suites pass;
  `t_smooth.py` passed in Codex's run and failed its known flaky check in mine.
- Shots: `shots/worlds-{pixel,smooth}.png` and `-900` versions.
- Not committed: a line in `AGENTS.md` about running `kandown work` appeared in the working
  tree during this task; it was not part of the task and is left for the owner.
- Open: nothing.

## 2026-10-24: screw tubes carry bars (Claude Code, with Codex)

Note: "how do you transport bars? belts work but tubes do not, should they?" The owner chose
tubes over a separate bar lift.

- `src/sim/items.js`: a bar whose eight cells are all tube cells is in the tube (derived from
  its position, so saves need nothing new). On the tube beat it steps along the tube, snaps
  into the next block's lane at corners, and leaves a mouth with a launch like grains. A bar
  pushed by a belt into a tube's side or back, falling into a down tube or sitting on an up
  tube's intake enters when the aligned cells are free. Grains and bars wait for each other.
  Gas and liquid don't block a bar (Claude added this after review: smoke near furnaces stalled
  bars in Codex's version).
- Text: `pieces.js` tube description, `notebook.js`, `docs/DESIGN.md`, `docs/PROJECT_MAP.md`.
- Tests: `t_conveyors` (55 checks): belt into a lift and out the top, right tube, up-to-right
  and right-to-down corners, blocked mouth waits without overlap, grains and a bar share a
  lift with sand conserved, save mid-tube, smoke and steam in the tube, melting in a tube,
  bars still ride and stop on belts and rest on top of sideways tubes. All 11 headless suites
  and `play.py` (131) pass; `t_smooth.py`'s half-resolution check failed on the slow fallback
  (known flaky).
- Shots: `tools/scenes/bar-lift.js`, `shots/bar-lift-{300,900}.png`: mold, belt, lift, corner,
  belt; the bars arrive at the top.
- Open: nothing.

## 2026-10-23: screw tubes use all their lanes (Claude Code, with Codex)

Note: "maybe the belts can move faster/carry more than the tubes… the tubes only carry one line of
materials on either side." The owner chose to fill every lane; belt speed is unchanged.

- Measured first (`test/t_conveyors.js`, a belt with a sand heap into the side of a 6-block up
  tube, 3,000 ticks): 1,459 grains out, lanes `[2940, 0, 0, 10]`; the edge lane was full (one
  lane carries at most ~1,500). Spreading only at intake changed little (1,486): belts push
  straight into the edge lane without going through `tubeIntake`.
- `src/sim/machine-tick.js`: on each tube beat a grain in a vertical tube may step sideways
  within its block to a lane with clearly fewer grains ahead (a strict gap stops shuffling).
  `src/sim/conveyors.js`: `tubeSpread`, shared by side intake (far lanes first) and corners.
- After: 2,071 out (92% of the 2,243 the same belt delivers to an open drop), lanes
  `[1937, 1116, 486, 872]`, sand conserved. Checks: all four lanes, conservation, at least 90%
  of the belt's delivery, a side-fed sideways tube fills more than one row. All 11 headless
  suites and `play.py` pass. `docs/DESIGN.md`, `docs/PROJECT_MAP.md`.
- Found on the way: Codex's sandbox (`workspace-write`) blocks network sockets, and headless
  Chrome crashes at startup without one (SIGTRAP core dumps in pairs). Codex runs now pass
  `-c sandbox_workspace_write.network_access=true`.
- Open: nothing.

## 2026-10-23: the build card describes the build tool, not the piece (Claude Code)

Note: "instead of the current info box for the build tool showing the current building selected,
it should just show a description about the build tool in general." The owner kept the R/F chips.

- `src/ui/selection.js`: the build card reads "Build", then the R/F chips, Replace · G and
  ghosts, a line on placing, the build keys and the status line. No piece name or cost.
- `test/play.py`: the card shows "Build" and the placing text, not the piece's name, cost or
  description; the other card checks are unchanged. `play.py` and all 11 headless suites pass.
- Shots: `shots/build-card-general-{pixel,smooth}.png`. `docs/DESIGN.md`.
- Open: nothing.

## 2026-10-23: world sounds are heard around the bot, not the view (Claude Code)

Note: "the sound effect is still heard when the player's zoom level is farther out… sounds
shouldnt be heard based on the viewport, but rather in a radius from the player."

- `src/ui/audio.js` `sfxPlace`: gain from the distance to the bot's centre (full within 24
  grains, silent at 96), pan from the horizontal offset; zoom no longer matters.
- `test/play.py`: the audio checks use points relative to the bot; a new check zooms out to 2
  and plays a visible sizzle 200 grains away: silent (it played before). `play.py` and all 11
  headless suites pass. `docs/DESIGN.md`.
- Open: nothing.

## 2026-10-23: F on an ingot mold does nothing and says nothing (Claude Code)

Note: "now that ingot molds empty themselves, F still makes a message…"

- `src/sim/actions.js` `interactAt` passes ingot molds by, so F reaches anything else nearby;
  `src/ui/actions.js` shows no "Nothing to use there." when the cursor is on an ingot mold.
- Tests: `t_molds` (F on an ingot mold in every state returns nothing, changes nothing, no
  sound) and `play.py` (F over an ingot mold leaves no status and no toast). All 11 headless
  suites and `play.py` pass.
- Open: nothing.

## 2026-10-23: pixel-art silo for the pixel look (Claude Code, with Codex)

Note: "the silo art looks the same for both pixelated graphiocs and smooth."

- `src/render/machines.js`: `drawSilo` is the smooth (illustrated) version as before; in the
  pixel look `drawSiloPixel` draws it grain by grain: a drum shaded by column with rib rows and
  ladder rungs, a cone to the slide gate, legs, the roof hatch with its lid, and the sight glass
  filling in the contents colour. Codex's first version read poorly (dark side columns made
  the drum look narrow); Claude redrew the grain layout.
- `tools/scenes/silo.js` (run after `clear-sky.js`): three silos, empty, half sand, nearly full
  copper ore with the gate open. Shots `shots/silo-*.png` at zoom 6 and 12; smooth shots match
  the before shots pixel for pixel. `docs/DESIGN.md`, `docs/PROJECT_MAP.md`.
- Tests: all 11 headless suites and `play.py` (129) pass. `t_smooth.py`'s half-resolution check
  failed twice on the slow fallback at load ~5 (known flaky; it passed in the previous task).
- Open: nothing.

## 2026-10-23: world sounds are local (Claude Code, with Codex)

Note: "there is a sound effect that seems to happen randomly but constantly… maybe the sound
effects are not local?" The owner then confirmed a new world is quiet.

- Measured headlessly over 5 game minutes: a fresh world emits no sounds; a lava and water
  contact 720+ grains from the bot emitted 2,889 sizzles (about 580 a minute), and `events.js`
  played every one at full volume. The sizzle (`heat.js` boiling and quenching) is the only sim
  sound with a position.
- `src/ui/audio.js`: `sfxPlace(x, y)` gives gain (1 in view, 0 one view width beyond it) and
  pan; `playSfx(name, arg, x, y)` skips silent ones before the rate limit and plays the rest
  through a gain and stereo panner (`AU.bus`). `src/ui/events.js` passes positions.
  `docs/DESIGN.md`, `docs/PROJECT_MAP.md`.
- `test/play.py`: far off-screen is silent, on screen plays, just off-screen is quieter and
  panned. All 11 headless suites, `play.py` and `t_smooth.py` pass.
- Checked afterwards with the owner's exported world (seed 622525): over one minute, 48 sizzles
  and no other sounds, all from water boiling on naturally hot deep rock (a pool on 580°C basalt
  and drips in a shaft and a cave, about 500 grains below the bot). Nothing built there. With
  the fix every one of them is silent from the surface.

## 2026-10-23: fewer, smaller messages (Claude Code, with Codex)

Note: "lets change the engot mold and other status messages. they are way too large and just too
many messages…". The owner chose molds, machine warnings and shorter status lines.

- `src/render/overlays.js`: every mold always shows a small percentage (`moldLabel`), gold
  `100% · F` for a set bit or tank, red when spoiled or when an ingot mold is blocked
  (`mo.blocked`, set in `ejectIngot`, not saved). Details are in the hover panel (`info.js`),
  which also carries the furnace-loading advice that F messages used to append.
- `src/ui/events.js`: no cards or toasts for bars dropping, blocked molds or set castings (the
  sounds stay). Fan heat is a red ! marker on the fan (`mm.hot`, hover shows °C); melted tubes
  and machines destroyed by heat get a short red floating label (`warn` event, coloured
  `spawnFloater`) instead of a toast.
- Status lines shortened to one sentence across `src/sim` (molds, actions, blocks, blueprints,
  drill, items, machines, tools, undo, vacuum) and `src/ui` (actions, selection).
- Tests: `play.py` records every status line and checks none is over 90 characters (longest
  now 55); mold label and no-toast checks. `t_molds` (message lengths, `mo.blocked`), `t_build`
  (hot fan: `mm.hot`, no toast), `t_world`. All 11 headless suites and `play.py` pass.
- Shots: `shots/messages-pixel.png`, `shots/messages-smooth.png` (filling, blocked, set bit,
  hot fan).
- Open: the first-ever bar drop no longer shows an explaining card; the notebook still explains.

## 2026-10-23: ingot molds need no F; removing a mold returns its metal (Claude Code, with Codex)

Note: "now that the ingot molds drop the ingots, what is the purpose of F…". The owner chose no F
on ingot molds.

- `src/sim/molds.js` `ejectIngot`: when nothing over or in the mold is molten, junk and mixed
  metal (as scrap) and plugs over the trough or frozen in the grate are written as loose grains
  into the 4 rows under the mold; what doesn't fit waits. Slag and ash are thrown away. After a
  drop or a clean-out the grate above is raked. Bars and the short stub work as before.
- `collectMold` on an ingot mold only says "Ingot molds empty themselves." Bit and tank molds
  keep F. `removeMold` (every kind) gives the set metal to the bag: whole bars as bars, the rest
  as scrap; molten metal is left alone.
- Text: `troughState` flaws, `info.js`, `overlays.js` (no "· F" on ingots), `pieces.js`,
  `progress.js`, `notebook.js`, `docs/systems/stone-and-liquids.md`, `docs/DESIGN.md`.
- Tests: `t_molds` 60 checks (F does nothing on ingots; plate, slag and rake cases now
  automatic; junk out, mixed metal out as scrap, junk waits when blocked and while molten;
  removal returns bars and scrap, leaves molten metal). `furnace_play.js` adapted. All 11
  headless suites and `play.py` (122) pass.
- Open: undoing a mold removal brings the mold back empty (its metal stays in the bag).

## 2026-10-23: compact build card top-left (Claude Code, with Codex)

Note: "build tool has no info box like the other tools." The owner chose a compact card.

- `src/ui/selection.js`: the build tool has a card again: Q, piece name, cost (red if short), E;
  variant chips, Replace · G and ghosts waiting; one how-to line; the status line. No long
  description; that stays in the hover popup.
- `src/ui/build-menu.js`: the `.bb-status` row is gone (its contents are on the card); the
  variant chips are shared through `buildVariantChips(P)`. `src/styles/style.css`,
  `docs/DESIGN.md`, `docs/PROJECT_MAP.md`.
- `test/play.py`: the card is visible while building with name and cost and without the
  description; status and Replace · G are read from the card; no `.bb-status` (122 checks pass).
  All 11 headless suites pass; `t_smooth.py`'s half-resolution check is still flaky under load.
- Shots: `shots/build-card-*` (belt and wall, both looks, 1280×800 and 900×700) look right.
- Open: at 900px wide the bottom-left hover panel still sits partly under the build bar.

## 2026-10-23: Replace is held G, not Ctrl (Claude Code, with Codex)

Note: "i see the issue was ctrl+w closing the tab… it happens for the bookmark hotkey too. maybe
change the replace key?" The owner chose holding G.

- `src/ui/input.js`: `GAME.ctrl` became `GAME.replace`, set while G is held. Ctrl no longer
  does anything on its own, so WASD works while replacing. Ctrl+S, Ctrl+Z, Ctrl+wheel and the
  leave-page prompt are unchanged.
- Flag users and text: `previews.js`, `actions.js` ("Hold G to replace it."), `build-menu.js`
  (Replace · G chip, key hints), `loop.js`, `state.js`, `notebook.js` controls, `docs/DESIGN.md`.
- `test/play.py`: the replace checks hold G; a new check flies right with D while G is held
  (121 checks pass). All 11 headless suites pass. `t_smooth.py`'s half-resolution check failed
  on the slow fallback again (known, flaky under load).
- Open: nothing.

## 2026-10-23: Ctrl+W no longer closes the game without asking (Claude Code)

Note: "crash: hoplding crtl and pressing W crashes the whole browser."

- Cause: Ctrl+W is the browser's close-tab shortcut (the whole window if it is the last tab);
  holding Ctrl for replace while pressing W to fly hits it. Chrome and Brave reserve it, so the
  page never sees the key and cannot cancel it.
- `src/ui/input.js`: a `beforeunload` handler makes the browser ask before leaving while
  `GAME.running`. The `visibilitychange` autosave no longer depends on the audio context
  existing (it used to return early before any sound had played).
- `test/play.py`: 2 checks (leaving asks while a world is open, not on the title screen); 120
  checks pass. All 11 headless suites pass; `t_smooth.py` passed on its rerun (the known flaky
  half-resolution check failed once at load ~5).
- Open: the real Ctrl+W prompt can only be seen in a headed browser; the owner checks it.

## 2026-10-23: build menu hover popup, top-left card hidden while building (Claude Code, with Codex)

Notes: "mouse hover reviels a new info popup…" and "dont need the short description in the build
menu…". The owner chose to hide the top-left card entirely while building.

- `src/ui/build-menu.js`: `#bb-pop` above the hovered tile (icon, name, cost, size, power, full
  description, how to place, filter help, build keys), clamped to the window; on a gamepad it
  shows for the selected tile. The `.bb-tip` row became `.bb-status`: variant (material,
  direction, filter) chips, Replace · Ctrl, ghosts waiting, and status messages. The bar now
  updates its parts separately so hovering doesn't rebuild the tiles.
- `src/ui/selection.js`: `#selcard` is hidden for the build tool unless copying, cutting or
  pasting; the build branch moved to the bar. `src/styles/style.css`: popup and status styles;
  on narrow screens the bottom-right panel moves up while building.
- `src/ui/notebook.js`, `docs/DESIGN.md`, `docs/PROJECT_MAP.md`.
- Tests: `play.py` 118 checks (popup on hover, switching tiles, hidden on leave, gamepad,
  no `.bb-tip`, card hidden while building and back for tools 1 and 2, status in the bar; the
  Ctrl replace check now reads the bar). All 11 headless suites pass. `t_smooth.py`'s
  half-resolution check is still flaky under load (~4.8): 1 pass in 3 runs, as before.
- Screenshots in `shots/` (build-popup-*) at 1280×800 and 900×700 in both looks.
- Open: at 900px wide the bottom-left hover-info panel sits partly under the build bar.

## 2026-10-23: a backed-up heap on a belt no longer shears (Claude Code, with Codex)

Note: "belts with a big pile on top only move the grains on the top of the stack…". The owner's
setup: a belt ending in a screw tube going up; material mixed into the bottom layers never moved.

- Cause: `carryColumn` (`src/sim/conveyors.js`) skipped a blocked grain and carried on up the
  column, so with the front backed up the top slid over a stuck bottom. Reproduced headlessly
  first: over 4,000 ticks, 8,815 shear steps at the tube (1,415 against a wall) and only 57 of
  192 bottom sand grains left the belt.
- Fix: on a belt, the walk stops at the first blocked powder grain, so the load above stands
  still. On a filter the walk goes on only so chosen materials can still drop through; nothing
  above a blocked grain is carried. After: no shear, 119 of 192 sand grains out through the tube.
- `test/t_conveyors.js`: 3 new checks (no shear into an up tube, none against a wall, bottom
  layer leaves through the tube). `tools/scenes/belt-heap.js` sets up the owner's case for
  `tools/shot.py`; shots at 30 and 900 ticks look right. `docs/DESIGN.md`, `docs/PROJECT_MAP.md`.
- Tests: all 11 headless suites and `play.py` (109) pass. `t_smooth.py` failed its
  half-resolution check once at load 6.5 and passed on the rerun.
- Open: `t_smooth.py`'s half-resolution check is flaky on a busy machine (the game drops to
  pixel when frames are slow). It is worth making that test immune to the slow fallback.

## 2026-10-23: ingots drop out of the bottom of the mold (Claude Code, with Codex)

Note: "copper bars go into the furnace. also they shouldnt require the player to press F…".
The owner confirmed the first part was a bug: the knock-out threw bars up into the furnace.

- `src/sim/molds.js`: `ejectIngot` runs from `checkMolds` (every 15 ticks). Bars spawn directly
  under the mold, only where `entFreeAt` says there is room; the rest wait in the trough. Metal
  is taken from the top so the short stub stays on the floor. `ingot_drop` and `ingot_blocked`
  events (blocked is said once per spell via `mo.blockedSaid`, not saved). F on a ready ingot
  never launches bars upward; F on an empty, short or spoiled mold still cleans and rakes.
- `src/ui/events.js`: first drop gets a card, later drops a toast; blocked gets a toast. Ingots
  no longer emit `cast_ready`.
- Text: `pieces.js`, `progress.js`, `notebook.js`, `info.js`, `docs/DESIGN.md` (with the
  decision), `docs/systems/stone-and-liquids.md`, `docs/PROJECT_MAP.md`.
- Tests: `t_molds` (38 checks: downward drops, stub kept, top-up, blocked and warned once,
  partial drops, full entity pool, slag), `t_basics`, `t_stone`, `furnace_play.js` and
  `play.py` section 6 changed to the new behaviour. All 11 headless suites, `play.py` (109) and
  `t_smooth.py` (8) pass. A shot of a mold with a belt one block below shows the bars riding off.
- Open: a belt touching the mold's underside counts as blocked; bars need one open block. The
  owner may want a belt directly under the mold to work.

## 2026-10-23: Shift+Q/E build tabs, G pick key removed (Claude Code, with Codex)

Notes: "add shift + q/e to move tabs" and "having the "g pick" there is confusing".

- `src/ui/build-menu.js`: `selectBuildCat(c)` (shared with tab clicks) and `cycleBuildCat(d)`;
  the menu hint now reads Q/E cycle · Shift+Q/E tab.
- `src/ui/input.js`: Shift+Q/E switch tabs with the build tool; G is unbound.
- Text: build card (`selection.js`), notebook controls (`notebook.js`), `actions.js` comment,
  `docs/DESIGN.md`, `docs/PROJECT_MAP.md` (regenerated, `--check` passes).
- `test/play.py`: checks for Shift+Q/E, wrapping, the remembered piece per tab, plain Q/E,
  and that G no longer calls pick while middle mouse still does (109 checks).
- Tests: all 11 headless suites and `play.py` pass; a `tools/shot.py` shot of the Moving tab
  shows the new hint. `t_smooth.py` passes, but its "half-resolution smooth still draws" check
  failed on several runs with the 'slow' pixel fallback while the machine was busy (load ~14).
  The previous commit fails the same way under the same load, so this change didn't cause it.
- Open: that check is timing-sensitive on a loaded machine; rerun it when the machine is idle.

## 2026-10-23: browser tests run locally (Claude Code, with Codex)

Request: get the browser tests, Playwright and Chromium working on the owner's Arch machine
(Python 3.14 refuses a system-wide pip install).

- Playwright 1.64.1 and its Chromium 156 (revision 1248) are in an in-repo `.venv/` made with uv.

- Added `tools/use_venv.py` and called it before Playwright imports in `tools/shot.py`,
  `tools/perf.py`, `tools/playthrough.py`, `test/play.py` and `test/t_smooth.py`.
  The `sys.prefix` guard prevents an exec loop if the venv lacks Playwright.
- Updated `requirements.txt`, `.gitignore`, `README.md`, `AGENTS.md`, `docs/TESTING.md`
  and `docs/PROJECT_MAP.md` for Playwright 1.64.1 and venv setup.
- Checks: regenerated the project map and ran `tools/map.py --check`.
- Tests run with plain `python3`: `test/run_all.py --browser` all passed (11 headless suites,
  `play.py` 101 checks, `t_smooth.py` 8 checks) on the newer Chromium; `tools/shot.py` shots
  of the stone-uses scene look right in both looks. With `.venv` moved away, `play.py` exits
  with the setup message instead of a traceback.
- Open: nothing.

## 2026-10-23: export from claude.ai to a local repository (Claude, in claude.ai)

Request: split the source into small files, deduplicate and tidy a little, test fully, and
write the notes, docs and agent instructions for local development.

- **Source split.**
  - `core.js` (3,688 lines), `main.js` (2,012) and `render.js` (938) became 52 files in
    `src/sim/`, `src/render/` and `src/ui/`, each with a header line. The build order lives in
    `src/manifest.json`.
  - Checked that the joined code was identical to before, apart from comments and blank lines.
  - Every headless suite gave the same output, line for line, before and after.
  - Screenshots of two scenes, in both looks, were pixel-identical before and after
    (`tools/shot.py --freeze`).
- **Tidying, behaviour unchanged.**
  - One scan loop in `simMove` instead of two mirrored copies.
  - A `drawHazardRims` helper replaced 3 copies.
  - The belt and filter shader code shares `beltLineSD`, `bandRidge` and `bandColor`.
  - Removed unused `refundCost`, `wakeArea`, `brushFor`, and the `PARTS` / `PART_BY_ID` aliases.
  - The suites share `suite()` (check / done) from the harness, and failures now set the exit
    code.
- **Tests.**
  - Six broken old scripts were deleted; they used functions removed long ago.
  - Experiments and reports moved to `tools/experiments/`.
  - `t_v2` / `t_v4` / `t_v9` were renamed `t_basics` / `t_world` / `t_conveyors`.
  - Added `test/run_all.py`.
- **Tools.**
  - `tools/shot.py`: repeatable screenshots of a scene set up in JavaScript.
  - `tools/where.py`: maps built-page line numbers back to source files.
  - `tools/map.py`: the project map's index.
  - `tools/scenes/`: example scenes.
  - The game skips cursor previews when `GAME.hidePreviews` is set, for clean screenshots.
- **World files outside claude.ai.**
  - Pause → Export world file only worked inside claude.ai, through its downloads hook.
  - It now falls back to an ordinary browser download (`exportWorld` in `src/ui/saves.js`).
  - `test/play.py` checks that an exported file downloads and imports back on the title screen
    as the same world.
- **Build.** `build.py` writes `dist/index.html` (a full page), `dist/deepgrain-foundry.html`
  (the same without the document wrapper, as the claude.ai artifact used) and
  `dist/bundle-lines.json`.
- **Docs.** `AGENTS.md`, `CLAUDE.md`, `README.md`, `NOTES.md`, `docs/PROJECT_MAP.md`,
  `docs/DESIGN.md`, `docs/TESTING.md`, `docs/systems/*`, `docs/history/requests.md`.
- **Tests run.** All 11 headless suites pass, as do `test/play.py` (101 checks) and
  `test/t_smooth.py` (8 checks).
- **Open.**
  - The owner reports that big piles on belts still only move at the top (NOTES.md); not yet
    reproduced.
  - The game's last claude.ai artifact was version 24.